Token-level serialized output training (t-SOT) was recently proposed to address the challenge of streaming multi-talker automatic speech recognition (ASR). T-SOT effectively handles overlapped speech by representing multi-talker transcriptions as a single token stream with ⟨cc⟩ symbols interspersed. However, the use of a naive neural transducer architecture significantly constrained its applicability for text-only adaptation. To overcome this limitation, we propose a novel t-SOT model structure that incorporates the idea of factorized neural transducers (FNT). The proposed method separates a language model (LM) from the transducer’s predictor and handles the unnatural token order resulting from the use of ⟨cc⟩ symbols in t-SOT. We achieve this by maintaining multiple hidden states and introducing special handling of the ⟨cc⟩ tokens within the LM. The proposed t-SOT FNT model achieves comparable performance to the original t-SOT model while retaining the ability to reduce word error rate (WER) on both single and multi-talker datasets through text-only adaptation.

With the success of the first Multi-channel Multi-party Meeting Transcription challenge (M2MeT), the second M2MeT challenge (M2MeT 2.0) held in ASRU2023 particularly aims to tackle the complex task of speaker-attributed ASR (SAASR), which directly addresses the practical and challenging problem of “who spoke what at when” at typical meeting scenario. We particularly established two sub-tracks. The fixed training condition sub-track, where the training data is constrained to predetermined datasets, but participants can use any open-source pre-trained model. The open training condition sub-track, which allows for the use of all available data and models without limitation. In addition, we release a new 10-hour test set for challenge ranking. This paper provides an overview of the dataset, track settings, results, and analysis of submitted systems, as a benchmark to show the current state of speaker-attributed ASR.

Large language models (LLMs) have achieved remarkable success in the field of natural language processing, enabling better human-computer interaction using natural language. However, the seamless integration of speech signals into LLMs has not been explored well. The “decoder-only“ architecture has also not been well studied for speech processing tasks. In this research, we introduce Speech-LLaMA, a novel approach that effectively incorporates acoustic information into text-based large language models. Our method leverages Connectionist Temporal Classification and a simple audio encoder to map the compressed acoustic features to the continuous semantic space of the LLM. In addition, we further probe the decoder-only architecture for speech-to-text tasks by training a smaller scale randomly initialized speech-LLaMA model from speech-text paired data alone. We conduct experiments on multilingual speech-to-text translation tasks and demonstrate a significant improvement over strong baselines, highlighting the potential advantages of decoder-only models for speech-to-text conversion.

Self-supervised learning (SSL) methods such as WavLM have shown promising speech separation (SS) results in small-scale simulation-based experiments. In this work, we extend the exploration of the SSL-based SS by massively scaling up both the pre-training data (more than 300K hours) and fine-tuning data (10K hours). We also investigate various techniques to efficiently integrate the pre-trained model with the SS network under a limited computation budget, including a low frame rate SSL model training setup and a fine-tuning scheme using only the part of the pre-trained model. Compared with a supervised baseline and the WavLM-based SS model using feature embeddings obtained with the previously released 94K hours trained WavLM, our proposed model obtains 15.9% and 11.2% of relative word error rate (WER) reductions, respectively, for a simulated far-field speech mixture test set. For conversation transcription on real meeting recordings using continuous speech separation, the proposed model achieves 6.8% and 10.6% of relative WER reductions over the purely supervised baseline on AMI and ICSI evaluation sets, respectively, while reducing the computational cost by 38%.

Self-supervised learning (SSL), which utilizes the input data itself for representation learning, has achieved state-of-the-art results for various downstream speech tasks. However, most of the previous studies focused on offline single-talker applications, with limited investigations in multi-talker cases, especially for streaming scenarios. In this paper, we investigate SSL for streaming multi-talker speech recognition, which generates transcriptions of overlapping speakers in a streaming fashion. Firstly, we observe that conventional SSL techniques do not work well on this task due to the poor representation of overlapping speech. We then propose a novel SSL training objective, referred to as bi-label masked speech prediction, which explicitly preserves representations of all speakers in overlapping speech. We investigate various aspects of the proposed system, including data configuration and quantizer selection. The proposed SSL setup achieves substantially better word error rates on the LibriSpeechMix dataset.

This paper presents a novel streaming automatic speech recognition (ASR) framework for multi-talker overlapping speech captured by a distant microphone array with an arbitrary geometry. Our framework, named t-SOT-VA, capitalizes on independently developed two recent technologies; array-geometry-agnostic continuous speech separation, or VarArray, and streaming multi-talker ASR based on token-level serialized output training (t-SOT). To combine the best of both technologies, we newly design a t-SOT-based ASR model that generates a serialized multi-talker transcription based on two separated speech signals from VarArray. We also propose a pre-training scheme for such an ASR model where we simulate VarArray’s output signals based on monaural single-talker ASR training data. Conversation transcription experiments using the AMI meeting corpus show that the system based on the proposed framework significantly outperforms conventional ones. Our system achieves the state-of-the-art word error rates of 13.7% and 15.5% for the AMI development and evaluation sets, respectively, in the multiple-distant-microphone setting while retaining the streaming inference capability.

