Human-robot collaboration towards a shared goal requires robots to understand human action and interaction with the surrounding environment. This paper focuses on humanrobot interaction (HRI) based on human-robot dialogue that relies on the robot action confirmation and action step generation using multimodal scene understanding. The state-of-theart approach uses multimodal transformers to generate robot action steps aligned with robot action confirmation from a single clip showing a task composed of multiple micro steps. Although actions towards a long-horizon task depend on each other throughout an entire video, the current approaches mainly focus on clip-level processing and do not leverage long-context information. This paper proposes a long-context Q-former incorporating left and right context dependency in full videos. Furthermore, this paper proposes a text-conditioning approach to feed text embeddings directly into the LLM decoder to mitigate the high abstraction of the information in text by Q-former. Experiments with the YouCook2 corpus show that the accuracy of confirmation generation is a major factor in the performance of action planning. Furthermore, we demonstrate that the longcontext Q-former improves the confirmation and action planning by integrating VideoLLaMA3.

Large-scale audio tagging datasets like AudioSet usually suffer from severe class imbalance comprising many audio examples for common sound classes but only few examples of rare sound classes. The latter, however, may yet be equally or even more important to recognize. Therefore, it is common practice to sample examples from rare classes more frequently during training. At the same time, the effects of such balancing on a model’s training and tagging performance are still little understood. In this work, we investigate how it affects training convergence and tagging performance. We consider varying degrees of balancing and investigate whether classes converge simultaneously or if there is a benefit from selecting different balancing rates for each class. Furthermore, we investigate data efficient oversampling, which keeps audio files from rare classes in memory, and repeats them in close succession over multiple batches, minimizing data loading from disk. Finally, we show that for AudioSet, the optimal amount of class balancing is different when fine-tuning a model pre-trained via self-supervised learning, versus training a supervised model from scratch.

We introduce O-EENC-SD: an end-to-end online speaker diarization system based on EEND-EDA, featuring a novel RNN-based stitching mechanism for online prediction. In particular, we develop a novel centroid refinement decoder whose usefulness is assessed through a rigorous ablation study. Our system provides key advantages over existing methods: a hyperparameter-free solution compared to unsupervised clustering approaches, and a more efficient alternative to current online end-to-end methods, which are computationally costly. We demonstrate that O-EENC-SD is competitive with the state of the art in the two-speaker conversational telephone speech domain, as tested on the CallHome dataset. Our results show that O-EENC-SD provides a great trade-off between DER and complexity, even when working on independent chunks with no overlap, making the system extremely efficient.

Understanding human actions could allow robots to perform a large spectrum of complex manipulation tasks and make collaboration with humans easier. Recently, multimodal scene understanding using audio-visual Transformers has been used to generate robot action sequences from videos of human demonstrations. However, automatic action sequence generation is not always perfect due to the distribution gap between the training and test environments. To bridge this gap, human intervention could be very effective, such as telling the robot agent what should be done. Motivated by this, we propose an error-correction-based action replanning approach that regenerates better action sequences using (1) automatically generated actions from a pretrained action generator and (2) human error-correction in natural language. We collected singlearm robot action sequences aligned to human action instruction for the cooking video dataset YouCook2. We trained the proposed errorcorrection-based action replanning model using a pre-trained multimodal LLM model (AVBLIP-2), generating a pair of (a) single-arm robot micro-step action sequences and (b) action descriptions in natural language simultaneously. To assess the performance of error correction, we collected human feedback on correcting errors in the automatically generated robot actions. Experiments show that our proposed interactive replanning model trained in a multitask manner using action sequence and description outperformed the baseline model in all types of scores.

Head-related transfer functions (HRTFs) with dense spatial grids are desired for immersive binaural audio generation, but their recording is time-consuming. Although HRTF spatial upsampling has shown remarkable progress with neural fields, spatial upsampling only from a few measured directions, e.g., 3 or 5 measurements, is still challenging. To tackle this problem, we propose a retrieval-augmented neural field (RANF). RANF retrieves a subject whose HRTFs are close to those of the target subject from a dataset. The HRTF of the retrieved subject at the desired direction is fed into the neural field in addition to the sound source direction itself. Furthermore, we present a neural network that can efficiently handle multiple retrieved subjects, inspired by a multi-channel processing technique called transform-average-concatenate. Our experiments confirm the benefits of RANF on the SONICOM dataset, and it is a key component in the winning solution of Task 2 of the listener acoustic personalization challenge 2024.

The goal of text-queried target sound extraction (TSE) is to extract from a mixture a sound source specified with a natural-language caption. While it is preferable to have access to large-scale text-audio pairs to address a variety of text queries, the limited number of available high-quality text-audio pairs hinders the data scaling. To this end, this work explores how to leverage audio-only data without any captions for the text-queried TSE task to potentially scale up the data amount. A straightforward way to do so is to use a joint audio-text embedding model, such as the contrastive language-audio pre-training (CLAP) model, as a query encoder and train a TSE model using audio embeddings obtained from the ground-truth audio. The TSE model can then accept text queries at inference time by switching to the text encoder. While this approach should work if the audio and text embedding spaces in CLAP were well aligned, in practice, the embeddings have domain-specific information that causes the TSE model to overfit to audio queries. We investigate several methods to avoid overfitting and show that simple embedding-manipulation methods such as dropout can effectively alleviate this issue. Extensive experiments demonstrate that using audio-only data with embedding dropout is as effective as using text captions during training, and audio-only data can be effectively leveraged to improve text-queried TSE models.

Several attempts have been made to handle multiple source separation tasks such as speech enhancement, speech separation, sound event separation, music source separation (MSS), or cinematic audio source separation (CASS) with a single model. These models are trained on large-scale data including speech, instruments, or sound events and can often successfully separate a wide range of sources. However, it is still challenging for such models to cover all separation tasks because some of them are contradictory (e.g., musical instruments are separated in MSS while they have to be grouped in CASS). To overcome this issue and support all the major separation tasks, we propose a task-aware unified source separation (TUSS) model. The model uses a variable number of learnable prompts to specify which source to separate, and changes its behavior depending on the given prompts, enabling it to handle all the major separation tasks including contradictory ones. Experimental results demonstrate that the proposed TUSS model successfully handles the five major separation tasks mentioned earlier. We also provide some audio examples, including both synthetic mixtures and real recordings, to demonstrate how flexibly the TUSS model changes its behavior at inference depending on the prompts.

Emitted sounds may drastically change when using different microphones, when properties of the sound sources change, or when recording in different acoustic environments. Ideally, anomalous sound detection (ASD) systems should be able to generalize well to unseen target domains by only providing a few target domain samples to define how normal data samples sound like, without needing to re-train or modify the system. In contrast with the source domain, for which many normal training samples are available, accurately estimating the underlying distribution of normal data after a domain shift based on very few samples is challenging. This usually leads to a mismatch between the corresponding anomaly scores of source and target domains and significantly reduces performance. In this work, we propose a framework for re-scaling anomaly scores based on the ratio between the cosine distance of a test sample to a normal reference sample and the distances to this sample’s next-closest neighbors in the reference set. In experimental evaluations, it is shown that the re-scaled anomaly scores reduce the domain mismatch for multiple domains. As a result, we obtain new state-of-the-art performances on the DCASE2020 and DCASE2023 ASD datasets.

International audience

We leverage Large Language Models (LLM) for zero-shot Semantic Audio Visual Navigation (SAVN). Existing methods utilize extensive training demonstrations for rein-forcement learning, yet achieve relatively low success rates and lack generalizability. The intermittent nature of au-ditory signals further poses additional obstacles to infer-ring the goal information. To address this challenge, we present the Reflective and Imaginative Language Agent (RILA). By employing multi-modal models to process sen-sory data, we instruct an LLM-based planner to actively ex-plore the environment. During the exploration, our agent adaptively evaluates and dismisses inaccurate perceptual descriptions. Additionally, we introduce an auxiliary LLM-based assistant to enhance global environmental compre-hension by mapping room layouts and providing strategic insights. Through comprehensive experiments and analy-sis, we show that our method outperforms relevant base-lines without training demonstrations from the environment and complementary semantic information.

Generative adversarial network (GAN) models can synthesize high-quality audio signals while ensuring fast sample generation. However, they are difficult to train and are prone to several issues including mode collapse and divergence. In this paper, we introduce SpecDiff-GAN, a neural vocoder based on HiFi-GAN, which was initially devised for speech synthesis from mel spectrogram. In our model, the training stability is enhanced by means of a forward diffusion process which consists in injecting noise from a Gaussian distribution to both real and fake samples before inputting them to the discriminator. We further improve the model by exploiting a spectrally-shaped noise distribution with the aim to make the discriminator's task more challenging. We then show the merits of our proposed model for speech and music synthesis on several datasets. Our experiments confirm that our model compares favorably in audio quality and efficiency compared to several baselines.

The introduction of audio latent diffusion models possessing the ability to generate realistic sound clips on demand from a text description has the potential to revolutionize how we work with audio. In this work, we make an initial attempt at understanding the inner workings of audio latent diffusion models by investigating how their audio outputs compare with the training data, similar to how a doctor auscultates a patient by listening to the sounds of their organs. Using text-to-audio latent diffusion models trained on the AudioCaps dataset, we systematically analyze memorization behavior as a function of training set size. We also evaluate different retrieval metrics for evidence of training data memorization, finding the similarity between mel spectrograms to be more robust in detecting matches than learned embedding vectors. In the process of analyzing memorization in audio latent diffusion models, we also discover a large amount of duplicated audio clips within the AudioCaps database.

Indoor monitoring systems are in high demand to protect vulnerable people, especially when they are alone at home, in nursing homes, hospitals, etc. Although surveillance systems in public spaces use cameras and microphones to find incidents, indoor monitoring in personal spaces needs to protect privacy. Such systems thus need to understand scenes without relying on direct sensing information, e.g., from audio-visual sensors, instead using indirect sensing information that is difficult to interpret by humans and may be insufficient to understand ongoing events precisely. To mitigate this drawback, this paper proposes a new indoor monitoring approach that attempts to realize scene understanding using only indirect sensors by transferring the learned inductive bias of a multimodal fusion model trained using direct and indirect sensing information to a model that uses only indirect information during inference. We collected direct (audio-visual) and indirect (infrared and Wi-Fi) sensing information of indoor human actions in daily life and manually annotated event captions. We build models that can generate event captions from various combinations of indirect and direct sensor data, and show that our transfer learning approach leads to significant improvements in caption quality when only indirect information is used at inference time.

Diffusion models are receiving a growing interest for a variety of signal generation tasks such as speech or music synthesis. WaveGrad, for example, is a successful diffusion model that conditionally uses the mel spectrogram to guide a diffusion process for the generation of high-fidelity audio. However, such models face important challenges concerning the noise diffusion process for training and inference, and they have difficulty generating high-quality speech for speakers that were not seen during training. With the aim of minimizing the conditioning error and increasing the efficiency of the noise diffusion process, we propose in this paper a new scheme called GLA-Grad, which consists in introducing a phase recovery algorithm such as the Griffin-Lim algorithm (GLA) at each step of the regular diffusion process. Furthermore, it can be directly applied to an already-trained waveform generation model, without additional training or fine-tuning. We show that our algorithm outperforms state-of-the-art diffusion models for speech generation, especially when generating speech for a previously unseen target speaker.

