The emergence of novel generative modeling paradigms, particularly audio language models, has significantly advanced the field of song generation. Although state-of-the-art models are capable of synthesizing both vocals and accompaniment tracks up to several minutes long concurrently, research about partial adjustments or editing of existing songs is still underexplored, which allows for more flexible and effective production. In this paper, we present SongEditor, the first song editing paradigm that introduces the editing capabilities into language-modeling song generation approaches, facilitating both segment-wise and track-wise modifications. SongEditor offers the flexibility to adjust lyrics, vocals, and accompaniments, as well as synthesizing songs from scratch. The core components of SongEditor include a music tokenizer, an autoregressive language model, and a diffusion generator, enabling generating an entire section, masked lyrics, or even separated vocals and background music. Extensive experiments demonstrate that the proposed SongEditor achieves exceptional performance in end-to-end song editing, as evidenced by both objective and subjective metrics.

Speech language models (Speech LMs) enable end-to-end speech-text modeling within a single model, offering a promising direction for spoken dialogue systems. The choice of speech-text jointly decoding paradigm plays a critical role in performance, efficiency, and alignment quality. In this work, we systematically compare representative joint speech-text decoding strategies—including the interleaved, and parallel generation paradigms—under a controlled experimental setup using the same base language model, speech tokenizer and training data. Our results show that the interleaved approach achieves the best alignment. However it suffers from slow inference due to long token sequence length. To address this, we propose a novel early-stop interleaved (ESI) pattern that not only significantly accelerates decoding but also yields slightly better performance. Additionally, we curate high-quality question answering (QA) datasets to further improve speech QA performance.

Recent advancements in text-to-speech (TTS) have shown that language model (LM)-based systems offer competitive performance to their counterparts. Further optimization can be achieved through preference alignment algorithms, which adjust LMs to align with the preferences of reward models, enhancing the desirability of the generated content. This study presents a thorough empirical evaluation of how preference alignment algorithms, particularly Direct Preference Optimization (DPO), enhance LM-based TTS. With a 1.15B parameter LM-based TTS model, we demonstrate that preference alignment consistently improves intelligibility, speaker similarity, and proxy subjective evaluation scores, with the latter two metrics surpassing even human speech in certain evaluations. We also show preference alignment is applicable to low-resource scenarios and effectively generalized to out-of-domain applications.

Recent zero-shot text-to-speech (TTS) systems face a common dilemma: autoregressive (AR) models suffer from slow generation and lack duration controllability, while non-autoregressive (NAR) models lack temporal modeling and typically require complex designs. In this paper, we introduce a novel pseudo-autoregressive (PAR) codec language modeling approach that unifies AR and NAR modeling. Combining explicit temporal modeling from AR with parallel generation from NAR, PAR generates dynamic-length spans at fixed time steps. Building on PAR, we propose PALLE, a two-stage TTS system that leverages PAR for initial generation followed by NAR refinement. In the first stage, PAR progressively generates speech tokens along the time dimension, with each step predicting all positions in parallel but only retaining the left-most span. In the second stage, low-confidence tokens are iteratively refined in parallel, leveraging the global contextual information. Experiments demonstrate that PALLE, trained on LibriTTS, outperforms state-of-the-art systems trained on large-scale data, including F5-TTS, E2-TTS, and MaskGCT, on the LibriSpeech test-clean set in terms of speech quality, speaker similarity, and intelligibility, while achieving up to ten times faster inference speed. Audio samples are available at https://microsoft.com/research/project/vall-e-x/palle.

Music generation is pivotal in multimedia, aiding creation and lowering the creative threshold. It focuses on generating music with clear vocals and harmonious accompaniment based on lyrics, combining high artistic creativity with technical challenges. The music codec is an important bridging component in large language model-based music generation, connecting language models with the generated music. However, existing neural codecs typically require token rates exceeding 50 Hz to achieve acceptable music quality, resulting in a context length that surpasses 12,000 tokens for a 4-minute song-a scale that is computationally demanding. This highlights the need for high-compression, high-fidelity music codecs that can reconstruct both vocals and accompaniment with high quality at low frame rates and bitrates, thereby better assisting music generation. To address this, we introduce MuCodec, designed for high-quality music reconstruction at ultra-low bitrates, facilitating more efficient music generation. MuCodec employs a two-stage training method, enabling its encoder, MuEncoder, to extract semantic and acoustic features in a unified representation. These features are discretized using residual vector quantization and converted into Mel-VAE features through flow matching, with reconstruction quality improved by representation alignment during training. The Mel-VAE features are then reconstructed into music using a pretrained Mel-VAE decoder and HiFi-GAN. To the best of our knowledge, MuCodec is the first codec capable of reconstructing 48kHz stereo music at an ultra-low bitrate of 0.35 kbps (25 Hz), achieving state-of-the-art performance in both subjective and objective evaluations, and can more effectively support music generation. Code and Demo: https://mucodec.github.io/Mucodec/.

Recent advances in text-to-speech synthesis have achieved notable success in generating high-quality short utterances for individual speakers. However, these systems still face challenges when extending their capabilities to long, multi-speaker, and spontaneous dialogues, typical of real-world scenarios such as podcasts. These limitations arise from two primary challenges: 1) long speech: podcasts typically span several minutes, exceeding the upper limit of most existing work; 2) spontaneity: podcasts are marked by their spontaneous, oral nature, which sharply contrasts with formal, written contexts; existing works often fall short in capturing this spontaneity. In this paper, we propose MoonCast, a solution for high-quality zero-shot podcast generation, aiming to synthesize spontaneous podcast-style speech from text-only sources (e.g., stories, technical reports, news in TXT, PDF, or Web URL formats) using the voices of unseen speakers. To enable long audio generation, we employ a language model with parameter, data, and context scaling to process sequences in an innovative format designed for modeling entire multi-speaker, multi-turn speech interactions. To enhance spontaneity, we observe that ASR transcripts capture spontaneous speech details (e.g., filler words indicating hesitations, and specific punctuation and spaces reflecting breathing pauses), suggesting that these transcripts can serve as a partial indicator of speech spontaneity. Building upon this assumption, we utilize a script generation module to generate scripts incorporating these spontaneous elements. Experiments show MoonCast outperforms baselines, with notable improvements in contextual coherence and spontaneity.