Recently, Transformer-based architectures have been explored for speaker embedding extraction. Although the Transformer employs the self-attention mechanism to efficiently model the global interaction between token embeddings, it is inadequate for capturing short-range local context, which is essential for the accurate extraction of speaker information. In this study, we enhance the Transformer with the enhanced locality modeling in two directions. First, we propose the Locality-Enhanced Conformer (LE-Confomer) by introducing depth-wise convolution and channel-wise attention into the Conformer blocks. Second, we present the Speaker Swin Transformer (SST) by adapting the Swin Transformer, originally proposed for vision tasks, into speaker embedding network. We evaluate the proposed approaches on the VoxCeleb datasets and a large-scale Microsoft internal multilingual (MS-internal) dataset. The proposed models achieve 0.75% EER on VoxCeleb 1 test set, outperforming the previously proposed Transformer-based models and CNN-based models, such as ResNet34 and ECAPA-TDNN. When trained on the MS-internal dataset, the proposed models achieve promising results with 14.6% relative reduction in EER over the Res2Net50 model.

This paper describes a speaker diarization model based on target speaker voice activity detection (TS-VAD) using transformers. To overcome the original TS-VAD model’s drawback of being unable to handle an arbitrary number of speakers, we investigate model architectures that use input tensors with variable-length time and speaker dimensions. Transformer layers are applied to the speaker axis to make the model output insensitive to the order of the speaker profiles provided to the TS-VAD model. Time-wise sequential layers are interspersed between these speaker-wise transformer layers to allow the temporal and cross-speaker correlations of the input speech signal to be captured. We also extend a diarization model based on end-to-end neural diarization with encoder-decoder based attractors (EEND-EDA) by replacing its dot-product-based speaker detection layer with the transformer-based TS-VAD. Experimental results on VoxConverse show that using the transformers for the cross-speaker modeling reduces the diarization error rate (DER) of TS-VAD by 11.3%, achieving a new state-of-the-art (SOTA) DER of 4.57%. Also, our extended EEND-EDA reduces DER by 6.9% on the CALLHOME dataset relative to the original EEND-EDA with a similar model size, achieving a new SOTA DER of 11.18% under a widely used training data setting.

Speaker change detection (SCD) is an important feature that improves the readability of the recognized words from an automatic speech recognition (ASR) system by breaking the word sequence into paragraphs at speaker change points. Existing SCD solutions either require additional ensemble for the time based decisions and recognized word sequences, or implement a tight integration between ASR and SCD, limiting the potential optimum performance for both tasks. To address these issues, we propose a novel framework for the SCD task, where an additional SCD module is built on top of an existing Transformer Transducer ASR (TT-ASR) network. Two variants of the SCD network are explored in this framework that naturally estimate speaker change probability for each word, while allowing the ASR and SCD to have independent optimization scheme for the best performance. Experiments show that our methods can significantly improve the F1 score on LibriCSS and Microsoft call center data sets without ASR degradation, compared with a joint SCD and ASR baseline.

Multi-talker automatic speech recognition (ASR) has been studied to generate transcriptions of natural conversation including over-lapping speech of multiple speakers. Due to the difficulty in acquiring real conversation data with high-quality human transcriptions, a naïve simulation of multi-talker speech by randomly mixing multiple utterances was conventionally used for model training. In this work, we propose an improved technique to simulate multi-talker overlap-ping speech with realistic speech overlaps, where an arbitrary pattern of speech overlaps is represented by a sequence of discrete tokens. With this representation, speech overlapping patterns can be learned from real conversations based on a statistical language model, such as N-gram, which can be then used to generate multi-talker speech for training. In our experiments, multi-talker ASR models trained with the proposed method show consistent improvement on the word error rates across multiple datasets.

Self-supervised learning (SSL) is a long-standing goal for speech processing, since it utilizes large-scale unlabeled data and avoids extensive human labeling. Recent years have witnessed great successes in applying self-supervised learning in speech recognition, while limited exploration was attempted in applying SSL for modeling speaker characteristics. In this paper, we aim to improve the existing SSL framework for speaker representation learning. Two methods are introduced for enhancing the unsupervised speaker information extraction. First, we apply multi-task learning to the current SSL framework, where we integrate utterance-wise contrastive loss with the SSL objective function. Second, for better speaker discrimination, we propose an utterance mixing strategy for data augmentation, where additional overlapped utterances are created unsupervisely and incorporated during training. We integrate the proposed methods into the HuBERT framework. Experiment results on the SUPERB benchmark show that the proposed system achieves state-of-the-art performance in universal representation learning, especially for speaker identification oriented tasks. An ablation study is performed verifying the efficacy of each proposed method. Finally, we scale up the training dataset to 94 thousand hours of public audio data and achieve further performance improvement in all SUPERB tasks.

An Xception model reaches state-of-the-art (SOTA) accuracy on the ESC-50 dataset for audio event detection through knowledge transfer from ImageNet weights, pretraining on AudioSet, and an on-the-fly data augmentation pipeline. This paper presents an ablation study that analyzes which components contribute to the boost in performance and training time. A smaller Xception model is also presented which nears SOTA performance with almost a third of the parameters.

While permutation invariant training (PIT) based continuous speech separation (CSS) significantly improves the conversation transcription accuracy, it often suffers from speech leakages and failures in separation at "hot spot" regions because it has a fixed number of output channels. In this paper, we propose to apply recurrent selective attention network (RSAN) to CSS, which generates a variable number of output channels based on active speaker counting. In addition, we propose a novel block-wise dependency extension of RSAN by introducing dependencies between adjacent processing blocks in the CSS framework. It enables the network to utilize the separation results from the previous blocks to facilitate the current block processing. Experimental results on the LibriCSS dataset show that the RSAN-based CSS (RSAN-CSS) network consistently improves the speech recognition accuracy over PIT-based models. The proposed block-wise dependency modeling further boosts the performance of RSAN-CSS.