Head-related transfer functions (HRTFs) are important for immersive audio, and their spatial interpolation has been studied to upsample finite measurements. Recently, neural fields (NFs) which map from sound source direction to HRTF have gained attention. Existing NF-based methods focused on estimating the magnitude of the HRTF from a given sound source direction, and the magnitude is converted to a finite impulse response (FIR) filter. We propose the neural infinite impulse response filter field (NIIRF) method that instead estimates the coefficients of cascaded IIR filters. IIR filters mimic the modal nature of HRTFs, thus needing fewer coefficients to approximate them well compared to FIR filters. We find that our method can match the performance of existing NF-based methods on multiple datasets, even outperforming them when measurements are sparse. We also explore approaches to personalize the NF to a subject and experimentally find low-rank adaptation to be effective.

Neuro-steered speaker extraction aims to extract the listener’s brainattended speech signal from a multi-talker speech signal, in which the attention is derived from the cortical activity. This activity is usually recorded using electroencephalography (EEG) devices. Though promising, current methods often have a high speaker confusion error, where the interfering speaker is extracted instead of the attended speaker, degrading the listening experience. In this work, we aim to reduce the speaker confusion error in the neuro-steered speaker extraction model through a jointly fine-tuned auxiliary auditory attention detection model. The latter reinforces the consistency between the extracted target speech signal and the EEG representation, and also improves the EEG representation. Experimental results show that the proposed network significantly outperforms the baseline in terms of speaker confusion and overall signal quality in two-talker scenarios.

Automated audio captioning (AAC) aims to generate informative descriptions for various sounds from nature and/or human activities. In recent years, AAC has quickly attracted research interest, with state-of-the-art systems now relying on a sequence-to-sequence (seq2seq) backbone powered by strong models such as Transformers. Following the macro-trend of applied machine learning research, in this work, we strive to improve the performance of seq2seq AAC models by extensively leveraging pretrained models and large language models (LLMs). Specifically, we utilize BEATS to extract fine-grained audio features. Then, we employ Instructor LLM to fetch text embeddings of captions, and infuse their language-modality knowledge into BEATs audio features via an auxiliary InfoNCE loss function. Moreover, we propose a novel data augmentation method that uses ChatGPT to produce caption mix-ups (i.e., grammatical and compact combinations of two captions) which, together with the corresponding audio mixtures, increase not only the amount but also the complexity and diversity of training data. During inference, we propose to employ nucleus sampling and a hybrid reranking algorithm, which has not been explored in AAC research. Combining our efforts, our model achieves a new state-of-the-art 32.6 SPIDEr-FL score on the Clotho evaluation split, and wins the 2023 DCASE AAC challenge.

International audience

We study the problem of multimodal physical scene understanding, where an embodied agent needs to find fallen objects by inferring object properties, direction, and distance of an impact sound source. Previous works adopt feed-forward neural networks to directly regress the variables from sound, leading to poor generalization and domain adaptation issues. In this paper, we illustrate that learning a disentangled model of acoustic formation, referred to as disentangled acoustic field (DAF), to capture the sound generation and propagation process, enables the embodied agent to construct a spatial uncertainty map over where the objects may have fallen. We demonstrate that our analysis-by-synthesis framework can jointly infer sound properties by explicitly decomposing and factorizing the latent space of the disentangled model. We further show that the spatial uncertainty map can significantly improve the success rate for the localization of fallen objects by proposing multiple plausible exploration locations.

Since diarization and source separation of meeting data are closely related tasks, we here propose an approach to perform the two objectives jointly. It builds upon the target-speaker voice activity detection (TS-VAD) diarization approach, which assumes that initial speaker embeddings are available. We replace the final combined speaker activity estimation network of TS-VAD with a network that produces speaker activity estimates at a time-frequency resolution. Those act as masks for source extraction, either via masking or via beamforming. The technique can be applied both for single-channel and multi-channel input and, in both cases, achieves a new state-of-the-art word error rate (WER) on the LibriCSS meeting data recognition task. We further compute speaker-aware and speaker-agnostic WERs to isolate the contribution of diarization errors to the overall WER performance.

Target speech extraction aims to extract, based on a given conditioning cue, a target speech signal that is corrupted by interfering sources, such as noise or competing speakers. Building upon the achievements of the state-of-the-art (SOTA) time-frequency speaker separation model TF-GridNet, we propose AV-GridNet, a visual-grounded variant that incorporates the face recording of a target speaker as a conditioning factor during the extraction process. Recognizing the inherent dissimilarities between speech and noise signals as interfering sources, we also propose SAV-GridNet, a scenario-aware model that identifies the type of interfering scenario first and then applies a dedicated expert model trained specifically for that scenario. Our proposed model achieves SOTA results on the second COG-MHEAR Audio-Visual Speech Enhancement Challenge, outperforming other models by a significant margin, objectively and in a listening test. We also perform an extensive analysis of the results under the two scenarios.

This paper proposes reverberation as supervision (RAS), a novel unsupervised loss function for single-channel reverberant speech separation. Prior methods for unsupervised separation required the synthesis of mixtures of mixtures or assumed the existence of a teacher model, making them difficult to consider as potential methods explaining the emergence of separation abilities in an animal’s auditory system. We assume the availability of two-channel mixtures at training time, and train a neural network to separate the sources given one of the channels as input such that the other channel may be predicted from the separated sources. As the relationship between the room impulse responses (RIRs) of each channel depends on the locations of the sources, which are unknown to the network, the network cannot rely on learning that relationship. Instead, our proposed loss function fits each of the separated sources to the mixture in the target channel via Wiener filtering, and compares the resulting mixture to the ground-truth one. We show that minimizing the scale-invariant signal-to-distortion ratio (SI-SDR) of the predicted right-channel mixture with respect to the ground truth implicitly guides the network towards separating the left-channel sources. On a semi-supervised reverberant speech separation task based on the WHAMR! dataset, using training data where just 5% (resp., 10%) of the mixtures are labeled with associated isolated sources, we achieve 70% (resp., 78%) of the SI-SDR improvement obtained when training with supervision on the full training set, while a model trained only on the labeled data obtains 43% (resp., 45%).

Traditional source separation approaches train deep neural network models end-to-end with all the data available at once by minimizing the empirical risk on the whole training set. On the inference side, after training the model, the user fetches a static computation graph and runs the full model on some specified observed mixture signal to get the estimated source signals. Additionally, many of those models consist of several basic processing blocks which are applied sequentially. We argue that we can significantly increase resource efficiency during both training and inference stages by re-formulating a model’s training and inference procedures as iterative mappings of latent signal representations. First, we can apply the same processing block more than once on its output to refine the input signal and consequently improve parameter efficiency. During training, we can follow a block-wise procedure which enables a reduction on memory requirements. Thus, one can train a very complicated network structure using significantly less computation compared to end-to-end training. During inference, we can dynamically adjust how many processing blocks and iterations of a specific block an input signal needs using a gating module.

We introduce a framework for audio source separation using embeddings on a hyperbolic manifold that compactly represent the hierarchical relationship between sound sources and time-frequency features. Inspired by recent successes modeling hierarchical relationships in text and images with hyperbolic embeddings, our algorithm obtains a hyperbolic embedding for each time-frequency bin of a mixture signal and estimates masks using hyperbolic softmax layers. On a synthetic dataset containing mixtures of multiple people talking and musical instruments playing, our hyperbolic model performed comparably to a Euclidean baseline in terms of source to distortion ratio, with stronger performance at low embedding dimensions. Furthermore, we find that time-frequency regions containing multiple overlapping sources are embedded towards the center (i.e., the most uncertain region) of the hyperbolic space, and we can use this certainty estimate to efficiently trade-off between artifact introduction and interference reduction when isolating individual sounds.

Recent research has shown remarkable performance in leveraging multiple extraneous conditional and non-mutually-exclusive semantic concepts for sound source separation, allowing the flexibility to extract a given target source based on multiple different queries. In this work, we propose a new optimal condition training (OCT) method for single-channel target source separation, based on greedy parameter updates using the highest performing condition among equivalent conditions associated with a given target source. Our experiments show that the complementary information carried by the diverse semantic concepts significantly helps to disentangle and isolate sources of interest much more efficiently compared to single-conditioned models. Moreover, we propose a variation of OCT with condition refinement, in which an initial condition vector is adapted to the given mixture and transformed to a more amenable representation for target source extraction. We showcase the effectiveness of OCT on diverse source separation experiments where it improves upon permutation invariant models with oracle assignment between estimated and target sources and obtains state-of-the-art performance in the more challenging task of text-based source separation, outperforming even dedicated text-only conditioned models.

Diffusion models have recently shown promising results for difficult enhancement tasks such as the conditional and unconditional restoration of natural images and audio signals. In this work, we explore the possibility of leveraging a recently proposed advanced iterative diffusion model, namely cold diffusion, to recover clean speech signals from noisy signals. The unique mathematical properties of the sampling process from cold diffusion could be utilized to restore high-quality samples from arbitrary degradations. Based on these properties, we propose an improved training algorithm and objective to help the model generalize better during the sampling process. We verify our proposed framework by investigating two model architectures. Experimental results on benchmark speech enhancement dataset VoiceBank-DEMAND demonstrate the strong performance of the proposed approach compared to representative discriminative models and diffusion-based enhancement models.

Emulating the human ability to solve the cocktail party problem, i.e., focus on a source of interest in a complex acoustic scene, is a long standing goal of audio source separation research. In this paper, we focus on the cocktail fork problem, which takes a three-pronged approach to source separation by separating an audio mixture such as a movie soundtrack or podcast into the three broad categories of speech, music, and sound effects (SFX - understood to include ambient noise and natural sound events). We evaluate several deep learning-based source separation models on this task using simple objective measures such as signal-to-distortion ratio (SDR) as well as objective metrics that better correlate with human perception. Furthermore, we thoroughly evaluate how source separation can influence the downstream transcription asks of speech recognition for speech and audio tagging for music and SFX. We also investigate the task of activity detection on the three sources as a way to further improve source separation and transcription. While we observe that source separation improves transcription performance in comparison to the original soundtrack, performance is still sub-optimal due to artifacts introduced by the separation process. Therefore, we thoroughly investigate how remixing of the three separated source stems at various relative levels can reduce artifacts and consequently improve transcription performance. We find that remixing music and SFX interferences at a target SNR of 17.5 dB reduces speech recognition word error rate, and similar impact from remixing is observed for tagging music and SFX content.