We introduce MMAR, a new benchmark designed to evaluate the deep reasoning capabilities of Audio-Language Models (ALMs) across massive multi-disciplinary tasks. MMAR comprises 1,000 meticulously curated audio-question-answer triplets, collected from real-world internet videos and refined through iterative error corrections and quality checks to ensure high quality. Unlike existing benchmarks that are limited to specific domains of sound, music, or speech, MMAR extends them to a broad spectrum of real-world audio scenarios, including mixed-modality combinations of sound, music, and speech. Each question in MMAR is hierarchically categorized across four reasoning layers: Signal, Perception, Semantic, and Cultural, with additional sub-categories within each layer to reflect task diversity and complexity. To further foster research in this area, we annotate every question with a Chain-of-Thought (CoT) rationale to promote future advancements in audio reasoning. Each item in the benchmark demands multi-step deep reasoning beyond surface-level understanding. Moreover, a part of the questions requires graduate-level perceptual and domain-specific knowledge, elevating the benchmark's difficulty and depth. We evaluate MMAR using a broad set of models, including Large Audio-Language Models (LALMs), Large Audio Reasoning Models (LARMs), Omni Language Models (OLMs), Large Language Models (LLMs), and Large Reasoning Models (LRMs), with audio caption inputs. The performance of these models on MMAR highlights the benchmark's challenging nature, and our analysis further reveals critical limitations of understanding and reasoning capabilities among current models. These findings underscore the urgent need for greater research attention in audio-language reasoning, including both data and algorithm innovation. We hope MMAR will serve as a catalyst for future advances in this important but little-explored area.

Generating music with coherent structure, harmonious instrumental and vocal elements remains a significant challenge in song generation. Existing language models and diffusion-based methods often struggle to balance global coherence with local fidelity, resulting in outputs that lack musicality or suffer from incoherent progression and mismatched lyrics. This paper introduces SongBloom, a novel framework for full-length song generation that leverages an interleaved paradigm of autoregressive sketching and diffusion-based refinement. SongBloom employs an autoregressive diffusion model that combines the high fidelity of diffusion models with the scalability of language models. Specifically, it gradually extends a musical sketch from short to long and refines the details from coarse to fine-grained. The interleaved generation paradigm effectively integrates prior semantic and acoustic context to guide the generation process. Experimental results demonstrate that SongBloom outperforms existing methods across both subjective and objective metrics and achieves performance comparable to the state-of-the-art commercial music generation platforms. Audio samples are available on our demo page: https://cypress-yang.github.io/SongBloom_demo.

Generating natural-sounding, multi-speaker dialogue is crucial for applications such as podcast creation, virtual agents, and multimedia content generation. However, existing systems struggle to maintain speaker consistency, model overlapping speech, and synthesize coherent conversations efficiently. In this paper, we introduce CoVoMix2, a fully non-autoregressive framework for zero-shot multi-talker dialogue generation. CoVoMix2 directly predicts mel-spectrograms from multi-stream transcriptions using a flow-matching-based generative model, eliminating the reliance on intermediate token representations. To better capture realistic conversational dynamics, we propose transcription-level speaker disentanglement, sentence-level alignment, and prompt-level random masking strategies. Our approach achieves state-of-the-art performance, outperforming strong baselines like MoonCast and Sesame in speech quality, speaker consistency, and inference speed. Notably, CoVoMix2 operates without requiring transcriptions for the prompt and supports controllable dialogue generation, including overlapping speech and precise timing control, demonstrating strong generalizability to real-world speech generation scenarios. Audio samples are available at https://www.microsoft.com/en-us/research/project/covomix/covomix2.

Speech emotions are crucial in human communication and are extensively used in fields like speech synthesis and natural language understanding. Most prior studies, such as speech emotion recognition, have categorized speech emotions into a fixed set of classes. Yet, emotions expressed in human speech are often complex, and categorizing them into predefined groups can be insufficient to adequately represent speech emotions. On the contrary, describing speech emotions directly by means of natural language may be a more effective approach. Regrettably, there are not many studies available that have focused on this direction. Therefore, this paper proposes a speech emotion captioning framework named SECap, aiming at effectively describing speech emotions using natural language. Owing to the impressive capabilities of large language models in language comprehension and text generation, SECap employs LLaMA as the text decoder to allow the production of coherent speech emotion captions. In addition, SECap leverages HuBERT as the audio encoder to extract general speech features and Q-Former as the Bridge-Net to provide LLaMA with emotion-related speech features. To accomplish this, Q-Former utilizes mutual information learning to disentangle emotion-related speech features and speech contents, while implementing contrastive learning to extract more emotion-related speech features. The results of objective and subjective evaluations demonstrate that: 1) the SECap framework outperforms the HTSAT-BART baseline in all objective evaluations; 2) SECap can generate high-quality speech emotion captions that attain performance on par with human annotators in subjective mean opinion score tests.

Current speaker recognition systems primarily rely on supervised approaches, constrained by the scale of labeled datasets. To boost the system performance, researchers leverage large pretrained models such as WavLM to transfer learned high-level features to the downstream speaker recognition task. However, this approach introduces extra parameters as the pretrained model remains in the inference stage. Another group of researchers directly apply self-supervised methods such as DINO to speaker embedding learning, yet they have not explored its potential on large-scale in-the-wild datasets. In this paper, we present the effectiveness of DINO training on the large-scale WenetSpeech dataset and its transferability in enhancing the supervised system performance on the CNCeleb dataset. Additionally, we introduce a confidence-based data filtering algorithm to remove unreliable data from the pretraining dataset, leading to better performance with less training data. The associated pretrained models, confidence files, pretraining and finetuning scripts will be made available in the Wespeaker toolkit.

The query-based audio separation usually employs specific queries to extract target sources from a mixture of audio signals. Currently, most query-based separation models need additional networks to obtain query embedding. In this way, separation model is optimized to be adapted to the distribution of query embedding. However, query embedding may exhibit mismatches with separation models due to inconsistent structures and independent information. In this paper, we present CaRE-SEP, a consistent and relevant embedding network for general sound separation to encourage a comprehensive reconsideration of query usage in audio separation. CaRE-SEP alleviates the potential mismatch between queries and separation in two aspects, including sharing network structure and sharing feature information. First, a Swin-Unet model with a shared encoder is conducted to unify query encoding and sound separation into one model, eliminating the network architecture difference and generating consistent distribution of query and separation features. Second, by initializing CaRE-SEP with a pretrained classification network and allowing gradient backpropagation, the query embedding is optimized to be relevant to the separation feature, further alleviating the feature mismatch problem. Experimental results indicate the proposed CaRE-SEP model substantially improves the performance of separation tasks. Moreover, visualizations validate the potential mismatch and how CaRE-SEP solves it.