Recently, self-supervised learning (SSL) has demonstrated strong performance in speaker recognition, even if the pretraining objective is designed for speech recognition.In this paper, we study which factor leads to the success of selfsupervised learning on speaker-related tasks, e.g.speaker verification (SV), through a series of carefully designed experiments.Our empirical results on the Voxceleb-1 dataset suggest that the benefit of SSL to SV task is from a combination of mask speech prediction loss, data scale, and model size, while the SSL quantizer has a minor impact.We further employ the integrated gradients attribution method and loss landscape visualization to understand the effectiveness of self-supervised learning for speaker recognition performance.

This paper presents a streaming speaker-attributed automatic speech recognition (SA-ASR) model that can recognize "who spoke what" with low latency even when multiple people are speaking simultaneously.Our model is based on token-level serialized output training (t-SOT) which was recently proposed to transcribe multi-talker speech in a streaming fashion.To further recognize speaker identities, we propose an encoderdecoder based speaker embedding extractor that can estimate a speaker representation for each recognized token not only from non-overlapping speech but also from overlapping speech.The proposed speaker embedding, named t-vector, is extracted synchronously with the t-SOT ASR model, enabling joint execution of speaker identification (SID) or speaker diarization (SD) with the multi-talker transcription with low latency.We evaluate the proposed model for a joint task of ASR and SID/SD by using LibriSpeechMix and LibriCSS corpora.The proposed model achieves substantially better accuracy than a prior streaming model and shows comparable or sometimes even superior results to the state-of-the-art offline SA-ASR model.

This paper proposes a token-level serialized output training (t-SOT), a novel framework for streaming multi-talker automatic speech recognition (ASR). Unlike existing streaming multi-talker ASR models using multiple output branches, the t-SOT model has only a single output branch that generates recognition tokens (e.g., words, subwords) of multiple speakers in chronological order based on their emission times. A special token that indicates the change of ``virtual'' output channels is introduced to keep track of the overlapping utterances. Compared to the prior streaming multi-talker ASR models, the t-SOT model has the advantages of less inference cost and a simpler model architecture. Moreover, in our experiments with LibriSpeechMix and LibriCSS datasets, the t-SOT-based transformer transducer model achieves the state-of-the-art word error rates by a significant margin to the prior results. For non-overlapping speech, the t-SOT model is on par with a single-talker ASR model in terms of both accuracy and computational cost, opening the door for deploying one model for both single- and multi-talker scenarios.

Speaker-attributed automatic speech recognition (SA-ASR) is a task to recognize “who spoke what” from multi-talker recordings. An SA-ASR system usually consists of multiple modules such as speech separation, speaker diarization and ASR. On the other hand, considering the joint optimization, an end-to-end (E2E) SA-ASR model has recently been proposed with promising results on simulation data. In this paper, we present our recent study on the comparison of such modular and joint approaches towards SA-ASR on real monaural recordings. We develop state-of-the-art SA-ASR systems for both modular and joint approaches by leveraging large-scale training data, including 75 thousand hours of ASR training data and the VoxCeleb corpus for speaker representation learning. We also propose a new pipeline that performs the E2E SA-ASR model after speaker clustering. Our evaluation on the AMI meeting corpus reveals that after fine-tuning with a small real data, the joint system performs 8.9-29.9% better in accuracy compared to the best modular system while the modular system performs better before such fine-tuning. We also conduct various error analyses to show the remaining issues for the monaural SA-ASR.

Continuous speech separation was recently proposed to deal with the overlapped speech in natural conversations. While it was shown to significantly improve the speech recognition performance for multichannel conversation transcription, its effectiveness has yet to be proven for a single-channel recording scenario. This paper examines the use of Conformer architecture in lieu of recurrent neural networks for the separation model. Conformer allows the separation model to efficiently capture both local and global context information, which is helpful for speech separation. Experimental results using the LibriCSS dataset show that the Conformer separation model achieves the state of the art results for both single-channel and multi-channel settings. Results for real meeting recordings are also presented, showing significant performance gains in both word error rate (WER) and speaker-attributed WER.

In the presence of a wide variety of dialects, training dialect-specific models for each dialect is a demanding task. Previous studies have explored training a single model that is robust across multiple dialects. These studies have used either multi-condition training, multi-task learning, end-to-end modeling, or ensemble modeling. In this study, we further explore using a single model for multi-dialect speech recognition using ensemble modeling. First, we build an ensemble of dialect-specific models (or experts). Then we linearly combine the outputs of the experts using attention weights generated by a long short-term memory (LSTM) network. For comparison purposes, we train a model that jointly learns to recognize and classify dialects using multi-task learning and a second model using multi-condition training. We train all of these models with about 60,000 hours of speech data collected in American English, Canadian English, British English, and Australian English. Experimental results reveal that our best proposed model achieved an average 4.74% word error rate reduction (WERR) compared to the strong baseline model.

This paper describes the Microsoft speaker diarization system for monaural multi-talker recordings in the wild, evaluated at the diarization track of the VoxCeleb Speaker Recognition Challenge (VoxSRC) 2020. We will first explain our system design to address issues in handling real multi-talker recordings. We then present the details of the components, which include Res2Net-based speaker embedding extractor, conformer-based continuous speech separation with leakage filtering, and a modified DOVER (short for Diarization Output Voting Error Reduction) method for system fusion. We evaluate the systems with the data set provided by VoxSRC challenge 2020, which contains real-life multi-talker audio collected from YouTube. Our best system achieves 3.71% and 6.23% of the diarization error rate (DER) on development set and evaluation set, respectively, being ranked the 1st at the diarization track of the challenge.