Deep learning based speech enhancement in the short-time Fourier transform (STFT) domain typically uses a large window length such as 32 ms. A larger window can lead to higher frequency resolution and potentially better enhancement. This however incurs an algorithmic latency of 32 ms in an online setup, because the overlap-add algorithm used in the inverse STFT (iSTFT) is also performed using the same window size. To reduce this inherent latency, we adapt a conventional dual-window-size approach, where a regular input window size is used for STFT but a shorter output window is used for overlap-add, for STFT-domain deep learning based frame-online speech enhancement. Based on this STFT-iSTFT configuration, we employ complex spectral mapping for frame-online enhancement, where a deep neural network (DNN) is trained to predict the real and imaginary (RI) components of target speech from the mixture RI components. In addition, we use the DNN-predicted RI components to conduct frame-online beamforming, the results of which are used as extra features for a second DNN to perform frame-online postfiltering. The frequency-domain beamformer can be easily integrated with our DNNs and is designed to not incur any algorithmic latency. Additionally, we propose a future-frame prediction technique to further reduce the algorithmic latency. Evaluation on noisy-reverberant speech enhancement shows the effectiveness of the proposed algorithms. Compared with Conv-TasNet, our STFT-domain system can achieve better enhancement performance for a comparable amount of computation, or comparable performance with less computation, maintaining strong performance at an algorithmic latency as low as 2 ms.

Spatio-temporal scene-graph approaches to video-based reasoning tasks, such as video question-answering (QA), typically construct such graphs for every video frame. These approaches often ignore the fact that videos are essentially sequences of 2D ``views'' of events happening in a 3D space, and that the semantics of the 3D scene can thus be carried over from frame to frame. Leveraging this insight, we propose a (2.5+1)D scene graph representation to better capture the spatio-temporal information flows inside the videos. Specifically, we first create a 2.5D (pseudo-3D) scene graph by transforming every 2D frame to have an inferred 3D structure using an off-the-shelf 2D-to-3D transformation module, following which we register the video frames into a shared (2.5+1)D spatio-temporal space and ground each 2D scene graph within it. Such a (2.5+1)D graph is then segregated into a static sub-graph and a dynamic sub-graph, corresponding to whether the objects within them usually move in the world. The nodes in the dynamic graph are enriched with motion features capturing their interactions with other graph nodes. Next, for the video QA task, we present a novel transformer-based reasoning pipeline that embeds the (2.5+1)D graph into a spatio-temporal hierarchical latent space, where the sub-graphs and their interactions are captured at varied granularity. To demonstrate the effectiveness of our approach, we present experiments on the NExT-QA and AVSD-QA datasets. Our results show that our proposed (2.5+1)D representation leads to faster training and inference, while our hierarchical model showcases superior performance on the video QA task versus the state of the art.

Graph-based temporal classification (GTC), a generalized form of the connectionist temporal classification loss, was recently proposed to improve automatic speech recognition (ASR) systems using graph-based supervision. For example, GTC was first used to encode an N-best list of pseudo-label sequences into a graph for semi-supervised learning. In this paper, we propose an extension of GTC to model the posteriors of both labels and label transitions by a neural network, which can be applied to a wider range of tasks. As an example application, we use the extended GTC (GTC-e) for the multi-speaker speech recognition task. The transcriptions and speaker information of multi-speaker speech are represented by a graph, where the speaker information is associated with the transitions and ASR outputs with the nodes. Using GTC-e, multi-speaker ASR modelling becomes very similar to single-speaker ASR modeling, in that tokens by multiple speakers are recognized as a single merged sequence in chronological order. For evaluation, we perform experiments on a simulated multi-speaker speech dataset derived from LibriSpeech, obtaining promising results with performance close to classical benchmarks for the task.

Pseudo-labeling (PL), a semi-supervised learning (SSL) method where a seed model performs self-training using pseudo-labels generated from untranscribed speech, has been shown to enhance the performance of end-to-end automatic speech recognition (ASR). Our prior work proposed momentum pseudo-labeling (MPL), which performs PL-based SSL via an interaction between online and offline models, inspired by the mean teacher framework. MPL achieves remarkable results on various semi-supervised settings, showing robustness to variations in the amount of data and domain mismatch severity. However, there is further room for improving the seed model used to initialize the MPL training, as it is in general critical for a PL-based method to start training from high-quality pseudo-labels. To this end, we propose to enhance MPL by (1) introducing the Conformer architecture to boost the overall recognition accuracy and (2) exploiting iterative pseudo-labeling with a language model to improve the seed model before applying MPL. The experimental results demonstrate that the proposed approaches effectively improve MPL performance, outperforming other PL-based methods. We also present in-depth investigations to make our improvements effective, e.g., with regard to batch normalization typically used in Conformer and LM quality.

The recurrent neural network transducer (RNN-T) objective plays a major role in building today’s best automatic speech recognition (ASR) systems for production. Similarly to the connectionist temporal classification (CTC) objective, the RNN-T loss uses specific rules that define how a set of alignments is generated to form a lattice for the full-sum training. However, it is yet largely unknown if these rules are optimal and do lead to the best possible ASR results. In this work, we present a new transducer objective function that generalizes the RNN-T loss to accept a graph representation of the labels, thus providing a flexible and efficient framework to manipulate training lattices, e.g., for studying different transition rules, implementing different transducer losses, or restricting alignments. We demonstrate that transducer-based ASR with CTC-like lattice achieves better results compared to standard RNN-T, while also ensuring a strictly monotonic alignment, which will allow better optimization of the decoding procedure. For example, the proposed CTC-like transducer achieves an improvement of 4.8% on the test-other condition of LibriSpeech relative to an equivalent RNN-T based system.

The cocktail party problem aims at isolating any source of interest within a complex acoustic scene, and has long inspired audio source separation research. Recent efforts have mainly focused on separating speech from noise, speech from speech, musical instruments from each other, or sound events from each other. However, separating an audio mixture (e.g., movie soundtrack) into the three broad categories of speech, music, and sound effects (understood to include ambient noise and natural sound events) has been left largely unexplored, despite a wide range of potential applications. This paper formalizes this task as the cocktail fork problem, and presents the Divide and Remaster (DnR) dataset to foster research on this topic. DnR is built from three well-established audio datasets (LibriSpeech, FMA, FSD50k), taking care to reproduce conditions similar to professionally produced content in terms of source overlap and relative loudness, and made available at CD quality. We benchmark standard source separation algorithms on DnR, and further introduce a new multi-resolution model to better address the variety of acoustic characteristics of the three source types. Our best model produces SI-SDR improvements over the mixture of 11.0 dB for music, 11.2 dB for speech, and 10.8 dB for sound effects.

In previous work, we have proposed the Audio-Visual Scene-Aware Dialog (AVSD) task, collected an AVSD dataset, developed AVSD technologies, and hosted an AVSD challenge track at both the 7th and 8th Dialog System Technology Challenges (DSTC7, DSTC8). In these challenges, the best-performing systems relied heavily on human-generated descriptions of the video content, which were available in the datasets but would be unavailable in real-world applications. To promote further advancements for real-world applications, we proposed a third AVSD challenge, at DSTC10, with two modifications: 1) the human-created description is unavailable at inference time, and 2) systems must demonstrate temporal reasoning by finding evidence from the video to support each answer. This paper introduces the new task that includes temporal reasoning and our new extension of the AVSD dataset for DSTC10, for which we collected human-generated temporal reasoning data. We also introduce a baseline system built using an AV-transformer, which we released along with the new dataset. Finally, this paper introduces a new system that extends our baseline system with attentional multimodal fusion, joint student-teacher learning (JSTL), and model combination techniques, achieving state-of-the-art performances on the AVSD datasets for DSTC7, DSTC8, and DSTC10. We also propose two temporal reasoning methods for AVSD: one attention-based, and one based on a time-domain region proposal network.

Existing systems for sound event localization and detection (SELD) typically operate by estimating a source location for all classes at every time instant. In this paper, we propose an alternative class-conditioned SELD model for situations where we may not be interested in localizing all classes all of the time. This class-conditioned SELD model takes as input the spatial and spectral features from the sound file, and also a one-hot vector indicating the class we are currently interested in localizing. We inject the conditioning information at several points in our model using feature-wise linear modulation (FiLM) layers. Through experiments on the DCASE 2020 Task 3 dataset, we show that the proposed class-conditioned SELD model performs better in terms of common SELD metrics than the baseline model that locates all classes simultaneously, and also outperforms specialist models that are trained to locate only a single class of interest. We also evaluate performance on the DCASE 2021 Task 3 dataset, which includes directional interference (sound events from classes we are not interested in localizing) and notice especially strong improvement from the class-conditioned model.

We introduce a new paradigm for single-channel target source separation where the sources of interest can be distinguished using non-mutually exclusive concepts (e.g., loudness, gender, language, spatial location, etc).Our proposed heterogeneous separation framework can seamlessly leverage datasets with large distribution shifts and learn cross-domain representations under a variety of concepts used as conditioning.Our experiments show that training separation models with heterogeneous conditions facilitates the generalization to new concepts with unseen out-of-domain data while also performing substantially higher than single-domain specialist models.Notably, such training leads to more robust learning of new harder source separation discriminative concepts and can yield improvements over permutation invariant training with oracle source selection.We analyze the intrinsic behavior of source separation training with heterogeneous metadata and propose ways to alleviate emerging problems with challenging separation conditions.We release the collection of preparation recipes for all datasets used to further promote research towards this challenging task.

Given an input video, its associated audio, and a brief caption, the audio-visual scene aware dialog (AVSD) task requires an agent to indulge in a question-answer dialog with a human about the audio-visual content. This task thus poses a challenging multi-modal representation learning and reasoning scenario, advancements into which could influence several human-machine interaction applications. To solve this task, we introduce a semantics-controlled multi-modal shuffled Transformer reasoning framework, consisting of a sequence of Transformer modules, each taking a modality as input and producing representations conditioned on the input question. Our proposed Transformer variant uses a shuffling scheme on their multi-head outputs, demonstrating better regularization. To encode fine-grained visual information, we present a novel dynamic scene graph representation learning pipeline that consists of an intra-frame reasoning layer producing spatio-semantic graph representations for every frame, and an inter-frame aggregation module capturing temporal cues. Our entire pipeline is trained end-to-end. We present experiments on the benchmark AVSD dataset, both on answer generation and selection tasks. Our results demonstrate state-of-the-art performances on all evaluation metrics.

Most music source separation systems require large collections of isolated sources for training, which can be difficult to obtain. In this work, we use musical scores, which are comparatively easy to obtain, as a weak label for training a source separation system. In contrast with previous score-informed separation approaches, our system does not require isolated sources, and score is used only as a training target, not required for inference. Our model consists of a separator that outputs a time-frequency mask for each instrument, and a transcriptor that acts as a critic, providing both temporal and frequency supervision to guide the learning of the separator. A harmonic mask constraint is introduced as another way of leveraging score information during training, and we propose two novel adversarial losses for additional fine-tuning of both the transcriptor and the separator. Results demonstrate that using score information outper-forms temporal weak-labels, and adversarial structures lead to further improvements in both separation and transcription performance.

The performance of automatic speech recognition (ASR) systems typically degrades significantly when the training and test data domains are mismatched. In this paper, we show that self-training (ST) combined with an uncertainty-based pseudo-label filtering approach can be effectively used for domain adaptation. We propose DUST, a dropout-based uncertainty-driven self-training technique which uses agreement between multiple predictions of an ASR system obtained for different dropout settings to measure the model’s uncertainty about its prediction. DUST excludes pseudo-labeled data with high uncertainties from the training, which leads to substantially improved ASR results compared to ST without filtering, and accelerates the training time due to a reduced training data set. Domain adaptation experiments using WSJ as a source domain and TED-LIUM 3 as well as SWITCHBOARD as the target domains show that up to 80% of the performance of a system trained on ground-truth data can be recovered.