Recently, the utilization of extensive open-sourced text data has significantly advanced the performance of text-based large language models (LLMs). However, the use of in-the-wild large-scale speech data in the speech technology community remains constrained. One reason for this limitation is that a considerable amount of the publicly available speech data is compromised by background noise, speech overlapping, lack of speech segmentation information, missing speaker labels, and incomplete transcriptions, which can largely hinder their usefulness. On the other hand, human annotation of speech data is both time-consuming and costly. To address this issue, we introduce an automatic in-the-wild speech data preprocessing framework (AutoPrep) in this paper, which is designed to enhance speech quality, generate speaker labels, and produce transcriptions automatically. The proposed AutoPrep framework comprises six components: speech enhancement, speech segmentation, speaker clustering, target speech extraction, quality filtering and automatic speech recognition. Experiments conducted on the open-sourced WenetSpeech and our self-collected AutoPrepWild corpora demonstrate that the proposed AutoPrep framework can generate preprocessed data with similar DNSMOS and PDNSMOS scores compared to several open-sourced TTS datasets. The corresponding TTS system can achieve up to 0.68 in-domain speaker similarity.1

Target speaker extraction (TSE) aims to extract the target speaker’s voice from the input mixture. Previous studies have concentrated on high-overlapping scenarios. However, real-world applications usually meet more complex scenarios like variable speaker overlapping and target speaker absence. In this paper, we introduces a framework to perform continuous TSE (C-TSE), comprising a target speaker voice activation detection (TSVAD) and a TSE model. This framework significantly improves TSE performance on similar speakers and enhances personalization, which is lacking in traditional diarization methods. In detail, unlike conventional TSVAD deployed to refine the diarization results, the proposed Attention-target speaker voice activation detection (A-TSVAD) to directly generate timestamps of the target speaker. We also explore some different integration methods of A-TSVAD and TSE by comparing the cascaded and parallel methods. The framework’s effectiveness is assessed using a range of metrics, including diarization and enhancement metrics. Our experiments demonstrate that A-TSVAD outperforms conventional methods in reducing diarization errors, when integrating A-TSVAD and TSE in a sequential cascaded manner further enhances extraction accuracy. Audio demos are available on our demo page1.

This report presents the development of Tencent AI Lab’s personalized speech enhancement system for the 2023 ICASSP Signal Processing Grand Challenge – deep noise suppression (DNS) challenge1, which includes the use of a modified band-split recurrent neural network (BSRNN) and a multi-resolution spectrogram discriminator to improve perceptual quality metrics. The proposed system outperforms the baseline system by 0.047 and 0.036 final scores in the headset track and speaker phone track, respectively, and ranks among the top-32in both tracks.

While recent developments on the design of neural networks have greatly advanced the state-of-the-art of speech enhancement and separation systems, practical applications of such networks often put extra constraints on their model size and computational complexity. Moreover, as different telecommunication services may have different transmission bandwidths which result in different signal sample rates, one model is typically designed for a particular sample rate. In this paper, we extend the usage of a recently proposed frequency-domain source separation model, the band-split RNN (BSRNN), to the task of universal-sample-rate resource efficient speech enhancement. BSRNN explicitly splits the spectrogram into different frequency bands and perform interleaved band-level and sequence-level modeling, and the bandwidths can be manually designed to balance the model size, computational cost, and performance. By properly designing the band-splitting scheme and the hyperparameters, a single BSRNN model can handle signals at a wide range of sample rates, and the computational cost required to process a lower-sample-rate signal can be smaller than that of a higher-sample-rate signal. Experiment results show that compared to various benchmark systems in speech enhancement and separation, our universal-sample-rate BSRNN (USR-BSRNN) achieves comparable or better signalto-noise ratio (SNR) performance at a same level of model size or computational cost.

Echo cancellation and noise reduction are essential for full-duplex communication, yet most existing neural networks have high computational costs and are inflexible in tuning model complexity. In this paper, we introduce time-frequency dual-path compression to achieve a wide range of compression ratios on computational cost. Specifically, for frequency compression, trainable filters are used to replace manually designed filters for dimension reduction. For time compression, only using frame skipped prediction causes large performance degradation, which can be alleviated by a post-processing network with full sequence modeling. We have found that under fixed compression ratios, dual-path compression combining both the time and frequency methods will give further performance improvement, covering compression ratios from 4x to 32x with little model size change. Moreover, the proposed models show competitive performance compared with fast FullSubNet and DeepFilterNet.

The performance of music source separation (MSS) models has been greatly improved in recent years thanks to the development of novel neural network architectures and training pipelines. However, recent model designs for MSS were mainly motivated by other audio processing tasks or other research fields, such as speech enhancement and speech separation, while the intrinsic characteristics and patterns of the music signals were not fully discovered. In this paper, we propose band-split RNN (BSRNN), a frequency-domain model that explictly splits the spectrogram of the mixture into subbands and perform interleaved band-level and sequence-level modeling. The choices of the bandwidths of the subbands can be determined by a priori knowledge or expert knowledge on the characteristics of the target source in order to optimize the performance on a certain type of target musical instrument. To better make use of unlabeled data, we also describe a semi-supervised model finetuning pipeline that can further improve the performance of the model. Experiment results show that BSRNN trained only on MUSDB18-HQ dataset significantly outperforms several top-ranking models in Music Demixing (MDX) Challenge 2021, and the semi-supervised finetuning stage further improves the performance on all four instrument tracks.

In automatic speech recognition (ASR) research, discriminative criteria have achieved superior performance in DNN-HMM systems. Given this success, the adoption of discriminative criteria is promising to boost the performance of end-to-end (E2E) ASR systems. With this motivation, previous works have introduced the minimum Bayesian risk (MBR, one of the discriminative criteria) into E2E ASR systems. However, the effectiveness and efficiency of the MBR-based methods are compromised: the MBR criterion is only used in system training, which creates a mismatch between training and decoding; the on-the-fly decoding process in MBR-based methods results in the need for pre-trained models and slow training speeds. To this end, novel algorithms are proposed in this work to integrate another widely used discriminative criterion, lattice-free maximum mutual information (LF-MMI), into E2E ASR systems not only in the training stage but also in the decoding process. The proposed LF-MMI training and decoding methods show their effectiveness on two widely used E2E frameworks: Attention-Based Encoder-Decoders (AEDs) and Neural Transducers (NTs). Compared with MBR-based methods, the proposed LF-MMI method: maintains the consistency between training and decoding; eschews the on-the-fly decoding process; trains from randomly initialized models with superior training efficiency. Experiments suggest that the LF-MMI method outperforms its MBR counterparts and consistently leads to statistically significant performance improvements on various frameworks and datasets from 30 hours to 14.3 k hours. The proposed method achieves state-of-the-art (SOTA) results on Aishell-1 (CER 4.10%) and Aishell-2 (CER 5.02%) datasets. Code is released1.