Scores from traditional confidence classifiers (CCs) in automatic speech recognition (ASR) systems lack universal interpretation and vary with updates to the underlying confidence or acoustic models (AMs).In this work, we build interpretable confidence scores with an objective to closely align with ASR accuracy.We propose a new sequence-level CC with a richer context providing CC scores highly correlated with ASR accuracy and scores stable across CC updates.Hence, expanding CC applications.Recently, AM customization has gained traction with the widespread use of unified models.Conventional adaptation strategies that customize AM expect wellmatched data for the target domain with gold-standard transcriptions.We propose a cost-effective method of using CC scores to select an optimal adaptation data set, where we maximize ASR gains from minimal data.We study data in various confidence ranges and optimally choose data for AM adaptation with KL-Divergence regularization.On the Microsoft voice search task, data selection for supervised adaptation using the sequence-level confidence scores achieves word error rate reduction (WERR) of 8.5% for row-convolution LSTM (RC-LSTM) and 5.2% for latency-controlled bidirectional LSTM (LC-BLSTM).In the semi-supervised case, with ASR hypotheses as labels, our method provides WERR of 5.9% and 2.8% for RC-LSTM and LC-BLSTM, respectively.

Transformer has been successfully applied to speech separation recently with its strong long-dependency modeling capacity using a self-attention mechanism. However, Transformer tends to have heavy run-time costs due to the deep encoder layers, which hinders its deployment on edge devices. A small Transformer model with fewer encoder layers is preferred for computational efficiency, but it is prone to performance degradation. In this paper, an ultra fast speech separation Transformer model is proposed to achieve both better performance and efficiency with teacher student learning (T-S learning). We introduce layer-wise T-S learning and objective shifting mechanisms to guide the small student model to learn intermediate representations from the large teacher model. Compared with the small Transformer model trained from scratch, the proposed T-S learning method reduces the word error rate (WER) by more than 5% for both multi-channel and single-channel speech separation on LibriCSS dataset. Utilizing more unlabeled speech data, our ultra fast speech separation models achieve more than 10% relative WER reduction.

In this paper, we present AISHELL-4, a sizable real-recorded Mandarin speech dataset collected by 8-channel circular microphone array for speech processing in conference scenario. The dataset consists of 211 recorded meeting sessions, each containing 4 to 8 speakers, with a total length of 120 hours. This dataset aims to bridge the advanced research on multi-speaker processing and the practical application scenario in three aspects. With real recorded meetings, AISHELL-4 provides realistic acoustics and rich natural speech characteristics in conversation such as short pause, speech overlap, quick speaker turn, noise, etc. Meanwhile, accurate transcription and speaker voice activity are provided for each meeting in AISHELL-4. This allows the researchers to explore different aspects in meeting processing, ranging from individual tasks such as speech front-end processing, speech recognition and speaker diarization, to multi-modality modeling and joint optimization of relevant tasks. Given most open source dataset for multi-speaker tasks are in English, AISHELL-4 is the only Mandarin dataset for conversation speech, providing additional value for data diversity in speech community. We also release a PyTorch-based training and evaluation framework as baseline system to promote reproducible research in this field.

Speech separation has been successfully applied as a frontend processing module of conversation transcription systems thanks to its ability to handle overlapped speech and its flexibility to combine with downstream tasks such as automatic speech recognition (ASR). However, a speech separation model often introduces target speech distortion, resulting in a sub-optimum word error rate (WER). In this paper, we describe our efforts to improve the performance of a single channel speech separation system. Specifically, we investigate a two-stage training scheme that firstly applies a feature level optimization criterion for pretraining, followed by an ASR-oriented optimization criterion using an end-to-end (E2E) speech recognition model. Meanwhile, to keep the model light-weight, we introduce a modified teacher-student learning technique for model compression. By combining those approaches, we achieve a absolute average WER improvement of 2.70% and 0.77% using models with less than 10M parameters compared with the previous state-of-the-art results on the LibriCSS dataset for utterance-wise evaluation and continuous evaluation, respectively

In this paper, we propose a multi-channel network for simultaneous speech dereverberation, enhancement and separation (DESNet). To enable gradient propagation and joint optimization, we adopt the attentional selection mechanism of the multi-channel features, which is originally proposed in end-to-end unmixing, fixed-beamforming and extraction (E2E-UFE) structure. Furthermore, the novel deep complex convolutional recurrent network (DCCRN) is used as the structure of the speech unmixing and the neural network based weighted prediction error (WPE) is cascaded before-hand for speech dereverberation. We also introduce the staged SNR strategy and symphonic loss for the training of the network to further improve the final performance. Experiments show that in non-dereverberated case, the proposed DESNet outperforms DCCRN and most state-of-the-art structures in speech enhancement and separation, while in dereverberated scenario, DESNet also shows improvements over the cascaded WPE-DCCRN networks.