Self-attention has become an important and widely used neural network component that helped to establish new state-of-the-art results for various applications, such as machine translation and automatic speech recognition (ASR). However, the computational complexity of self-attention grows quadratically with the input sequence length. This can be particularly problematic for applications such as ASR, where an input sequence generated from an utterance can be relatively long. In this work, we propose a combination of restricted self-attention and a dilation mechanism, which we refer to as dilated self-attention. The restricted self-attention allows attention to neighboring frames of the query at a high resolution, and the dilation mechanism summarizes distant information to allow attending to it with a lower resolution. Different methods for summarizing distant frames are studied, such as subsampling, mean-pooling, and attention-based pooling. ASR results demonstrate substantial improvements compared to restricted self-attention alone, achieving similar results compared to full-sequence based self-attention with a fraction of the computational costs.

Semi-supervised learning has demonstrated promising results in automatic speech recognition (ASR) by self-training using a seed ASR model with pseudo-labels generated for unlabeled data. The effectiveness of this approach largely relies on the pseudo-label accuracy, for which typically only the 1-best ASR hypothesis is used. However, alternative ASR hypotheses of an N-best list can provide more accurate labels for an unlabeled speech utterance and also reflect uncertainties of the seed ASR model. In this paper, we propose a generalized form of the connectionist temporal classification (CTC) objective that accepts a graph representation of the training labels. The newly proposed graph-based temporal classification (GTC) objective is applied for self-training with WFST-based supervision, which is generated from an N-best list of pseudo-labels. In this setup, GTC is used to learn not only a temporal alignment, similarly to CTC, but also a label alignment to obtain the optimal pseudo-label sequence from the weighted graph. Results show that this approach can effectively exploit an N-best list of pseudo-labels with associated scores, considerably outperforming standard pseudo-labeling, with ASR results approaching an oracle experiment in which the best hypotheses of the N-best lists are selected manually.

State-of-the-art approaches for visually-guided audio source separation typically assume sources that have characteristic sounds, such as musical instruments. These approaches often ignore the visual context of these sound sources or avoid modeling object interactions that may be useful to better characterize the sources, especially when the same object class may produce varied sounds from distinct interactions. To address this challenging problem, we propose Audio Visual Scene Graph Segmenter (AVSGS), a novel deep learning model that embeds the visual structure of the scene as a graph and segments this graph into subgraphs, each subgraph being associated with a unique sound obtained by co-segmenting the audio spectrogram. At its core, AVSGS uses a recursive neural network that emits mutually-orthogonal sub-graph embeddings of the visual graph using multi-head attention. These embeddings are used for conditioning an audio encoder-decoder towards source separation. Our pipeline is trained end-to-end via a self-supervised task consisting of separating audio sources using the visual graph from artificially mixed sounds.In this paper, we also introduce an “in the wild” video dataset for sound source separation that contains multiple non-musical sources, which we call Audio Separation in the Wild (ASIW). This dataset is adapted from the AudioCaps dataset, and provides a challenging, natural, and daily-life setting for source separation. Thorough experiments on the proposed ASIW and the standard MUSIC datasets demonstrate state-of-the-art sound separation performance of our method against recent prior approaches.

Pseudo-labeling (PL) has been shown to be effective in semisupervised automatic speech recognition (ASR), where a base model is self-trained with pseudo-labels generated from unlabeled data.While PL can be further improved by iteratively updating pseudo-labels as the model evolves, most of the previous approaches involve inefficient retraining of the model or intricate control of the label update.We present momentum pseudo-labeling (MPL), a simple yet effective strategy for semisupervised ASR.MPL consists of a pair of online and offline models that interact and learn from each other, inspired by the mean teacher method.The online model is trained to predict pseudo-labels generated on the fly by the offline model.The offline model maintains a momentum-based moving average of the online model.MPL is performed in a single training process and the interaction between the two models effectively helps them reinforce each other to improve the ASR performance.We apply MPL to an end-to-end ASR model based on the connectionist temporal classification.The experimental results demonstrate that MPL effectively improves over the base model and is scalable to different semi-supervised scenarios with varying amounts of data or domain mismatch.

Video captioning is an essential technology to understand scenes and describe events in natural language.To apply it to real-time monitoring, a system needs not only to describe events accurately but also to produce the captions as soon as possible.Low-latency captioning is needed to realize such functionality, but this research area for online video captioning has not been pursued yet.This paper proposes a novel approach to optimize each caption's output timing based on a trade-off between latency and caption quality.An audio-visual Transformer is trained to generate ground-truth captions using only a small portion of all video frames, and to mimic outputs of a pre-trained Transformer to which all the frames are given.A CNN-based timing detector is also trained to detect a proper output timing, where the captions generated by the two Transformers become sufficiently close to each other.With the jointly trained Transformer and timing detector, a caption can be generated in the early stages of an event-triggered video clip, as soon as an event happens or when it can be forecasted.Experiments with the ActivityNet Captions dataset show that our approach achieves 94% of the caption quality of the upper bound given by the pre-trained Transformer using the entire video clips, using only 28% of frames from the beginning.

This paper addresses end-to-end automatic speech recognition (ASR) for long audio recordings such as lecture and conversational speeches.Most end-to-end ASR models are designed to recognize independent utterances, but contextual information (e.g., speaker or topic) over multiple utterances is known to be useful for ASR.In our prior work, we proposed a contextexpanded Transformer that accepts multiple consecutive utterances at the same time and predicts an output sequence for the last utterance, achieving 5-15% relative error reduction from utterance-based baselines in lecture and conversational ASR benchmarks.Although the results have shown remarkable performance gain, there is still potential to further improve the model architecture and the decoding process.In this paper, we extend our prior work by (1) introducing the Conformer architecture to further improve the accuracy, (2) accelerating the decoding process with a novel activation recycling technique, and (3) enabling streaming decoding with triggered attention.We demonstrate that the extended Transformer provides state-of-the-art end-to-end ASR performance, obtaining a 17.3% character error rate for the HKUST dataset and 12.0%/6.3%word error rates for the Switchboard-300 Eval2000 CallHome/Switchboard test sets.The new decoding method reduces decoding time by more than 50% and further enables streaming ASR with limited accuracy degradation.

Attention-based end-to-end automatic speech recognition (ASR) systems have recently demonstrated state-of-the-art results for numerous tasks. However, the application of self-attention and attention-based encoder-decoder models remains challenging for streaming ASR, where each word must be recognized shortly after it was spoken. In this work, we present the dual causal/non-causal self-attention (DCN) architecture, which in contrast to restricted self-attention prevents the overall context to grow beyond the look-ahead of a single layer when used in a deep architecture. DCN is compared to chunk-based and restricted self-attention using streaming transformer and conformer architectures, showing improved ASR performance over restricted self-attention and competitive ASR results compared to chunk-based self-attention, while providing the advantage of frame-synchronous processing. Combined with triggered attention, the proposed streaming end-to-end ASR systems obtained state-of-the-art results on the LibriSpeech, HKUST, and Switchboard ASR tasks.

Deep neural network (DNN) based end-to-end optimization in the complex time-frequency (T-F) domain or time domain has shown considerable potential in monaural speech separation. Many recent studies optimize loss functions defined solely in the time or complex domain, without including a loss on magnitude. Although such loss functions typically produce better scores if the evaluation metrics are objective time-domain metrics, they however produce worse scores on speech quality and intelligibility metrics and usually lead to worse speech recognition performance, compared with including a loss on magnitude. While this phenomenon has been experimentally observed by many studies, it is often not accurately explained and there lacks a thorough understanding on its fundamental cause. This paper provides a novel view from the perspective of the implicit compensation between estimated magnitude and phase. Analytical results based on monaural speech separation and robust automatic speech recognition (ASR) tasks in noisy-reverberant conditions support the validity of our view.

A promising approach for speech dereverberation is based on supervised learning, where a deep neural network (DNN) is trained to predict the direct sound from noisy-reverberant speech. This data-driven approach is based on leveraging prior knowledge of clean speech patterns, and seldom explicitly exploits the linear-filter structure in reverberation, i.e., that reverberation results from a linear convolution between a room impulse response (RIR) and a dry source signal. In this work, we propose to exploit this linear-filter structure within a deep learning based monaural speech dereverberation framework. The key idea is to first estimate the direct-path signal of the target speaker using a DNN and then identify signals that are decayed and delayed copies of the estimated direct-path signal, as these can be reliably considered as reverberation. They can be either directly removed for dereverberation, or used as extra features for another DNN to perform better dereverberation. To identify the copies, we estimate the underlying filter (or RIR) by efficiently solving a linear regression problem per frequency in the time-frequency domain. We then modify the proposed algorithm for speaker separation in reverberant and noisy-reverberant conditions. State-of-the-art speech dereverberation and speaker separation results are obtained on the REVERB, SMS-WSJ, and WHAMR! datasets.

Recently, fully recurrent neural network (RNN) based end-to-end models have been proven to be effective for multi-speaker speech recognition in both the single-channel and multi-channel scenarios. In this work, we explore the use of Transformer models for these tasks by focusing on two aspects. First, we replace the RNN-based encoder-decoder in the speech recognition model with a Transformer architecture. Second, in order to use the Transformer in the masking network of the neural beamformer in the multi-channel case, we modify the self-attention component to be restricted to a segment rather than the whole sequence in order to reduce computation. Besides the model architecture improvements, we also incorporate an external dereverberation preprocessing, the weighted prediction error (WPE), enabling our model to handle reverberated signals. Experiments on the spatialized wsj1-2mix corpus show that the Transformer-based models achieve 40.9% and 25.6% relative WER reduction, down to 12.1% and 6.4% WER, under the anechoic condition in single-channel and multi-channel tasks, respectively, while in the reverberant case, our methods achieve 41.5% and 13.8% relative WER reduction, down to 16.5% and 15.2% WER.

While significant advances have been made with respect to the separation of overlapping speech signals, studies have been largely constrained to mixtures of clean, near anechoic speech, not representative of many real-world scenarios. Although the WHAM! dataset introduced noise to the ubiquitous wsj0-2mix dataset, it did not include reverberation, which is generally present in indoor recordings outside of recording studios. The spectral smearing caused by reverberation can result in significant performance degradation for standard deep learning-based speech separation systems, which rely on spectral structure and the sparsity of speech signals to tease apart sources. To address this, we introduce WHAMR!, an augmented version of WHAM! with synthetic reverberated sources, and provide a thorough baseline analysis of current techniques as well as novel cascaded architectures on the newly introduced conditions.