Generating sound effects that people want is an important topic. However, there are limited studies in this area for sound generation. In this study, we investigate generating sound conditioned on a text prompt and propose a novel text-to-sound generation framework that consists of a text encoder, a Vector Quantized Variational Autoencoder (VQ-VAE), a token-decoder, and a vocoder. The framework first uses the token-decoder to transfer the text features extracted from the text encoder to a mel-spectrogram with the help of VQ-VAE, and then the vocoder is used to transform the generated mel-spectrogram into a waveform. We found that the token-decoder significantly influences the generation performance. Thus, we focus on designing a good token-decoder in this study. We begin with th21e traditional autoregressive (AR) token-decoder, which has shown state-of-the-art performance in previous sound generation works. However, the AR token-decoder always predicts the mel-spectrogram tokens one by one in order, which may introduce the unidirectional bias and accumulation of errors problems. Moreover, with the AR token-decoder, the sound generation time increases linearly with the sound duration. To overcome the shortcomings introduced by AR token-decoders, we propose a non-autoregressive token-decoder based on the discrete diffusion model, named Diffsound. Specifically, the Diffsound model predicts all of the mel-spectrogram tokens in one step and then refines the predicted tokens in the next step, so the best-predicted results can be obtained by iteration. Our experiments show that our proposed Diffsound model not only produces better text-to-sound generation results when compared with the AR token-decoder but also has a faster generation speed,i.e., MOS: 3.56v.s2.786, and the generation speed is five times faster than the AR decoder. Furthermore, to automatically assess the quality of generated samples, we define three different objective evaluation metricsi.e., Fréchet Inception Distance (FID), Kullback-Leibler (KL), and audio caption loss, which can comprehensively assess the relevance and fidelity of the generated samples.

Despite the rapid advance of automatic speech recognition (ASR) technologies, accurate recognition of cocktail party speech characterised by the interference from overlapping speakers, background noise and room reverberation remains a highly challenging task to date. Motivated by the invariance of visual modality to acoustic signal corruption, audio-visual speech enhancement techniques have been developed, although predominantly targeting overlapping speech separation and recognition tasks. In this paper, an audio-visual multi-channel speech separation, dereverberation and recognition approach featuring a full incorporation of visual information into all three stages of the system is proposed. The advantage of the additional visual modality over using audio only is demonstrated on two neural dereverberation approaches based on DNN-WPE and spectral mapping respectively. The learning cost function mismatch between the separation and dereverberation models and their integration with the back-end recognition system is minimised using fine-tuning on the MSE and LF-MMI criteria. Experiments conducted on the LRS2 dataset suggest that the proposed audio-visual multi-channel speech separation, dereverberation and recognition system outperforms the baseline audio-visual multi-channel speech separation and recognition system containing no dereverberation module by a statistically significant word error rate (WER) reduction of 2.06 % absolute (8.77 % relative).

Recently, End-to-End (E2E) frameworks have achieved remarkable results on various Automatic Speech Recognition (ASR) tasks. However, Lattice-Free Maximum Mutual Information (LF-MMI), as one of the discriminative training criteria that show superior performance in hybrid ASR systems, is rarely adopted in E2E ASR frameworks. In this work, we propose a novel approach to introduce LF-MMI criterion into E2E ASR frameworks in both training and decoding stages. The proposed approach shows its effectiveness on two of the most widely used E2E frameworks including Attention-Based Encoder-Decoders (AEDs) and Neural Transducers (NTs). Experiments suggest that the introduction of the LF-MMI criterion consistently leads to significant performance improvements on various datasets and different E2E ASR frameworks. The best of our models achieves competitive CER of 4.1% / 4.4% on Aishell-1 dev/test set; significant error reduction is also achieved on Aishell-2 and Librispeech datasets over strong baselines. Code is released1.

Recognition of overlapped speech has been a highly challenging task to date. State-of-the-art multi-channel speech separation system are becoming increasingly complex and expensive for practical applications. To this end, low-bit neural network quantization provides a powerful solution to dramatically reduce their model size. However, current quantization methods are based on uniform precision and fail to account for the varying performance sensitivity at different model components to quantization errors. In this paper, novel mixed precision DNN quantization methods are proposed by applying locally variable bit-widths to individual TCN components of a TF masking based multi-channel speech separation system. The optimal local precision settings are automatically learned using three techniques. The first two approaches utilize quantization sensitivity metrics based on either the mean square error (MSE) loss function curvature, or the KL-divergence measured between full precision and quantized separation models. The third approach is based on mixed precision neural architecture search. Experiments conducted on the LRS3-TED corpus simulated overlapped speech data suggest that the proposed mixed precision quantization techniques consistently outperform the uniform precision baseline speech separation systems of comparable bit-widths in terms of SI-SNR and PESQ scores as well as word error rate (WER) reductions up to 2.88% absolute (8% relative).

Speaker identification for overlapped speech presents a great challenge for speaker diarization tasks in meeting scenarios. In order to overcome such challenges, several overlap-aware resegmentation methods based on deep learning have been integrated into speaker diarization systems. In this paper we propose two multi-channel diarization systems which have enhanced capability in detecting overlapped speech and identify speakers via learning spatial features. The first system applies a multi-look strategy to train networks without given the speakers’ direction of arrival(DOA), and the other system estimates the DOA of target speakers based on existing diarization results. Both systems aim to estimate the voice activity of speakers in different directions to handle overlapped speech. Experimental results on the AMI corpus show that the relative improvements of both systems can reach 9.4% and 18.1% in term of diarization error rate (DER) against an overlap-aware single-channel system with a BeamformIt front-end.

Prosodic boundary plays an important role in text-to-speech synthesis (TTS) in terms of naturalness and readability.However, the acquisition of prosodic boundary labels relies on manual annotation, which is costly and time-consuming.In this paper, we propose to automatically extract prosodic boundary labels from text-audio data via a neural text-speech model with pre-trained audio encoders.This model is pre-trained on text and speech data separately and jointly fine-tuned on TTS data in a triplet format: {speech, text, prosody}.The experimental results on both automatic evaluation and human evaluation demonstrate that: 1) the proposed text-speech prosody annotation framework significantly outperforms text-only baselines; 2) the quality of automatic prosodic boundary annotations is comparable to human annotations; 3) TTS systems trained with model-annotated boundaries are slightly better than systems that use manual ones.Code is released 1 .