The IEEE Spoken Language Technology Workshop (SLT) 2021 Alpha-mini Speech Challenge (ASC) is intended to improve research on keyword spotting (KWS) and sound source location (SSL) on humanoid robots. Many publications report significant improvements in deep learning based KWS and SSL on open source datasets in recent years. For deep learning model training, it is necessary to expand the data coverage to improve the model robustness. Thus, simulating multi-channel noisy and reverberant data from single-channel speech, noise, echo and room impulsive response (RIR) is widely adopted. However, this approach may generate mismatch between simulated data and recorded data in real application scenarios, especially echo data. In this challenge, we open source a sizable speech, keyword, echo and noise corpus for promoting data-driven methods, particularly deep-learning approaches on KWS and SSL. We also choose Alpha-mini, a humanoid robot produced by UBTECH equipped with a built-in four-microphone array on its head, to record development and evaluation sets under the actual Alpha-mini robot application scenario, including environ-mental noise as well as echo and mechanical noise generated by the robot itself for model evaluation. Furthermore, we illustrate the rules, evaluation methods and baselines for re-searchers to quickly assess their achievements and optimize their models.

The front-end module in multi-channel automatic speech recognition (ASR) systems mainly use microphone array techniques to produce enhanced signals in noisy conditions with reverberation and echos. Recently, neural network (NN) based front-end has shown promising improvement over the conventional signal processing methods. In this paper, we propose to adopt the architecture of deep complex Unet (DCUnet) - a powerful complex-valued Unet-structured speech enhancement model - as the front-end of the multi-channel acoustic model, and integrate them in a multi-task learning (MTL) framework along with cascaded framework for comparison. Meanwhile, we investigate the proposed methods with several training strategies to improve the recognition accuracy on the 1000-hours real-world XiaoMi smart speaker data with echos. Experiments show that our proposed DCUnet-MTL method brings about 12.2% relative character error rate (CER) reduction compared with the traditional approach with array processing plus single-channel acoustic model. It also achieves superior performance than the recently proposed neural beamforming method.

The ResNet-based architecture has been widely adopted to extract speaker embeddings for text-independent speaker verification systems. By introducing the residual connections to the CNN and standardizing the residual blocks, the ResNet structure is capable of training deep networks to achieve highly competitive recognition performance. However, when the input feature space becomes more complicated, simply increasing the depth and width1of the ResNet network may not fully realize its performance potential. In this paper, we present two extensions of the ResNet architecture, ResNeXt and Res2Net, for speaker verification. Originally proposed for image recognition, the ResNeXt and Res2Net introduce two more dimensions, cardinality and scale, in addition to depth and width, to improve the model's representation capacity. By increasing the scale dimension, the Res2Net model can represent multi-scale features with various granularities, which particularly facilitates speaker verification for short utterances. We evaluate our proposed systems on three speaker verification tasks. Experiments on the VoxCeleb test set demonstrated that the ResNeXt and Res2Net can significantly outperform the conventional ResNet model. The Res2Net model achieved superior performance by reducing the EER by 18.5% relative. Experiments on the other two internal test sets of mismatched conditions further confirmed the generalization of the ResNeXt and Res2Net architectures against noisy environment and segment length variations.

This paper describes a dataset and protocols for evaluating continuous speech separation algorithms. Most prior speech separation studies use pre-segmented audio signals, which are typically generated by mixing speech utterances on computers so that they fully overlap. Also, the separation algorithms have often been evaluated based on signal-based metrics such as signal-to-distortion ratio. However, in natural conversations, speech signals are continuous and contain both overlapped and overlap-free regions. In addition, the signal-based metrics only have weak correlation with automatic speech recognition (ASR) accuracy. Not only does this make it hard to assess the practical relevance of the tested algorithms, it also hinders researchers from developing systems that can be readily applied to real scenarios. In this paper, we define continuous speech separation (CSS) as a task of generating a set of non-overlapped speech signals from a continuous audio stream that contains multiple utterances that are partially overlapped by a varying degree. A new real recording dataset, called LibriCSS, is derived from LibriSpeech by concatenating the corpus utterances to simulate conversations and capturing the audio replays with far-field microphones. A Kaldi-based ASR evaluation protocol is established by using a well-trained multi-conditional acoustic model. A recently proposed speaker-independent CSS algorithm is investigated by using LibriCSS. The dataset and evaluation scripts are made available to facilitate the research in this direction1.

With the advent of recurrent neural network transducer (RNN-T) model, the performance of keyword spotting (KWS) systems has greatly improved. However, the KWS systems, employed for wake-word detection, still rely on the availability of keyword specific training data for achieving reasonable performance on each keyword. With a goal to improve the KWS performance for these keywords without having to collect additional natural speech data, we explore Text-To-Speech (TTS) technology to synthetically generate training data for such keywords. Employing an RNN-T based KWS model, already well trained on large keyword-independent natural speech dataset, as a seed model, we run adaptation experiments using the generated keyword-specific TTS data. Besides observing a considerable improvement in the overall performance for the low-resource keywords, we find that the performance improvement with TTS-generated training data, similar to natural speech data, depends on speaker diversity, amount of data per speaker and data simulation. We get additional improvement in performance by selectively adapting specific parts of the RNN-T model and gain key insights into different architectural constructs of RNN-T model.

Deep speaker embedding models have been commonly used as a building block for speaker diarization systems; however, the speaker embedding model is usually trained according to a global loss defined on the training data, which could be suboptimal for distinguishing speakers locally in a specific meeting session. In this work we present the first use of graph neural networks (GNNs) for the speaker diarization problem, utilizing a GNN to refine speaker embeddings locally using the structural information between speech segments inside each session. The speaker embeddings extracted by a pre-trained model are remapped into a new embedding space, in which the different speakers within a single session are better separated. The model is trained for linkage prediction in a supervised manner by minimizing the difference between the affinity matrix constructed by the refined embeddings and the ground-truth adjacency matrix. Spectral clustering is then applied on top of the refined embeddings. We show that the clustering performance of the refined speaker embeddings outperforms the original embeddings significantly on both simulated and real meeting data, and our system achieves the state-of-the-art result on the NIST SRE 2000 CALLHOME database.