Encoder-decoder based sequence-to-sequence models have demonstrated state-of-the-art results in end-to-end automatic speech recognition (ASR). Recently, the transformer architecture, which uses self-attention to model temporal context information, has been shown to achieve significantly lower word error rates (WERs) compared to recurrent neural network (RNN) based system architectures. Despite its success, the practical usage is limited to offline ASR tasks, since encoder-decoder architectures typically require an entire speech utterance as input. In this work, we propose a transformer based end-to-end ASR system for streaming ASR, where an output must be generated shortly after each spoken word. To achieve this, we apply time-restricted self-attention for the encoder and triggered attention for the encoder-decoder attention mechanism. Our proposed streaming transformer architecture achieves 2.8% and 7.3% WER for the "clean" and "other" test data of LibriSpeech, which to our knowledge is the best published streaming end-to-end ASR result for this task.

Deep learning models for monaural audio source separation are typically trained on large collections of isolated sources, which may not be available in domains such as environmental monitoring. We propose objective functions and network architectures that enable training a source separation system with weak labels. In contrast with strong time-frequency (TF) labels, weak labels only indicate the time periods where different sources are active in this scenario. We train a separator that outputs a TF mask for each type of sound event, using a classifier to pool label estimates across frequency. Our objective function requires the classifier applied to a separated source to output weak labels for the class corresponding to that source and zeros for all other classes. The objective function also enforces that the separated sources sum to the mixture. We benchmark performance using synthetic mixtures of overlapping sound events recorded in urban environments. Compared to training on mixtures and their isolated sources, our model still achieves significant SDR improvement.

We propose an unsupervised speaker adaptation method inspired by the neural Turing machine for end-to-end (E2E) automatic speech recognition (ASR). The proposed model contains a memory block that holds speaker i-vectors extracted from the training data and reads relevant i-vectors from the memory through an attention mechanism. The resulting memory vector (M-vector) is concatenated to the acoustic features or to the hidden layer activations of an E2E neural network model. The E2E ASR system is based on the joint connectionist temporal classification and attention-based encoder-decoder architecture. M-vector and i-vector results are compared for inserting them at different layers of the encoder neural network using the WSJ and TED-LIUM2 ASR benchmarks. We show that M-vectors, which do not require an auxiliary speaker embedding extraction system at test time, achieve similar word error rates (WERs) compared to i-vectors for single speaker utterances and significantly lower WERs for utterances in which there are speaker changes.

Various adversarial audio attacks have recently been developed to fool automatic speech recognition (ASR) systems. We here propose a defense against such attacks based on the uncertainty introduced by dropout in neural networks. We show that our defense is able to detect attacks created through optimized perturbations and frequency masking on a state-of-the-art end-to-end ASR system. Furthermore, the defense can be made robust against attacks that are immune to noise reduction. We test our defense on Mozilla's CommonVoice dataset, the UrbanSound dataset, and an excerpt of the LibriSpeech dataset, showing that it achieves high detection accuracy in a wide range of scenarios.

While there has been much recent progress using deep learning techniques to separate speech and music audio signals, these systems typically require large collections of isolated sources during the training process. When extending audio source separation algorithms to more general domains such as environmental monitoring, it may not be possible to obtain isolated signals for training. Here, we propose objective functions and network architectures that enable training a source separation system with weak labels. In this scenario, weak labels are defined in contrast with strong time-frequency (TF) labels such as those obtained from isolated sources, and refer either to frame-level weak labels where one only has access to the time periods when different sources are active in an audio mixture, or to clip-level weak labels that only indicate the presence or absence of sounds in an entire audio clip. We train a separator that estimates a TF mask for each type of sound event, using a sound event classifier as an assessor of the separator's performance to bridge the gap between the TF-level separation and the ground truth weak labels only available at the frame or clip level. Our objective function requires the separator to estimate a source such that the classifier applied to it will assign high probability to the class corresponding to that source and low probability to all other classes. The objective function also enforces that the separated sources sum up to the mixture. We benchmark the performance of our algorithm using synthetic mixtures of overlapping events created from a database of sounds recorded in urban environments, and show that the method can also be applied to other tasks such as music source separation. Compared to training a network using isolated sources, our model achieves somewhat lower but still significant SI-SDR improvement, even in scenarios with significant sound event overlap.

Recently, the end-to-end approach has proven its efficacy in monaural multi-speaker speech recognition. However, high word error rates (WERs) still prevent these systems from being used in practical applications. On the other hand, the spatial information in multi-channel signals has proven helpful in far-field speech recognition tasks. In this work, we propose a novel neural sequence-to-sequence (seq2seq) architecture, MIMO-Speech, which extends the original seq2seq to deal with multi-channel input and multi-channel output so that it can fully model multi-channel multi-speaker speech separation and recognition. MIMO-Speech is a fully neural end-to-end framework, which is optimized only via an ASR criterion. It is comprised of: 1) a monaural masking network, 2) a multi-source neural beamformer, and 3) a multi-output speech recognition model. With this processing, the input overlapped speech is directly mapped to text sequences. We further adopted a curriculum learning strategy, making the best use of the training set to improve the performance. The experiments on the spatialized wsj1-2mix corpus show that our model can achieve more than 60% WER reduction compared to the single-channel system with high quality enhanced signals (SI-SDR = 23.1 dB) obtained by the above separation function.

In this paper, we present a one-pass decoding algorithm for streaming recognition with joint connectionist temporal classification (CTC) and attention-based end-to-end automatic speech recognition (ASR) models. The decoding scheme is based on a frame-synchronous CTC prefix beam search algorithm and the recently proposed triggered attention concept. To achieve a fully streaming end-to-end ASR system, the CTC-triggered attention decoder is combined with a unidirectional encoder neural network based on parallel time-delayed long short-term memory (PTDLSTM) streams, which has demonstrated superior performance compared to various other streaming encoder architectures in earlier work. A new type of pre-training method is studied to further improve our streaming ASR models by adding residual connections to the encoder neural network and layer-wise removing them during the training process. The proposed joint CTC-triggered attention decoding algorithm, which enables streaming recognition of attention-based ASR systems, achieves similar ASR results compared to offline CTC-attention decoding and significantly better results compared to CTC prefix beam search decoding alone.

The recently-proposed deep clustering algorithm introduced significant advances in monaural speaker-independent multi-speaker speech separation. Deep clustering operates on magnitude spectro-grams using bidirectional recurrent networks and K-means clustering, both of which require offline operation, i.e., algorithm latency is longer than utterance length. This paper evaluates architectures for reduced latency deep clustering by combining: (1) block processing to efficiently propagate the memory encoded by the recurrent network, and (2) teacher-student learning, where low-latency models learn from an offline teacher. Compared to our best performing offline model, we only lose 0.3 dB SDR at a latency of 1.2 seconds and 0.7 dB SDR at a latency of 0.6 seconds on the publicly available wsj0-2mix dataset. Moreover, by providing a detailed analysis of the failure cases for our low-latency speech separation models, we show that the cause of this performance gap is related to frame-level permutation errors, where the network fails to accurately track speaker identity throughout an utterance.

This paper presents a method to train end-to-end automatic speech recognition (ASR) models using unpaired data. Although the end-to-end approach can eliminate the need for expert knowledge such as pronunciation dictionaries to build ASR systems, it still requires a large amount of paired data, i.e., speech utterances and their transcriptions. Cycle-consistency losses have been recently proposed as a way to mitigate the problem of limited paired data. These approaches compose a reverse operation with a given transformation, e.g., text-to-speech (TTS) with ASR, to build a loss that only requires unsupervised data, speech in this example. Applying cycle consistency to ASR models is not trivial since fundamental information, such as speaker traits, are lost in the intermediate text bottleneck. To solve this problem, this work presents a loss that is based on the speech encoder state sequence instead of the raw speech signal. This is achieved by training a Text-To-Encoder model and defining a loss based on the encoder reconstruction error. Experimental results on the LibriSpeech corpus show that the proposed cycle-consistency training reduced the word error rate by 14.7% from an initial model trained with 100-hour paired data, using an additional 360 hours of audio data without transcriptions. We also investigate the use of text-only data mainly for language modeling to further improve the performance in the unpaired data training scenario.

A new system architecture for end-to-end automatic speech recognition (ASR) is proposed that combines the alignment capabilities of the connectionist temporal classification (CTC) approach and the modeling strength of the attention mechanism. The proposed system architecture, named triggered attention (TA), uses a CTC-based classifier to control the activation of an attention-based decoder neural network. This allows for a frame-synchronous decoding scheme with an adjustable look-ahead parameter to control the induced delay and opens the door to streaming recognition with attention-based end-to-end ASR systems. We present ASR results of the TA model on three data sets of different size and language and compare the scores to a well-tuned attention-based end-to-end ASR baseline system, which consumes input frames in the traditional full-sequence manner. The proposed triggered attention (TA) decoder concept achieves similar or better ASR results in all experiments compared to the full-sequence attention model, while also limiting the decoding delay to two look-ahead frames, which in our setup corresponds to an output delay of 80 ms.

In speech enhancement and source separation, signal-to-noise ratio is a ubiquitous objective measure of denoising/separation quality. A decade ago, the BSS_eval toolkit was developed to give researchers worldwide a way to evaluate the quality of their algorithms in a simple, fair, and hopefully insightful way: it attempted to account for channel variations, and to not only evaluate the total distortion in the estimated signal but also split it in terms of various factors such as remaining interference, newly added artifacts, and channel errors. In recent years, hundreds of papers have been relying on this toolkit to evaluate their proposed methods and compare them to previous works, often arguing that differences on the order of 0.1 dB proved the effectiveness of a method over others. We argue here that the signal-to-distortion ratio (SDR) implemented in the BSS_eval toolkit has generally been improperly used and abused, especially in the case of single-channel separation, resulting in misleading results. We propose to use a slightly modified definition, resulting in a simpler, more robust measure, called scale-invariant SDR (SI-SDR). We present various examples of critical failure of the original SDR that SI-SDR overcomes.

Deep learning based speech enhancement and source separation systems have recently reached unprecedented levels of quality, to the point that performance is reaching a new ceiling. Most systems rely on estimating the magnitude of a target source, either directly or by computing a real-valued mask to be applied to a time-frequency representation of the mixture signal. A limiting factor in such approaches is a lack of phase estimation: the phase of the mixture is most often used when reconstructing the estimated time-domain signal. We propose to estimate phase using "phasebook", a new type of layer based on a discrete representation of the phase difference between the mixture and the target. We also introduce "combook", a similar type of layer that directly estimates a complex mask. We present various training and inference schemes involving these representations, and explain in particular how to include them in an end-to-end learning framework. We also present an oracle study to assess upper bounds on performance for various types of masks using discrete phase representations. We evaluate the proposed methods on the wsj0-2mix dataset, a well-studied corpus for single-channel speaker-independent speaker separation, matching the performance of state-of-the-art mask-based approaches without requiring additional phase reconstruction steps.

Separating an audio scene into isolated sources is a fundamental problem in computer audition, analogous to image segmentation in visual scene analysis. Source separation systems based on deep learning are currently the most successful approaches for solving the underdetermined separation problem, where there are more sources than channels. Such systems are normally trained on sound mixtures where the ground truth decomposition is already known. In this work, we use an unsupervised spatial source separation on stereo mixtures which generates initial decompositions of mixtures to train a deep learning source separation model. These estimated decompositions vary greatly in quality across the training mixtures. To overcome this, we weight the data during training using a confidence measure that assesses which mixtures or parts of mixtures are well-separated by the unsupervised algorithm. Once trained, the model can be applied to separate single-channel mixtures, where no source direction information is available. The idea is to use simple, low-level processing to separate sources in an unsupervised fashion, identify easy conditions, and then use that knowledge to bootstrap a (self-)supervised source separation model for difficult conditions. We also explore using the two approaches in an ensemble.