Despite the rapid progress in automatic speech recognition (ASR) research, recognizing multilingual speech using a unified ASR system remains highly challenging. Previous works on multilingual speech recognition mainly focus on two directions: recognizing multiple monolingual speech or recognizing code-switched speech that uses different languages interchangeably within a single utterance. However, a pragmatic multilingual recognizer is expected to be compatible with both directions. In this work, a novel language-aware encoder (LAE) architecture is proposed to handle both situations by disentangling language-specific information and generating frame-level language-aware representations during encoding. In the LAE, the primary encoding is implemented by the shared block while the language-specific blocks are used to extract specific representations for each language. To learn language-specific information discriminatively, a language-aware training method is proposed to optimize the language-specific blocks in LAE. Experiments conducted on Mandarin-English code-switched speech suggest that the proposed LAE is capable of discriminating different languages in frame-level and shows superior performance on both monolingual and multilingual ASR tasks. With either a real-recorded or simulated code-switched dataset, the proposed LAE achieves statistically significant improvements on both CTC and neural transducer systems. Code is released

Despite the rapid progress of end-to-end (E2E) automatic speech recognition (ASR), it has been shown that incorporating external language models (LMs) into the decoding can further improve the recognition performance of E2E ASR systems. To align with the modelingunits adopted in E2E ASR systems, subword-level (e.g., characters, BPE) LMs are usually used to cooperate with current E2E ASR systems. However, the use of subword-level LMs will ignore the word-level information, which may limit the strength of the external LMs in E2E ASR. Although several methods have been proposed to incorporate word-level external LMs in E2E ASR, these methods are mainly designed for languages with clear word boundaries such as English and cannot be directly applied to languages like Mandarin, in which each character sequence can have multiple corresponding word sequences. To this end, we propose a novel decoding algorithm where a word-level lattice is constructed on-the-fly to consider all possible word sequences for each partial hypothesis. Then, the LM score of the hypothesis is obtained by intersecting the generated lattice with an external word N-gram LM. The proposed method is examined on both Attention-based Encoder-Decoder (AED) and Neural Transducer (NT) frameworks. Experiments suggest that our method consistently outperforms subword-level LMs, including N-gram LM and neural network LM. We achieve state-of-the-art results on both Aishell-1 (CER 4.18%) and Aishell-2 (CER 5.06%) datasets and reduce CER by 14.8% relatively on a 21K-hour Mandarin dataset. Code is released.

State-of-the-art automatic speech recognition (ASR) system development is data and computation intensive. The optimal design of deep neural networks (DNNs) for these systems often require expert knowledge and empirical evaluation. In this paper, a range of neural architecture search (NAS) techniques are used to automatically learn two types of hyper-parameters of factored time delay neural networks (TDNN-Fs): i) the left and right splicing context offsets; and ii) the dimensionality of the bottleneck linear projection at each hidden layer. These techniques include the differentiable neural architecture search (DARTS) method integrating architecture learning with lattice-free MMI training; Gumbel-Softmax and pipelined DARTS methods reducing the confusion over candidate architectures and improving the generalization of architecture selection; and Penalized DARTS incorporating resource constraints to balance the trade-off between performance and system complexity. Parameter sharing among TDNN-F architectures allows an efficient search over up to$7^{28}$different systems. Statistically significant word error rate (WER) reductions of up to 1.2% absolute and relative model size reduction of 31% were obtained over a state-of-the-art 300-hour Switchboard corpus trained baseline LF-MMI TDNN-F system featuring speed perturbation, i-Vector and learning hidden unit contribution (LHUC) based speaker adaptation as well as RNNLM rescoring. Performance contrasts on the same task against recent end-to-end systems reported in the literature suggest the best NAS auto-configured system achieves state-of-the-art WERs of 9.9% and 11.1% on the NIST Hub5’ 00 and Rt03 s test sets respectively with up to 96% model size reduction. Further analysis using Bayesian learning shows that the proposed NAS approaches can effectively minimize the structural redundancy in the TDNN-F systems and reduce their model parameter uncertainty. Consistent performance improvements were also obtained on a UASpeech dysarthric speech recognition task.

With the global population ageing rapidly, Alzheimer's disease (AD) is particularly prominent in older adults, which has an insidious onset followed by gradual, irreversible deterioration in cognitive domains (memory, communication, etc). Thus the detection of Alzheimer's disease is crucial for timely intervention to slow down disease progression. This paper presents a comparative study of different acoustic and linguistic features for the AD detection using various classifiers. Experimental results on ADReSS dataset reflect that the proposed models using ComParE, X-vector, Linguistics, TFIDF and BERT features are able to detect AD with high accuracy and sensitivity, and are comparable with the state-of-the-art results reported. While most previous work used manual transcripts, our results also indicate that similar or even better performance could be obtained using automatically recognized transcripts over manually collected ones. This work achieves accuracy scores at 0.67 for acoustic features and 0.88 for linguistic features on either manual or ASR transcripts on the ADReSS Challenge1test set.

State-of-the-art neural language models represented by Transformers are becoming increasingly complex and expensive for practical applications. Low-bit deep neural network quantization techniques provides a powerful solution to dramatically reduce their model size. Current low-bit quantization methods are based on uniform precision and fail to account for the varying performance sensitivity at different parts of the system to quantization errors. To this end, novel mixed precision DNN quantization methods are proposed in this paper. The optimal local precision settings are automatically learned using two techniques. The first is based on a quantization sensitivity metric in the form of Hessian trace weighted quantization perturbation. The second is based on mixed precision Transformer architecture search. Alternating direction methods of multipliers (ADMM) are used to efficiently train mixed precision quantized DNN systems. Experiments conducted on Penn Treebank (PTB) and a Switchboard corpus trained LF-MMI TDNN system suggest the proposed mixed precision Transformer quantization techniques achieved model size compression ratios of up to 16 times over the full precision baseline with no recognition performance degradation. When being used to compress a larger full precision Transformer LM with more layers, overall word error rate (WER) reductions up to 1.7% absolute (18% relative) were obtained.

State-of-the-art neural language models (LMs) represented by Transformers are highly complex. Their use of fixed, deterministic parameter estimates fail to account for model uncertainty and lead to over-fitting and poor generalization when given limited training data. In order to address these issues, this paper proposes a full Bayesian learning framework for Transformer LM estimation. Efficient variational inference based approaches are used to estimate the latent parameter posterior distributions associated with different parts of the Transformer model architecture including multi-head self-attention, feed forward and embedding layers. Statistically significant word error rate (WER) reductions up to 0.5% absolute (3.18% relative) and consistent perplexity gains were obtained over the baseline Transformer LMs on state-of-the-art Switchboard corpus trained LF-MMI factored TDNN systems with i-Vector speaker adaptation. Performance improvements were also obtained on a cross domain LM adaptation task requiring porting a Transformer LM trained on the Switchboard and Fisher data to a low-resource DementiaBank elderly speech corpus.

Early diagnosis of Neurocognitive Disorder (NCD) is crucial in facilitating preventive care and timely treatment to delay further progression. This paper presents the development of a state-of-the-art automatic speech recognition (ASR) system built on the Dementia-Bank Pitt corpus for automatic NCD detection. Speed perturbation based audio data augmentation expanded the limited elderly speech data by four times. Large quantities of out-of-domain, non-aged adult speech were exploited by cross-domain adapting a 1000-hour LibriSpeech corpus trained LF-MMI factored TDNN system to DementiaBank. The variability among elderly speakers was modeled using i-Vector and learning hidden unit contributions (LHUC) based speaker adaptive training. Robust Bayesian estimation of TDNN systems and LHUC transforms were used in both cross-domain and speaker adaptation. A Transformer language model was also built to improve the final system performance. A word error rate (WER) reduction of 11.72% absolute (26.11% relative) was obtained over the baseline i-Vector adapted LF-MMI TDNN system on the evaluation data of 48 elderly speakers. The best NCD detection accuracy of 88%, comparable to that using the ground truth speech transcripts, was obtained using the textual features extracted from the final ASR system outputs.