Automatic recognition of overlapped speech remains a highly challenging task to date. Motivated by the bimodal nature of human speech perception, this paper investigates the use of audio-visual technologies for overlapped speech recognition. Three issues associated with the construction of audio-visual speech recognition (AVSR) systems are addressed. First, the basic architecture designs i.e. end-to-end and hybrid of AVSR systems are investigated. Second, purposefully designed modality fusion gates are used to robustly integrate the audio and visual features. Third, in contrast to a traditional pipelined architecture containing explicit speech separation and recognition components, a streamlined and integrated AVSR system optimized consistently using the lattice-free MMI (LF-MMI) discriminative criterion is also proposed. The proposed LF-MMI time-delay neural network (TDNN) system establishes the state-of-the-art for the LRS2 dataset. Experiments on overlapped speech simulated from the LRS2 dataset suggest the proposed AVSR system outperformed the audio only baseline LF-MMI DNN system by up to 29.98% absolute in word error rate (WER) reduction, and produced recognition performance comparable to a more complex pipelined system. Consistent performance improvements of 4.89% absolute in WER reduction over the baseline AVSR system using feature fusion are also obtained.

Deep CNN networks have shown great success in various tasks for text-independent speaker recognition. In this paper, we explore two approaches for modeling long temporal contexts to improve the performance of the ResNet networks. The first approach is simply integrating the utterance-level mean and variance normalization into the ResNet architecture. Secondly, we combine the BLSTM and ResNet into one unified architecture. The BLSTM layers model long range, supposedly phonetically aware, context information, which could facilitate the ResNet to learn the optimal attention weight and suppress the environmental variations. The BLSTM outputs are projected into multiple-channel feature maps and fed into the ResNet network. Experiments on the VoxCeleb1 and the internal MS-SV tasks show that with attentive pooling, the proposed approaches achieve up to 23-28% relative improvement in EER over a well-trained ResNet.

Speech enhancement has benefited from the success of deep learning in terms of intelligibility and perceptual quality.Conventional time-frequency (TF) domain methods focus on predicting TF-masks or speech spectrum, via a naive convolution neural network (CNN) or recurrent neural network (RNN).Some recent studies use complex-valued spectrogram as a training target but train in a real-valued network, predicting the magnitude and phase component or real and imaginary part, respectively.Particularly, convolution recurrent network (CRN) integrates a convolutional encoder-decoder (CED) structure and long short-term memory (LSTM), which has been proven to be helpful for complex targets.In order to train the complex target more effectively, in this paper, we design a new network structure simulating the complex-valued operation, called Deep Complex Convolution Recurrent Network (DCCRN), where both CNN and RNN structures can handle complex-valued operation.The proposed DCCRN models are very competitive over other previous networks, either on objective or subjective metric.With only 3.7M parameters, our DCCRN models submitted to the Interspeech 2020 Deep Noise Suppression (DNS) challenge ranked first for the real-time-track and second for the non-real-time track in terms of Mean Opinion Score (MOS).

This paper presents a new input format, channel-wise subband input (CWS), for convolutional neural networks (CNN) based music source separation (MSS) models in the frequency domain. We aim to address the major issues in CNN-based high-resolution MSS model: high computational cost and weight sharing between distinctly different bands. Specifically, in this paper, we decompose the input mixture spectra into several bands and concatenate them channel-wise as the model input. The proposed approach enables effective weight sharing in each subband and introduces more flexibility between channels. For comparison purposes, we perform voice and accompaniment separation (VAS) on models with different scales, architectures, and CWS settings. Experiments show that the CWS input is beneficial in many aspects. We evaluate our method on musdb18hq test set, focusing on SDR, SIR and SAR metrics. Among all our experiments, CWS enables models to obtain 6.9% performance gain on the average metrics. With even a smaller number of parameters, less training data, and shorter training time, our MDenseNet with 8-bands CWS input still surpasses the original MMDenseNet with a large margin. Moreover, CWS also reduces computational cost and training time to a large extent.

Speaker attribution is required in many real-world applications, such as meeting transcription, where speaker identity is assigned to each utterance according to speaker voice profiles. In this paper, we propose to solve the speaker attribution problem by using graph-based semi-supervised learning methods. A graph of speech segments is built for each session, on which segments from voice profiles are represented by labeled nodes while segments from test utterances are unlabeled nodes. The weight of edges between nodes is evaluated by the similarities between the pretrained speaker embeddings of speech segments. Speaker attribution then becomes a semi-supervised learning problem on graphs, on which two graph-based methods are applied: label propagation (LP) and graph neural networks (GNNs). The proposed approaches are able to utilize the structural information of the graph to improve speaker attribution performance. Experimental results on real meeting data show that the graph based approaches reduce speaker attribution error by up to 68% compared to a baseline speaker identification approach that processes each utterance independently.

Although mask based adaptive beamforming technique benefits speech recognition in far-field, noisy and multi-talker scenarios, it depends on the long time context to estimate target and interference statistics, thus when applied in applications with low latency requirement, its performance usually drops drastically. In contrast, the fixed beamformers do not import time delay but usually have limited capability in acoustic cancellation of interfering source. In this work, we propose a novel multi-channel speech separation system that targets at overlapped speech recognition with low latency processing, which includes four jointly optimized components: a pre-separator, a set of fixed beamformer, an attentional selection module and neural post filtering. With proposed model, low latency processing is achieved by utilizing the known microphone geometry information, while keeps the high quality separation through neural post filtering and end-to-end optimization. In our experiments, we show that the proposed system achieves comparable performance in offline evaluation with the mask based MVDR and speech extraction system, while yield remarkable improvements in the online evaluation.