Isolating individual instruments in a musical mixture has a myriad of potential applications, and seems imminently achievable given the levels of performance reached by recent deep learning methods. While most musical source separation techniques learn an independent model for each instrument, we propose using a common embedding space for the time-frequency bins of all instruments in a mixture inspired by deep clustering and deep attractor networks. Additionally, an auxiliary network is used to generate parameters of a Gaussian mixture model (GMM) where the posterior distribution over GMM components in the embedding space can be used to create a mask that separates individual sources from a mixture. In addition to outperforming a mask-inference baseline on the MUSDB-18 dataset, our embedding space is easily interpretable and can be used for query-based separation.

Recent progress in separating the speech signals from multiple overlapping speakers using a single audio channel has brought us closer to solving the cocktail party problem. However, most studies in this area use a constrained problem setup, comparing performance when speakers overlap almost completely, at artificially low sampling rates, and with no external background noise. In this paper, we strive to move the field towards more realistic and challenging scenarios. To that end, we created the WSJ0 Hipster Ambient Mixtures (WHAM!) dataset, consisting of two speaker mixtures from the wsj0-2mix dataset combined with real ambient noise samples. The samples were collected in coffee shops, restaurants, and bars in the San Francisco Bay Area, and are made publicly available. We benchmark various speech separation architectures and objective functions to evaluate their robustness to noise. While separation performance decreases as a result of noise, we still observe substantial gains relative to the noisy signals for most approaches.

Recently, there has been growing interest in multi-speaker speech recognition, where the utterances of multiple speakers are recognized from their mixture.Promising techniques have been proposed for this task, but earlier works have required additional training data such as isolated source signals or senone alignments for effective learning.In this paper, we propose a new sequence-to-sequence framework to directly decode multiple label sequences from a single speech sequence by unifying source separation and speech recognition functions in an end-toend manner.We further propose a new objective function to improve the contrast between the hidden vectors to avoid generating similar hypotheses.Experimental results show that the model is directly able to learn a mapping from a speech mixture to multiple label sequences, achieving 83.1% relative improvement compared to a model trained without the proposed objective.Interestingly, the results are comparable to those produced by previous endto-end works featuring explicit separation and recognition modules.

End-to-end automatic speech recognition (ASR) can significantly reduce the burden of developing ASR systems for new languages, by eliminating the need for linguistic information such as pronunciation dictionaries. This also creates an opportunity to build a monolithic multilingual ASR system with a language-independent neural network architecture. In our previous work, we proposed a monolithic neural network architecture that can recognize multiple languages, and showed its effectiveness compared with conventional language-dependent models. However, the model is not guaranteed to properly handle switches in language within an utterance, thus lacking the flexibility to recognize mixed-language speech such as code-switching. In this paper, we extend our model to enable dynamic tracking of the language within an utterance, and propose a training procedure that takes advantage of a newly created mixed-language speech corpus. Experimental results show that the extended model outperforms both language-dependent models and our previous model without suffering from performance degradation that could be associated with language switching.

Current advances in deep learning have resulted in a convergence of methods across a wide range of tasks, opening the door for tighter integration of modules that were previously developed and optimized in isolation. Recent ground-breaking works have produced end-to-end deep network methods for both speech separation and end-to-end automatic speech recognition (ASR). Speech separation methods such as deep clustering address the challenging cocktail-party problem of distinguishing multiple simultaneous speech signals. This is an enabling technology for real-world human machine interaction (HMI). However, speech separation requires ASR to interpret the speech for any HMI task. Likewise, ASR requires speech separation to work in an unconstrained environment. Although these two components can be trained in isolation and connected after the fact, this paradigm is likely to be sub-optimal, since it relies on artificially mixed data. In this paper, we develop the first fully end-to-end, jointly trained deep learning system for separation and recognition of overlapping speech signals. The joint training framework synergistically adapts the separation and recognition to each other. As an additional benefit, it enables training on more realistic data that contains only mixed signals and their transcriptions, and thus is suited to large scale training on existing transcribed data.

The recently-proposed deep clustering algorithm represents a fundamental advance towards solving the cocktail party problem in the single-channel case. When multiple microphones are available, spatial information can be leveraged to differentiate signals from different directions. This study combines spectral and spatial features in a deep clustering framework so that the complementary spectral and spatial information can be simultaneously exploited to improve speech separation. We find that simply encoding inter-microphone phase patterns as additional input features during deep clustering provides a significant improvement in separation performance, even with random microphone array geometry. Experiments on a spatial-ized version of the wsj0-2mix dataset show the strong potential of the proposed algorithm for speech separation in reverberant environments.

The recently proposed deep clustering framework represents a significant step towards solving the cocktail party problem. This study proposes and compares a variety of alternative objective functions for training deep clustering networks. In addition, whereas the original deep clustering work relied on k-means clustering for test-time inference, here we investigate inference methods that are matched to the training objective. Furthermore, we explore the use of an improved chimera network architecture for speech separation, which combines deep clustering with mask-inference networks in a multiobjective training scheme. The deep clustering loss acts as a regularizer while training the end-to-end mask inference network for best separation. With further iterative phase reconstruction, our best proposed method achieves a state-of-the-art 11.5 dB signal-to-distortion ratio (SDR) result on the publicly available wsj0-2mix dataset, with a much simpler architecture than the previous best approach.

This paper proposes an end-to-end approach for single-channel speaker-independent multi-speaker speech separation, where time-frequency (T-F) masking, the short-time Fourier transform (STFT), and its inverse are represented as layers within a deep network.Previous approaches, rather than computing a loss on the reconstructed signal, used a surrogate loss based on the target STFT magnitudes.This ignores reconstruction error introduced by phase inconsistency.In our approach, the loss function is directly defined on the reconstructed signals, which are optimized for best separation.In addition, we train through unfolded iterations of a phase reconstruction algorithm, represented as a series of STFT and inverse STFT layers.While mask values are typically limited to lie between zero and one for approaches using the mixture phase for reconstruction, this limitation is less relevant if the estimated magnitudes are to be used together with phase reconstruction.We thus propose several novel activation functions for the output layer of the T-F masking, to allow mask values beyond one.On the publiclyavailable wsj0-2mix dataset, our approach achieves state-ofthe-art 12.6 dB scale-invariant signal-to-distortion ratio (SI-SDR) and 13.1 dB SDR, revealing new possibilities for deep learning based phase reconstruction and representing a fundamental progress towards solving the notoriously-hard cocktail party problem.

This paper presents a new hybrid approach for polyphonic Sound Event Detection (SED) which incorporates a temporal structure modeling technique based on a hidden Markov model (HMM) with a frame-by-frame detection method based on a bidirectional long short-term memory (BLSTM) recurrent neural network (RNN). The proposed BLSTM-HMM hybrid system makes it possible to model sound event-dependent temporal structures and also to perform sequence-by-sequence detection without having to resort to thresholding such as in the conventional frame-by-frame methods. Furthermore, to effectively reduce insertion errors of sound events, which often occurs under noisy conditions, we additionally implement a binary mask post-processing using a sound activity detection (SAD) network to identify segments with any sound event activity. We conduct an experiment using the DCASE 2016 task 2 dataset to compare our proposed method with typical conventional methods, such as non-negative matrix factorization (NMF) and a standard BLSTM-RNN. Our proposed method outperforms the conventional methods and achieves an F1-score 74.9 % (error rate of 44.7 %) on the event-based evaluation, and an F1-score of 80.5 % (error rate of 33.8 %) on the segment-based evaluation, most of which also outperforms the best reported result in the DCASE 2016 task 2 challenge.

Deep clustering is the first method to handle general audio separation scenarios with multiple sources of the same type and an arbitrary number of sources, performing impressively in speaker-independent speech separation tasks. However, little is known about its effectiveness in other challenging situations such as music source separation. Contrary to conventional networks that directly estimate the source signals, deep clustering generates an embedding for each time-frequency bin, and separates sources by clustering the bins in the embedding space. We show that deep clustering outperforms conventional networks on a singing voice separation task, in both matched and mismatched conditions, even though conventional networks have the advantage of end-to-end training for best signal approximation, presumably because its more flexible objective engenders better regularization. Since the strengths of deep clustering and conventional network architectures appear complementary, we explore combining them in a single hybrid network trained via an approach akin to multi-task learning. Remarkably, the combination significantly outperforms either of its components.

Recent advances in distant-talking ASR research have confirmed that speech enhancement is an essential technique for improving the ASR performance, especially in the multichannel scenario. However, speech enhancement inevitably distorts speech signals, which can cause significant degradation when enhanced signals are used as training data. Thus, distant-talking ASR systems often resort to using the original noisy signals as training data and the enhanced signals only at test time, and give up on taking advantage of enhancement techniques in the training stage. This paper proposes to make use of enhanced features in the student-teacher learning paradigm. The enhanced features are used as input to a teacher network to obtain soft targets, while a student network tries to mimic the teacher network's outputs using the original noisy features as input, so that speech enhancement is implicitly performed within the student network. Compared with conventional student-teacher learning, which uses a better network as teacher, the proposed self-supervised method uses better (enhanced) inputs to a teacher. This setup matches the above scenario of making use of enhanced features in network training. Experiments with the CHiME-4 challenge real dataset show significant ASR improvements with an error reduction rate of 12% in the single-channel track and 15% in the 2-channel track, respectively, by using 6-channel beamformed features for the teacher model.

This paper presents a new hybrid approach called duration-controlled long short-term memory (LSTM) for polyphonic sound event detection (SED). It builds upon a state-of-the-art SED method that performs frame-by-frame detection using a bidirectional LSTM recurrent neural network (BLSTM), and incorporates a duration-controlled modeling technique based on a hidden semi-Markov model. The proposed approach makes it possible to model the duration of each sound event precisely and to perform sequence-by-sequence detection without having to resort to thresholding, as in conventional frame-by-frame methods. Furthermore, to effectively reduce sound event insertion errors, which often occur under noisy conditions, we also introduce a binary-mask-based postprocessing that relies on a sound activity detection network to identify segments with any sound event activity, an approach inspired by the well-known benefits of voice activity detection in speech recognition systems. We conduct an experiment using the DCASE2016 task 2 dataset to compare our proposed method with typical conventional methods, such as nonnegative matrix factorization and standard BLSTM. Our proposed method outperforms the conventional methods both in an event-based evaluation, achieving a 75.3% F1 score and a 44.2% error rate, and in a segment-based evaluation, achieving an 81.1% F1 score, and a 32.9% error rate, outperforming the best results reported in the DCASE2016 task 2 Challenge.