In multi-talker cases, overlapped speech degrades the speaker verification (SV) performance dramatically. To tackle this challenging problem, speech separation with multi-channel techniques can be adopted to extract each speaker’s signals to improve the SV performance. In this paper, a joint training framework of the front-end multi-look speech separator and the back-end speaker embedding extractor is proposed for multi-channel overlapped speech. To better leverage the complementarity between the speech separator and the speaker embedding extractor, several training strategies are proposed to jointly optimize the two modules. Experimental results show that the proposed joint training framework significantly outperforms the individual SV system by around 52% relative EER reduction. Additionally, the robustness of the proposed framework is further evaluated under different conditions.

Automatic recognition of disordered speech remains a highly challenging task to date. Sources of variability commonly found in normal speech including accent, age or gender, when further compounded with the underlying causes of speech impairment and varying severity levels, create large diversity among speakers. To this end, speaker adaptation techniques play a vital role in current speech recognition systems. Motivated by the spectro-temporal level differences between disordered and normal speech that systematically manifest in articulatory imprecision, decreased volume and clarity, slower speaking rates and increased dysfluencies, novel spectro-temporal subspace basis embedding deep features derived by SVD decomposition of speech spectrum are proposed to facilitate both accurate speech intelligibility assessment and auxiliary feature based speaker adaptation of state-of-the-art hybrid DNN and end-to-end disordered speech recognition systems. Experiments conducted on the UASpeech corpus suggest the proposed spectro-temporal deep feature adapted systems consistently outperformed baseline i-Vector adaptation by up to 2.63% absolute (8.6% relative) reduction in word error rate (WER) with or without data augmentation. Learning hidden unit contribution (LHUC) based speaker adaptation was further applied. The final speaker adapted system using the proposed spectral basis embedding features gave an overall WER of 25.6% on the UASpeech test set of 16 dysarthric speakers

Automatic recognition of disordered speech remains a highly challenging task to date. The underlying neuro-motor conditions, often compounded with co-occurring physical disabilities, lead to the difficulty in collecting large quantities of impaired speech required for ASR system development. To this end, data augmentation techniques play a vital role in current disordered speech recognition systems. In contrast to existing data augmentation techniques only modifying the speaking rate or overall shape of spectral contour, fine-grained spectro-temporal differences between disordered and normal speech are modelled using deep convolutional generative adversarial networks (DCGAN) during data augmentation to modify normal speech spectra into those closer to disordered speech. Experiments conducted on the UASpeech corpus suggest the proposed adversarial data augmentation approach consistently outperformed the baseline augmentation methods using tempo or speed perturbation on a state-of-the-art hybrid DNN system. An overall word error rate (WER) reduction up to 3.05\% (9.7\% relative) was obtained over the baseline system using no data augmentation. The final learning hidden unit contribution (LHUC) speaker adapted system using the best adversarial augmentation approach gives an overall WER of 25.89% on the UASpeech test set of 16 dysarthric speakers.

In this paper, we exploit the effective way to leverage contextual information to improve the speech dereverberation performance in real-world reverberant environments. We propose a temporal-contextual attention approach on the deep neural network (DNN) for environment-aware speech dereverberation, which can adaptively attend to the contextual information. More specifically, a FullBand based Temporal Attention approach (FTA) is proposed, which models the correlations between the fullband information of the context frames. In addition, considering the difference between the attenuation of high frequency bands and low frequency bands (high frequency bands attenuate faster than low frequency bands) in the room impulse response (RIR), we also propose a SubBand based Temporal Attention approach (STA). In order to guide the network to be more aware of the reverberant environments, we jointly optimize the dereverberation network and the reverberation time (RT60) estimator in a multi-task manner. Our experimental results indicate that the proposed method outperforms our previously proposed reverberation-time-aware DNN and the learned attention weights are fully physical consistent. We also report a preliminary yet promising dereverberation and recognition experiment on real test data.

Discriminative training techniques define state-of-the-art performance for automatic speech recognition systems. However, they are inherently prone to overfitting, leading to poor generalization performance when using limited training data. In order to address this issue, this paper presents a full Bayesian framework to account for model uncertainty in sequence discriminative training of factored TDNN acoustic models. Several Bayesian learning based TDNN variant systems are proposed to model the uncertainty over weight parameters and choices of hidden activation functions, or the hidden layer outputs. Efficient variational inference approaches using as few as one single parameter sample ensure their computational cost in both training and evaluation time comparable to that of the baseline TDNN systems. Statistically significant word error rate (WER) reductions of 0.4%-1.8% absolute (5%-11% relative) were obtained over a state-of-the-art 900 h speed perturbed Switchboard corpus trained baseline LF-MMI factored TDNN system using multiple regularization methods including F-smoothing, L2 norm penalty, natural gradient, model averaging and dropout, in addition to i-Vector plus learning hidden unit contribution (LHUC) based speaker adaptation and RNNLM rescoring. The efficacy of the proposed Bayesian techniques is further demonstrated in a comparison against the state-of-the-art performance obtained on the same task using the most recent hybrid and end-to-end systems reported in the literature. Consistent performance improvements were also obtained on a 450-h HKUST conversational Mandarin telephone speech recognition task. On a third cross domain adaptation task requiring rapidly porting a 1000-h LibriSpeech data trained system to a small DementiaBank elderly speech corpus, the proposed Bayesian TDNN LF-MMI systems outperformed the baseline system using direct weight fine-tuning by up to 2.5% absolute WER reduction.

Despite the rapid progress of automatic speech recognition (ASR) technologies in the past few decades, recognition of disordered speech remains a highly challenging task to date. Disordered speech presents a wide spectrum of challenges to current data intensive deep neural networks (DNNs) based ASR technologies that predominantly target normal speech. This paper presents recent research efforts at the Chinese University of Hong Kong (CUHK) to improve the performance of disordered speech recognition systems on the largest publicly available UASpeech dysarthric speech corpus. A set of novel modelling techniques including neural architectural search, data augmentation using spectra-temporal perturbation, model based speaker adaptation and cross-domain generation of visual features within an audio-visual speech recognition (AVSR) system framework were employed to address the above challenges. The combination of these techniques produced the lowest published word error rate (WER) of 25.21% on the UASpeech test set 16 dysarthric speakers, and an overall WER reduction of 5.4% absolute (17.6% relative) over the CUHK 2018 dysarthric speech recognition system featuring a 6-way DNN system combination and cross adaptation of out-of-domain normal speech data trained systems. Bayesian model adaptation further allows rapid adaptation to individual dysarthric speakers to be performed using as little as 3.06 seconds of speech. The efficacy of these techniques were further demonstrated on a CUDYS Cantonese dysarthric speech recognition task.