This paper describes the NPU system submitted to Interspeech 2020 Far-Field Speaker Verification Challenge (FFSVC). We particularly focus on far-field text-dependent SV from single (task1) and multiple microphone arrays (task3). The major challenges in such scenarios are short utterance and cross-channel and distance mismatch for enrollment and test. With the belief that better speaker embedding can alleviate the effects from short utterance, we introduce a new speaker embedding architecture - ResNet-BAM, which integrates a bottleneck attention module with ResNet as a simple and efficient way to further improve the representation power of ResNet. This contribution brings up to 1% EER reduction. We further address the mismatch problem in three directions. First, domain adversarial training, which aims to learn domain-invariant features, can yield to 0.8% EER reduction. Second, front-end signal processing, including WPE and beamforming, has no obvious contribution, but together with data selection and domain adversarial training, can further contribute to 0.5% EER reduction. Finally, data augmentation, which works with a specifically-designed data selection strategy, can lead to 2% EER reduction. Together with the above contributions, in the middle challenge results, our single submission system (without multi-system fusion) achieves the first and second place on task 1 and task 3, respectively.

Audio-visual multi-modal modeling has been demonstrated to be effective in many speech related tasks, such as speech recognition and speech enhancement. This paper introduces a new time-domain audio-visual architecture for target speaker extraction from monaural mixtures. The architecture generalizes the previous TasNet (time-domain speech separation network) to enable multi-modal learning and at meanwhile it extends the classical audio-visual speech separation from frequency-domain to time-domain. The main components of proposed architecture include an audio encoder, a video encoder that extracts lip embedding from video streams, a multi-modal separation network and an audio decoder. Experiments on simulated mixtures based on recently released LRS2 dataset show that our method can bring 3dB+ and 4dB+ Si-SNR improvements on two- and three-speaker cases respectively, compared to audio-only TasNet and frequency-domain audio-visual networks.

Text-independent speaker verification imposes no constraints on the spoken content and usually needs long observations to make reliable prediction. In this paper, we propose two speaker embedding approaches by integrating the phonetic information into the attention-based residual convolutional neural network (CNN). Phonetic features are extracted from the bottleneck layer of a pretrained acoustic model. In implicit phonetic attention (IPA), the phonetic features are projected by a transformation network into multi-channel feature maps, and then combined with the raw acoustic features as the input of the CNN network. In explicit phonetic attention (EPA), the phonetic features are directly connected to the attentive pooling layer through a separate 1-dim CNN to generate the attention weights. With the incorporation of spoken content and attention mechanism, the system can not only distill the speaker-discriminant frames but also actively normalize the phonetic variations. Multi-head attention and discriminative objectives are further studied to improve the system. Experiments on the VoxCeleb corpus show our proposed system could outperform the state-of-the-art by around 43% relative.

Speech separation has been studied widely for single-channel close-talk microphone recordings over the past few years; developed solutions are mostly in frequency-domain.Recently, a raw audio waveform separation network (TasNet) is introduced for single-channel data, with achieving high Si-SNR (scale-invariant source-to-noise ratio) and SDR (sourceto-distortion ratio) comparing against the state-of-the-art solution in frequency-domain.In this study, we incorporate effective components of the TasNet into a frequency-domain separation method.We compare both for alternative scenarios.We introduce a solution for directly optimizing the separation criterion in frequency-domain networks.In addition to speech separation objective and subjective measurements, we evaluate the separation performance on a speech recognition task as well.We study the speech separation problem for far-field data (more similar to naturalistic audio streams) and develop multi-channel solutions for both frequency and time-domain separators with utilizing spectral, spatial and speaker location information.For our experiments, we simulated multi-channel spatialized reverberate WSJ0-2mix dataset.Our experimental results show that spectrogram separation can achieve competitive performance with better network design.Multi-channel framework as well is shown to improve the single-channel performance relatively up to +35.5% and +46% in terms of WER and SDR, respectively.

This paper summarizes several follow-up contributions for improving our submitted NWPU speaker-dependent system for CHiME-5 challenge, which aims to solve the problem of multi-channel, highly-overlapped conversational speech recognition in a dinner party scenario with reverberations and nonstationary noises.We adopt a speaker-aware training method by using i-vector as the target speaker information for multi-talker speech separation.With only one unified separation model for all speakers, we achieve a 10% absolute improvement in terms of word error rate (WER) over the previous baseline of 80.28% on the development set by leveraging our newly proposed data processing techniques and beamforming approach.With our improved back-end acoustic model, we further reduce WER to 60.15% which surpasses the result of our submitted CHiME-5 challenge system without applying any fusion techniques.