We address the problem of "cocktail-party" source separation in a deep learning framework called deep clustering. Previous deep network approaches to separation have shown promising performance in scenarios with a fixed number of sources, each belonging to a distinct signal class, such as speech and noise. However, for arbitrary source classes and number, "class-based" methods are not suitable. Instead, we train a deep network to assign contrastive embedding vectors to each time-frequency region of the spectrogram in order to implicitly predict the segmentation labels of the target spectrogram from the input mixtures. This yields a deep network-based analogue to spectral clustering, in that the embeddings form a low-rank pair-wise affinity matrix that approximates the ideal affinity matrix, while enabling much faster performance. At test time, the clustering step "decodes" the segmentation implicit in the embeddings by optimizing K-means with respect to the unknown assignments. Preliminary experiments on single-channel mixtures from multiple speakers show that a speaker-independent model trained on two-speaker mixtures can improve signal quality for mixtures of held-out speakers by an average of 6dB. More dramatically, the same model does surprisingly well with three-speaker mixtures.

Deep unfolding has recently been proposed to derive novel deep network architectures from model-based approaches. In this paper, we consider its application to multichannel source separation. We unfold a multichannel Gaussian mixture model (MCGMM), resulting in a deep MCGMM computational network that directly processes complex-valued frequency-domain multichannel audio and has an architecture defined explicitly by a generative model, thus combining the advantages of deep networks and model-based approaches. We further extend the deep MCGMM by modeling the GMM states using an MRF, whose unfolded mean-field inference updates add dynamics across layers. Experiments on source separation for multichannel mixtures of two simultaneous speakers shows that the deep MCGMM leads to improved performance with respect to the original MCGMM model.

Deep clustering is a recently introduced deep learning architecture that uses discriminatively trained embeddings as the basis for clustering. It was recently applied to spectrogram segmentation, resulting in impressive results on speaker-independent multi-speaker separation. In this paper we extend the baseline system with an end-to-end signal approximation objective that greatly improves performance on a challenging speech separation. We first significantly improve upon the baseline system performance by incorporating better regularization, larger temporal context, and a deeper architecture, culminating in an overall improvement in signal to distortion ratio (SDR) of 10.3 dB compared to the baseline of 6.0 dB for two-speaker separation, as well as a 7.1 dB SDR improvement for three-speaker separation. We then extend the model to incorporate an enhancement layer to refine the signal estimates, and perform end-to-end training through both the clustering and enhancement stages to maximize signal fidelity. We evaluate the results using automatic speech recognition. The new signal approximation objective, combined with end-to-end training, produces unprecedented performance, reducing the word error rate (WER) from 89.1% down to 30.8%. This represents a major advancement towards solving the cocktail party problem.

Recurrent neural networks are powerful models for processing sequential data, but they are generally plagued by vanishing and exploding gradient problems. Unitary recurrent neural networks (uRNNs), which use unitary recurrence matrices, have recently been proposed as a means to avoid these issues. However, in previous experiments, the recurrence matrices were restricted to be a product of parameterized unitary matrices, and an open question remains: when does such a parameterization fail to represent all unitary matrices, and how does this restricted representational capacity limit what can be learned? To address this question, we propose full-capacity uRNNs that optimize their recurrence matrix over all unitary matrices, leading to significantly improved performance over uRNNs that use a restricted-capacity recurrence matrix. Our contribution consists of two main components. First, we provide a theoretical argument to determine if a unitary parameterization has restricted capacity. Using this argument, we show that a recently proposed unitary parameterization has restricted capacity for hidden state dimension greater than 7. Second,we show how a complete, full-capacity unitary recurrence matrix can be optimized over the differentiable manifold of unitary matrices. The resulting multiplicative gradient step is very simple and does not require gradient clipping or learning rate adaptation. We confirm the utility of our claims by empirically evaluating our new full-capacity uRNNs on both synthetic and natural data, achieving superior performance compared to both LSTMs and the original restricted-capacity uRNNs.

We present an advanced dialog state tracking system designed for the 5th Dialog State Tracking Challenge (DSTC5). The main task of DSTC5 is to track the dialog state in a human-human dialog. For each utterance, the tracker emits a frame of slot-value pairs considering the full history of the dialog up to the current turn. Our system includes an encoder-decoder architecture with an attention mechanism to map an input word sequence to a set of semantic labels, i.e., slot-value pairs. This handles the problem of the unknown alignment between the utterances and the labels. By combining the attention-based tracker with rule-based trackers elaborated for English and Chinese, the F-score for the development set improved from 0.475 to 0.507 compared to the rule-only trackers. Moreover, we achieved 0.517 F-score by refining the combination strategy based on the topic and slot level performance of each tracker. In this paper, we also validate the efficacy of each technique and report the test set results submitted to the challenge.

This paper introduces the MERL/SRI system designed for the 3rd CHiME speech separation and recognition challenge (CHiME-3). Our proposed system takes advantage of recurrent neural networks (RNNs) throughout the model from the front speech enhancement to the language modeling. Two different types of beamforming are used to combine multi-microphone signals to obtain a single higher quality signal. Beamformed signal is further processed by a single-channel bi-directional long short-term memory (LSTM) enhancement network which is used to extract stacked mel-frequency cepstral coefficients (MFCC) features. In addition, two proposed noise-robust feature extraction methods are used with the beamformed signal. The features are used for decoding in speech recognition systems with deep neural network (DNN) based acoustic models and large-scale RNN language models to achieve high recognition accuracy in noisy environments. Our training methodology includes data augmentation and speaker adaptive training, whereas at test time model combination is used to improve generalization. Results on the CHiME-3 benchmark show that the full cadre of techniques substantially reduced the word error rate (WER). Combining hypotheses from different robust-feature systems ultimately achieved 9.10% WER for the real test data, a 72.4% reduction relative to the baseline of 32.99% WER.

Separation of speech embedded in non-stationary interference is a challenging problem that has recently seen dramatic improvements using deep network-based methods. Previous work has shown that estimating a masking function to be applied to the noisy spectrum is a viable approach that can be improved by using a signal-approximation based objective function. Better modeling of dynamics through deep recurrent networks has also been shown to improve performance. Here we pursue both of these directions. We develop a phase-sensitive objective function based on the signal-to-noise ratio (SNR) of the reconstructed signal, and show that in experiments it yields uniformly better results in terms of signal-to-distortion ratio (SDR). We also investigate improvements to the modeling of dynamics, using bidirectional recurrent networks, as well as by incorporating speech recognition outputs in the form of alignment vectors concatenated with the spectral input features. Both methods yield further improvements, pointing to tighter integration of recognition with separation as a promising future direction.

Non-negative matrix factorization (NMF) has been widely used for challenging single-channel audio source separation tasks. However, inference in NMF-based models relies on iterative inference methods, typically formulated as multiplicative updates. We propose “deep NMF”, a novel non-negative deep network architecture which results from unfolding the NMF iterations and untying its parameters. This architecture can be discriminatively trained for optimal separation performance. To optimize its non-negative parameters, we show how a new form of back-propagation, based on multiplicative updates, can be used to preserve non-negativity, without the need for constrained optimization. We show on a challenging speech separation task that deep NMF improves in terms of accuracy upon NMF and is competitive with conventional sigmoid deep neural networks, while requiring a tenth of the number of parameters.

Speech and audio signal processing research is a tale of data collection efforts and evaluation campaigns. Large benchmark datasets for automatic speech recognition (ASR) have been instrumental in the advancement of speech recognition technologies. However, when it comes to robust ASR, source separation, and localization, especially using microphone arrays, the perfect dataset is out of reach, and many different data collection efforts have each made different compromises between the conflicting factors in terms of realism, ground truth, and costs. Our goal here is to escape some of the most difficult trade-offs by proposing MICbots, a low-cost method of collecting large amounts of realistic data where annotations and ground truth are readily available. Our key idea is to use freely moving robots equiped with microphones and loudspeakers, playing recorded utterances from existing (already annotated) speech datasets. We give an overview of previous data collection efforts and the trade-offs they make, and describe the benefits of using our robot-based approach. We finally explain the use of this method to collect room impulse response measurement.

Model-based speech enhancement methods, which rely on separately modeling the speech and the noise, have been shown to be powerful in many different problem settings. When the structure of the noise can be arbitrary, which is often the case in practice, modelbased methods have to focus on developing good speech models, whose quality will be key to their performance. In this study, we propose a novel probabilistic model for speech enhancement which precisely models the speech by taking into account the underlying speech production process as well as its dynamics. The proposed model follows a source-filter approach where the excitation and filter parts are modeled as non-negative dynamical systems. We present convergence-guaranteed update rules for each latent factor. In order to assess performance, we evaluate our model on a challenging speech enhancement task where the speech is observed under non-stationary noises recorded in a car. We show that our model outperforms state-of-the-art methods in terms of objective measures.

State-of-the-art automatic speech recognition (ASR) systems are very complex, combining multiple techniques and involving many types of tuning parameters (e.g., numbers of states and Gaussians in HMMs, numbers of neurons/layers and learning rates in neural networks, etc.). To reach optimal performance in such systems, deep understanding and expertise of each component is necessary, thus limiting the development of ASR systems to skilled experts. To overcome the problem, this paper studies the use of black box optimization, which automatically tunes systems without any prior knowledge. We consider an ASR system as a function with tuning parameters as input and speech recognition performance (e.g., word accuracy) as output, and we investigate two probabilistic black box optimization techniques: Covariance Mean Adaptation Evolution Strategy (CMA-ES) and Bayesian optimization using Gaussian process. Middle-vocabulary speech recognition experiments show the effectiveness of black box optimization, as performance approaching that of fine-tuned systems obtained by experts and/or outperforming that of sub-optimal systems can be automatically obtained.

Linear discriminant analysis (LDA) is a simple and effective feature transformation technique that aims to improve discrim-inability by maximizing the ratio of the between-class variance to the within-class variance. However, LDA does not explicitly consider the sequential discriminative criterion which consists in directly reducing the errors of a speech recognizer. This pa-per proposes a simple extension of LDA that is called sequen-tial LDA (sLDA) based on a sequential discriminative criterion computed from the Gaussian statistics, which are obtained from sequential maximum mutual information (MMI) training. Al-though the objective function of the proposed LDA can be re-garded as a special case of various discriminative feature trans-formation techniques (for example, f-MPE or the bottom layer of a neural network), the transformation matrix can be ob-tained as the closed-form solution to a generalized eigenvalue problem, in contrast to the gradient-descent-based optimization methods usually used in these techniques. Experiments on large vocabulary continuous speech recognition (Corpus of Sponta-neous Japanese) and noisy speech recognition task (2nd CHiME challenge) show consistent improvements from standard LDA due to the sequential discriminative training. In addition, the proposed method, despite its simple and fast computation, im-proved the performance in combination with discriminative fea-ture transformation (f-bMMI), perhaps by providing a good ini-tialization to f-bMMI. Index Terms: Maximum mutual information, linear dis-criminant analysis, region dependent linear transformation

The objective of single-channel source separation is to accurately recover source signals from mixtures. Non-negative matrix fac-torization (NMF) is a popular approach for this task, yet previous NMF approaches have not optimized directly this objective, de-spite some efforts in this direction. Our paper introduces discrim-inative training of the NMF basis functions such that, given the coefficients obtained on a mixture, a desired source is optimally recovered. We approach this optimization by generalizing the model to have separate analysis and reconstruction basis func-tions. This generalization frees us to optimize reconstruction ob-jectives that incorporate the filtering step and SNR performance criteria. A novel multiplicative update algorithm is presented for the optimization of the reconstruction basis functions accord-ing to the proposed discriminative objective functions. Results

This paper proposes a generalized discriminative training framework for system combination, which encompasses acoustic modeling (Gaussian mixture models and deep neural networks) and discriminative feature transformation. To improve the performance by combining base systems with complementary systems, complementary systems should have reasonably good performance while tending to have different outputs compared with the base system. Although it is difficult to balance these two somewhat opposite targets in conventional heuristic combination approaches, our framework provides a new objective function that enables to adjust the balance within a sequential discriminative training criterion. We also describe how the proposed method relates to boosting methods. Experiments on highly noisy middle vocabulary speech recognition task (2nd CHiME challenge track 2) and LVCSR task (Corpus of Spontaneous Japanese) show the effectiveness of the proposed method, compared with a conventional system combination approach.