State-of-the-art language models (LMs) represented by long-short term memory recurrent neural networks (LSTM-RNNs) and Transformers are becoming increasingly complex and expensive for practical applications. Low-bit neural network quantization provides a powerful solution to dramatically reduce their model size. Current quantization methods are based on uniform precision and fail to account for the varying performance sensitivity at different parts of LMs to quantization errors. To this end, novel mixed precision neural network LM quantization methods are proposed in this paper. The optimal local precision choices for LSTM-RNN and Transformer based neural LMs are automatically learned using three techniques. The first two approaches are based on quantization sensitivity metrics in the form of either the KL-divergence measured between full precision and quantized LMs, or Hessian trace weighted quantization perturbation that can be approximated efficiently using matrix free techniques. The third approach is based on mixed precision neural architecture search. In order to overcome the difficulty in using gradient descent methods to directly estimate discrete quantized weights, alternating direction methods of multipliers (ADMM) are used to efficiently train quantized LMs. Experiments were conducted on state-of-the-art LF-MMI CNN-TDNN systems featuring speed perturbation, i-Vector and learning hidden unit contribution (LHUC) based speaker adaptation on two tasks: Switchboard telephone speech and AMI meeting transcription. The proposed mixed precision quantization techniques achieved “lossless” quantization on both tasks, by producing model size compression ratios of up to approximately 16 times over the full precision LSTM and Transformer baseline LMs, while incurring no statistically significant word error rate increase.

Automatic speech recognition (ASR) technologies have been significantly advanced in the past few decades. However, recognition of overlapped speech remains a highly challenging task to date. To this end, multi-channel microphone array data are widely used in current ASR systems. Motivated by the invariance of visual modality to acoustic signal corruption and the additional cues they provide to separate the target speaker from the interfering sound sources, this paper presents an audio-visual multi-channel based recognition system for overlapped speech. It benefits from a tight integration between a speech separation front-end and recognition back-end, both of which incorporate additional video input. A series of audio-visual multi-channel speech separation front-end components based on TF masking, Filter&Sum and mask-based MVDR neural channel integration approaches are developed. To reduce the error cost mismatch between the separation and the recognition components, the entire system is jointly fine-tuned using a multi-task criterion interpolation of the scale-invariant signal to noise ratio (Si-SNR) with either the connectionist temporal classification (CTC), or lattice-free maximum mutual information (LF-MMI) loss function. Experiments suggest that: the proposed audio-visual multi-channel recognition system outperforms the baseline audio-only multi-channel ASR system by up to 8.04% (31.68% relative) and 22.86% (58.51% relative) absolute WER reduction on overlapped speech constructed using either simulation or replaying of the LRS2 dataset respectively. Consistent performance improvements are also obtained using the proposed audio-visual multi-channel recognition system when using occluded video input with the lip region randomly covered up to 60%.

This work investigates the vulnerability of Gaussian Mixture Model (GMM) i-vector based speaker verification systems to adversarial attacks, and the transferability of adversarial samples crafted from GMM i-vector based systems to x-vector based systems. In detail, we formulate the GMM i-vector system as a scoring function of enrollment and testing utterance pairs. Then we leverage the fast gradient sign method (FGSM) to optimize testing utterances for adversarial samples generation. These adversarial samples are used to attack both GMM i-vector and x-vector systems. We measure the system vulnerability by the degradation of equal error rate and false acceptance rate. Experiment results show that GMM i-vector systems are seriously vulnerable to adversarial attacks, and the crafted adversarial samples are proved to be transferable and pose threats to neural network speaker embedding based systems (e.g. x-vector systems).

Dysarthric speech reconstruction (DSR) is a challenging task due to difficulties in repairing unstable prosody and correcting imprecise articulation. Inspired by the success of sequence-to-sequence (seq2seq) based text-to-speech (TTS) synthesis and knowledge distillation (KD) techniques, this paper proposes a novel end-to-end voice conversion (VC) method to tackle the reconstruction task. The proposed approach contains three components. First, a seq2seq based TTS is first trained with the transcribed normal speech. Second, with the text-encoder of this trained TTS system as "teacher", a teacher-student framework is proposed for cross-modal KD by training a speech-encoder to extract appropriate linguistic representations from the transcribed dysarthric speech. Third, the speech-encoder of the previous component is concatenated with the attention and decoder of the first component (TTS) to perform the DSR task, by directly mapping the dysarthric speech to its normal version. Experiments demonstrate that the proposed method can generate the speech with high naturalness and intelligibility, where the comparisons of human speech recognition between the reconstructed speech and the original dysarthric speech show that 35.4% and 48.7% absolute word error rate (WER) reduction can be achieved for dysarthric speakers with low and very low speech intelligibility, respectively.

The high memory consumption and computational costs of Recurrent neural network language models (RNNLMs) limit their wider application on resource constrained devices. In recent years, neural network quantization techniques that are capable of producing extremely low-bit compression, for example, binarized RNNLMs, are gaining increasing research interests. Directly training of quantized neural networks is difficult. By formulating quantized RNNLMs training as an optimization problem, this paper presents a novel method to train quantized RNNLMs from scratch using alternating direction methods of multipliers (ADMM). This method can also flexibly adjust the trade-off between the compression rate and model performance using tied low-bit quantization tables. Experiments on two tasks: Penn Treebank (PTB), and Switchboard (SWBD) suggest the proposed ADMM quantization achieved a model size compression factor of up to 31 times over the full precision baseline RNNLMs. Faster convergence of 5 times in model training over the baseline binarized RNNLM quantization was also obtained.

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.

Disordered speech recognition is a highly challenging task. The underlying neuro-motor conditions of people with speech disorders, often compounded with co-occurring physical disabilities, lead to the difficulty in collecting large quantities of speech required for system development. This paper investigates a set of data augmentation techniques for disordered speech recognition, including vocal tract length perturbation (VTLP), tempo perturbation and speed perturbation. Both normal and disordered speech were exploited in the augmentation process. Variability among impaired speakers in both the original and augmented data was modeled using learning hidden unit contributions (LHUC) based speaker adaptive training. The final speaker adapted system constructed using the UASpeech corpus and the best augmentation approach based on speed perturbation produced up to 2.92% absolute (9.3% relative) word error rate (WER) reduction over the baseline system without data augmentation, and gave an overall WER of 26.37% on the test set containing 16 dysarthric speakers.