This paper proposes and compares four cross-lingual and bilingual automatic speech recognition techniques under the constraint that only the acoustic model (AM) of the native language is used at runtime. The first three techniques fall into the category of lexicon conversion where each phoneme sequence (PHS) in the foreign language (FL) lexicon is mapped into the native language (NL) phoneme sequence. The first technique determines the PHS mapping through the international phonetic alphabet (IPA) features; The second and third techniques are data-driven. They determine the mapping by converting the PHS into corresponding context-independent and context-dependent hidden Markov models (HMMs) respectively and searching for the NL PHS with the least Kullback-Leibler divergence (KLD) between the HMMs. The fourth technique falls into the category of AM merging where the FL's AM is merged into the NL's AM by mapping each senone in the FL's AM to the senone in the NL's AM with the minimum KLD. We discuss the strengths and limitations of each technique developed, report empirical evaluation results on recognizing English utterances with a Korean recognizer, and demonstrate the high correlation between the average KLD and the word error rate (WER). The results show that the AM merging technique performs the best, achieving 60% relative WER reduction over the IPA-based technique.

Recently, we successfully developed and reported a new unsupervised online adaptation technique, which jointly compensates for additive and convolutive distortions with vector Taylor series (JAC/VTS), to adjust (uncompressed) HMMs under acoustically distorted environments [1]. In this paper, we extend that technique to adapt compressed HMMs using JAC/VTS where limited computation and/or memory resources are available for speech recognition (e.g., on mobile devices). Subspace coding (SSC) is developed and used to quantize each dimension of the multivariate Gaussians in the compressed HMMs. Three algorithmic design options are proposed and evaluated that combine SSC with JAC/VTS, where three different types of tradeoffs are made between recognition accuracy and the required computation/memory/storage resources. The strengths and weaknesses of these three options are discussed and shown on the Aurora2 task of noise-robust speech recognition. The first option greatly reduces the storage space and gives 93.2% accuracy, which is the same as the baseline accuracy but with little reduction in the run-time computation/memory cost. The second option reduces about 79.9% of the computation cost and about 33.5% of the memory requirement at a very small price of 0.5% decrease of accuracy (to 92.7%). The third option cuts about 89.2% of the computation cost and about 65.5% of the memory requirement while reducing recognition accuracy by 2.7% (to 90.5%).

We present a non-linear feature-domain noise reduction algorithm based on the minimum mean square error (MMSE) criterion on Mel-frequency cepstra (MFCC) for environment-robust speech recognition. Distinguishing from the MMSE enhancement in log spectral amplitude proposed by Ephraim and Malah (E&M) [7], the new algorithm presented in this paper develops the suppression rule that applies to power spectral magnitude of the filter-banks’ outputs and to MFCC directly, making it demonstrably more effective in noise-robust speech recognition. The noise variance in the new algorithm contains a significant term resulting from instantaneous phase asynchrony between clean speech and mixing noise, missing in the E&M algorithm. Speech recognition experiments on the standard Aurora-3 task demonstrate a reduction of word error rate by 48% against the ICSLP02 baseline, by 26% against the cepstral mean normalization baseline, and by 13% against the conventional E&M log-MMSE noise suppressor. The new algorithm is also much more efficient than E&M noise suppressor since the number of the channels in the Mel-frequency filter bank is much smaller (23 in our case) than the number of bins in the FFT domain (256). The results also show that our algorithm performs slightly better than the ETSI AFE on the well-matched and mid-mismatched settings.

We present an efficient and effective nonlinear feature-domain noise suppression algorithm, motivated by the minimum-mean-square-error (MMSE) optimization criterion, for noise-robust speech recognition. Distinguishing from the log-MMSE spectral amplitude noise suppressor proposed by Ephraim and Malah (E&M), our new algorithm is aimed to minimize the error expressed explicitly for the Mel-frequency cepstra instead of discrete Fourier transform (DFT) spectra, and it operates on the Mel-frequency filter bank's output. As a consequence, the statistics used to estimate the suppression factor become vastly different from those used in the E&M log-MMSE suppressor. Our algorithm is significantly more efficient than the E&M's log-MMSE suppressor since the number of the channels in the Mel-frequency filter bank is much smaller (23 in our case) than the number of bins (256) in DFT. We have conducted extensive speech recognition experiments on the standard Aurora-3 task. The experimental results demonstrate a reduction of the recognition word error rate by 48% over the standard ICSLP02 baseline, 26% over the cepstral mean normalization baseline, and 13% over the popular E&M's log-MMSE noise suppressor. The experiments also show that our new algorithm performs slightly better than the ETSI advanced front end (AFE) on the well-matched and mid-mismatched settings, and has 8% and 10% fewer errors than our earlier SPLICE (stereo-based piecewise linear compensation for environments) system on these settings, respectively.

Two feature extraction and compensation algorithms, feature-space minimum phone error (fMPE), which contributed to the recent significant progress in conversational speech recognition, and stereo-based piecewise linear compensation for environments (SPLICE), which has been used successfully in noise-robust speech recognition, are analyzed and compared. These two algorithms have been developed by very different motivations and been applied to very different speech-recognition tasks as well. While the mathematical construction of the two algorithms is ostensibly different, in this report, we establish a direct link between them. We show that both algorithms in the run-time operation accomplish feature extraction/compensation by adding a posterior-based weighted sum of "correction vectors," or equivalently the column vectors in the fMPE projection matrix, to the original, uncompensated features. Although the published fMPE algorithm empirically motivates such a feature extraction, operation as "a reasonable starting point for training" our analysis proves that it is a natural consequence of the rigorous minimum mean square error (MMSE) optimization rule as developed in SPLICE. Further, we review and compare related speech-recognition results with the use of fMPE and SPLICE algorithms. The results demonstrate the effectiveness of discriminative training on the feature extraction parameters (i.e., projection matrix in fMPE and equivalently correction vectors in SPLICE). The analysis and comparison of the two algorithms provide useful insight into the strong success of fMPE and point to further algorithm improvement and extension.