Distant-microphone automatic speech recognition (ASR) remains a challenging goal in everyday environments involving multiple background sources and reverberation. This paper reports on the results of the 2nd ‘CHiME’ Challenge, an initiative designed to analyse and evaluate the performance of ASR systems in a real-world domestic environment. We discuss the rationale for the challenge and provide a summary of the datasets, tasks and baseline systems. The paper overviews the systems that were entered for the two challenge tracks: small-vocabulary with moving talker and medium-vocabulary with stationary talker. We present a summary of the challenge findings including novel results produced by challenge system combination. Possible directions for future challenges are discussed.

Non-negative data arise in a variety of important signal processing domains, such as power spectra of signals, pixels in images, and count data. This paper introduces a novel non-negative dynamical system (NDS) for sequences of such data, and describes its application to modeling speech and audio power spectra. The NDS model can be interpreted both as an adaptation of linear dynamical systems (LDS) to non-negative data, and as an extension of non-negative matrix factorization (NMF) to support Markovian dynamics. Learning and inference algorithms were derived and experiments on speech enhancement were conducted by training sparse non-negative dynamical systems on speech data and adapting a noise model to the unknown noise condition. Results show that the model can capture the dynamics of speech in a useful way.

In this paper, we demonstrate that recently-developed sparse recovery algorithms can be used to improve source localization in reverberant environments. By formulating the localization problem in the frequency domain, we are able to efficiently incorporate information that exploits the reverberation instead of considering it a nuisance to be eliminated. In this formulation, localization becomes a joint-sparsity support recovery problem which can be solved using model-based methods. We also develop a location model which further improves performance. Using our approach, we are able to recover more sources that the number of sensors. In contrast to conventional wisdom, we demonstrate that reverberation is beneficial in source localization, as long as it known and properly accounted for.

Distant-microphone automatic speech recognition (ASR) remains a challenging goal in everyday environments involving multiple background sources and reverberation. This paper is intended to be a reference on the 2nd `CHiME' Challenge, an initiative designed to analyze and evaluate the performance of ASR systems in a real-world domestic environment. Two separate tracks have been proposed: a small-vocabulary task with small speaker movements and a medium-vocabulary task without speaker movements. We discuss the rationale for the challenge and provide a detailed description of the datasets, tasks and baseline performance results for each track.

Wiener filtering is one of the most ubiquitous tools in signal processing, in particular for signal denoising and source separation. In the context of audio, it is typically applied in the time-frequency domain by means of the short-time Fourier transform (STFT). Such processing does generally not take into account the relationship between STFT coefficients in different time-frequency bins due to the redundancy of the STFT, which we refer to as consistency. We propose to enforce this relationship in the design of the Wiener filter, either as a hard constraint or as a soft penalty. We derive two conjugate gradient algorithms for the computation of the filter coefficients and show improved audio source separation performance compared to the classical Wiener filter both in oracle and in blind conditions.

Model-based speech enhancement methods, such as vector-Taylor series-based methods (VTS) [1, 2], share a common methodology: they estimate speech using the expected value of the clean speech given the noisy speech under a statistical model. We show that it may be better to use the expected value of the noise under the model and subtract it from the noisy observation to form an indirect estimate of the speech. Interestingly, for VTS, this methodology turns out to be related to the application of an SNR-dependent gain to the direct VTS speech estimate. In results obtained on an automotive noise task, this methodology produces an average improvement of 1.6 dB signal-to-noise ratio (SNR), relative to conventional methods.

This paper presents a nonparametric Bayesian extension of non-negative matrix factorization (NMF) for music signal analysis. Instrument sounds often exhibit non-stationary spectral characteristics. We introduce infinite-state spectral bases into NMF to represent time-varying spectra in polyphonic music signals. We describe our extension of NMF with infinite-state spectral bases generated by the Dirichlet process in a statistical framework, derive an efficient optimization algorithm based on collapsed variational inference, and validate the framework on audio data.

Newborns must learn to structure incoming acoustic information into segments, words, phrases, etc., before they can start to learn language. This process is thought to rely on modulation structure of the speech waveform induced by segmental or prosodic regularities within the speech heard by the infant. Here, we investigate the process by which the initial acoustic processing required by modulation analysis can itself be tuned by exposure to the regularities of speech. Starting from the classic definition of modulation, as applied within channels of the peripheral filter, we formulate a mathematical framework in which the structure of initial spectral filtering is adapted for modulation analysis. Our working hypothesis is that the human ear and brain are adapted to the analysis of modulation, via a data-driven learning process on the scale of development (or possibly evolution). Simulation results are presented and a comparison with filterbanks classically used in signal processing is done.

How does one extract unknown but stereotypical events that are linearly superimposed within a signal with variable latencies and variable amplitudes? One could think of using template matching or matching pursuit to find the arbitrarily shifted linear components. However, traditional matching approaches require that the templates be known a priori. To overcome this restriction we use instead semi Non-Negative Matrix Factorization (semi-NMF) that we extend to allow for time shifts when matching the templates to the signal. The algorithm estimates templates directly from the data along with their non-negative amplitudes. The resulting method can be thought of as an adaptive template matching procedure. We demonstrate the procedure on the task of extracting spikes from single channel extracellular recordings. On these data the algorithm essentially performs spike detection and unsupervised spike clustering. Results on simulated data and extracellular recordings indicate that the method performs well for signal-to-noise ratios of 6dB or higher and that spike templates are recovered accurately provided they are sufficiently different.

We present a method for removing environmental noise from physiological recordings such as magnetoencephalography (MEG) for which noise-sensitive reference channels are available. Sensor signals are projected on a subspace spanned by the reference channels augmented by time-shifted and/or nonlinearly transformed versions of the same, and the projections are removed to obtain "clean" sensor signals. The method compensates for scalar, convolutional or non-linear mismatches between sensor and reference channels by synthesizing, for each reference/sensor pair, a filter that is optimal in a least-squares sense for removal of the artifact. The method was tested with synthetic and real MEG data, typically removing up to 98% of noise variance. It offers an alternative to bulky and costly magnetic shielding (multiple layers of aluminium and mu-metal) for present scientific and medical applications and future developments such as brain-machine interfaces (BMI).

We present in this paper a novel F0contour estimation method based on a parametric description of the wavelet power spectrum of speech that accounts for its structure simultaneously in time and frequency directions. We model the speech spectrum as a sequence of spectral clusters governed by a smooth common F0contour expressed as a spline curve. The harmonic and temporal structure of these clusters and their common F0contour are estimated simultaneously. Through experimental comparisons with existing methods, we show that our algorithm is competitive on clean single-speaker speech, and that it outperforms existing methods both in the presence of noise and for the estimation of multiple F0contours of cochannel concurrent speech.

The minimum classification error (MCE) framework for discriminative training is a simple and general formalism for directly optimizing recognition accuracy in pattern recognition problems. The framework applies directly to the optimization of hidden Markov models (HMMs) used for speech recognition problems. However, few if any studies have reported results for the application of MCE training to large-vocabulary, continuous-speech recognition tasks. This article reports significant gains in recognition performance and model compactness as a result of discriminative training based on MCE training applied to HMMs, in the context of three challenging large-vocabulary (up to 100 k word) speech recognition tasks: the Corpus of Spontaneous Japanese lecture speech transcription task, a telephone-based name recognition task, and the MIT Jupiter telephone-based conversational weather information task. On these tasks, starting from maximum likelihood (ML) baselines, MCE training yielded relative reductions in word error ranging from 7% to 20%. Furthermore, this paper evaluates the use of different methods for optimizing the MCE criterion function, as well as the use of precomputed recognition lattices to speed up training. An overview of the MCE framework is given, with an emphasis on practical implementation issues

This paper proposes a novel $F_{0}$ contour estimation algorithm based on a precise parametric description of the voiced parts of speech derived from the power spectrum. The algorithm is able to perform in a wide variety of noisy environments as well as to estimate the $F_{0}$ s of cochannel concurrent speech. The speech spectrum is modeled as a sequence of spectral clusters governed by a common $F_{0}$ contour expressed as a spline curve. These clusters are obtained by an unsupervised 2-D time-frequency clustering of the power density using a new formulation of the EM algorithm, and their common $F_{0}$ contour is estimated at the same time. A smooth $F_{0}$ contour is extracted for the whole utterance, linking together its voiced parts. A noise model is used to cope with nonharmonic background noise, which would otherwise interfere with the clustering of the harmonic portions of speech. We evaluate our algorithm in comparison with existing methods on several tasks, and show 1) that it is competitive on clean single-speaker speech, 2) that it outperforms existing methods in the presence of noise, and 3) that it outperforms existing methods for the estimation of multiple $F_{0}$ contours of cochannel concurrent speech.

We have been working on a new speech analyzer based on a parametric representation of speech governed by the F0 parameter, towards practical human-machine interfaces. As a precise estimation of the frequency response of the vocal tract from a real speech signal requires the power of each component of the harmonic structure to be accurately estimated, one hopes to have a high-precision estimation of F0. At the same time, under the empirical constraint that speech spectral envelopes are usually smooth in the power domain, half pitch errors can be significantly avoided. Therefore, F0 and the envelope should be estimated jointly rather than separately through an optimal estimation of the spectral envelope and the spectral fine structure. In this article, we introduce a new speech analysis method using a spectral model with a composite function of envelope and fine structure models. Index Terms: parametric speech analyzer, speech synthesis, pitch estimation, spectral envelope estimation.

Discriminative training applied to hidden Markov model (HMM) design can yield significant benefits in recognition accuracy and model compactness. However, compared to Maximum Likelihood based methods, discriminative training typically requires much more computation, as all competing candidates must be considered, not just the correct one. The choice of the algorithm used to optimize the discriminative criterion function is thus a key issue. We investigated several algorithms and used them for discriminative training based on the Minimum Classification Error (MCE) framework. In particular, we examined on-line, batch, and semi-batch Probabilistic Descent (PD), as well as Quickprop, Rprop and BFGS.We describe each algorithm and present comparative results on the TIMIT phone classification task and on the 230 hour Corpus of Spontaneous Japanese (CSJ) 30K word continuous speech recognition task.