Automatic speech recognition (ASR) of overlapped speech remains a highly challenging task to date. To this end, multi-channel microphone array data are widely used in state-of-the-art ASR systems. Motivated by the invariance of visual modality to acoustic signal corruption, this paper presents an audio-visual multi-channel overlapped speech recognition system featuring tightly integrated separation front-end and recognition back-end. A series of audio-visual multi-channel speech separation front-end components based on \textit{TF masking}, \textit{filter\&sum} and \textit{mask-based MVDR} beamforming approaches were developed. To reduce the error cost mismatch between the separation and recognition components, they were jointly fine-tuned using the connectionist temporal classification (CTC) loss function, or a multi-task criterion interpolation with scale-invariant signal to noise ratio (Si-SNR) error cost. Experiments suggest that the proposed multi-channel AVSR system outperforms the baseline audio-only ASR system by up to 6.81\% (26.83\% relative) and 22.22\% (56.87\% relative) absolute word error rate (WER) reduction on overlapped speech constructed using either simulation or replaying of the lipreading sentence 2 (LRS2) dataset respectively.

State-of-the-art text-to-speech (TTS) synthesis models can produce monolingual speech with high intelligibility and naturalness. However, when the models are applied to synthesize code-switched (CS) speech, the performance declines seriously. Conventionally, developing a CS TTS system requires multilingual data to incorporate language-specific and cross-lingual knowledge. Recently, end-to-end (E2E) architecture has achieved satisfactory results in monolingual TTS. The architecture enables the training from one end of alphabetic text input to the other end of acoustic feature output. In this paper, we explore the use of E2E framework for CS TTS, using a combination of Mandarin and English monolingual speech corpus uttered by two female speakers. To handle alphabetic input from different languages, we explore two kinds of encoders: (1) shared multilingual encoder with explicit language embedding (LDE); (2) separated monolingual encoder (SPE) for each language. The two systems use identical decoder architecture, where a discriminative code is incorporated to enable the model to generate speech in one speaker's voice consistently. Experiments confirm the effectiveness of the proposed modifications on the E2E TTS framework in terms of quality and speaker similarity of the generated speech. Moreover, our proposed systems can generate controllable foreign-accented speech at character-level using only mixture of monolingual training data.

The hidden activation functions inside deep neural networks (DNNs) play a vital role in learning high level discriminative features and controlling the information flows to track longer history. However, the fixed model parameters used in standard DNNs can lead to over-fitting and poor generalization when given limited training data. Furthermore, the precise forms of activations used in DNNs are often manually set at a global level for all hidden nodes, thus lacking an automatic selection method. In order to address these issues, Bayesian neural networks (BNNs) acoustic models are proposed in this paper to explicitly model the uncertainty associated with DNN parameters. Gaussian Process (GP) activations based DNN and LSTM acoustic models are also used in this paper to allow the optimal forms of hidden activations to be stochastically learned for individual hidden nodes. An efficient variational inference based training algorithm is derived for BNN, GPNN and GPLSTM systems. Experiments were conducted on a LVCSR system trained on a 75 hour subset of Switchboard I data. The best BNN and GPNN systems outperformed both the baseline DNN systems constructed using fixed form activations and their combination via frame level joint decoding by 1% absolute in word error rate.

Recurrent neural network language models (RNNLMs) have shown superior performance across a range of speech recognition tasks. At the heart of all RNNLMs, the activation functions play a vital role to control the information flows and tracking longer history contexts that are useful for predicting the following words. Long short-term memory (LSTM) units are well known for such ability and thus widely used in current RNNLMs. However, the deterministic parameter estimates in LSTM RNNLMs are prone to over-fitting and poor generalization when given limited training data. Furthermore, the precise forms of activations in LSTM have been largely empirically set for all cells at a global level. In order to address these issues, this paper introduces Gaussian process (GP) LSTM RNNLMs. In addition to modeling parameter uncertainty under a Bayesian framework, it also allows the optimal forms of gates being automatically learned for individual LSTM cells. Experiments were conducted on three tasks: the Penn Treebank (PTB) corpus, Switchboard conversational telephone speech (SWBD) and the AMI meeting room data. The proposed GP-LSTM RNNLMs consistently outperform the baseline LSTM RNNLMs in terms of both perplexity and word error rate.

Speech emotion recognition (SER) is a fundamental step towards fluent human-machine interaction. One challenging problem in SER is obtaining utterance-level feature representation for classification. Recent works on SER have made significant progress by using spectrogram features and introducing neural network methods, e.g., convolutional neural networks (CNNs). However the fundamental problem of CNNs is that the spatial information in spectrograms is not captured, which are basically position and relationship information of low-level features like pitch and formant frequencies. This paper presents a novel architecture based on the capsule networks (CapsNets) for SER. The proposed system can take into account the spatial relationship of speech features in spectrograms, and provide an effective pooling method for obtaining utterance global features. We also introduce a recurrent connection to CapsNets to improve the model's time sensitivity. We compare the proposed model to previous published results based on combined CNN-long short-term memory (CNN-LSTM) models on the benchmark corpus IEMOCAP over four emotions, i.e., neutral, angry, happy and sad. Experimental results show that our model achieves better results than the baseline system on weighted accuracy (WA) (72.73% vs. 68.8%) and un-weighted accuracy (UA) (59.71% vs. 59.4%), which demonstrates the effectiveness of CapsNets for SER.

Recurrent neural network language models (RNNLMs) have become an increasing popular choice for state-of-the-art speech recognition systems. RNNLMs are normally trained by minimizing the cross entropy (CE) using the stochastic gradient descent (SGD) algorithm. However, the SGD method doesn't consider the correlation between parameters and therefore can lead to unstable and slow convergence in training. Second-order optimization methods provide a possible solution to this issue. However these methods are either computationally heavy or do not have competitive performance. In this paper, a novel optimization method - stochastic natural gradient based on minimum variance assumption (SNGM) is proposed for training RNNLMs. It allows the natural gradient method to operate at a comparable training efficiency to the SGD method. By modifying the gradient according to the local curvature of the KL-divergence between current and updated probabilistic distributions, the proposed SNGM approach is shown to outperform both the SGD and limited memory BFGS methods across three tasks: Penn Treebank, Switchboard conversational speech recognition and AMI meeting room transcription in terms of both perplexity and word error rate.

Recurrent neural network language models (RNNLM) have become an increasingly popular choice for state-of-the-art speech recognition systems. RNNLMs are normally trained by minimizing the cross entropy (CE) using the stochastic gradient descent (SGD) algorithm. The SGD method only uses first-order derivatives and no higher order gradient information is used to consider the correlation between model parameters. It is unable to fully capture the curvature of the error cost function. This can lead to slow convergence in model training. In this paper, a limited-memory Broyden Fletcher Goldfarb Shannon (L-BFGS) based second order optimization technique is proposed for RNNLMs. This method efficiently approximates the matrix-vector product between the inverse Hessian and gradient vector via a recursion over past gradients with a compact memory requirement. Consistent perplexity and error rate reductions are obtained over the SGD method on two speech recognition tasks: Switchboard English and Babel Cantonese. A faster convergence and speed up in RNNLM training time was also obtained.