Streaming end-to-end multi-talker speech recognition aims at transcribing the overlapped speech from conversations or meetings with an all-neural model in a streaming fashion, which is fundamentally different from a modular-based approach that usually cascades the speech separation and the speech recognition models trained independently. Previously, we proposed the Streaming Unmixing and Recognition Transducer (SURT) model based on recurrent neural network transducer (RNN-T) for this problem and presented promising results. However, for real applications, the speech recognition system is also required to determine the times-tamp when a speaker finishes speaking for prompt system response. This problem, known as endpoint (EP) detection, has not been studied previously for multi-talker end-to-end models. In this work, we address the EP detection problem in the SURT framework by introducing an end-of-sentence token as an output unit, following the practice of single-talker end-to-end models. Furthermore, we also present a latency penalty approach that can significantly cut down the EP detection latency. Our experimental results based on the 2-speaker LibrispeechMix dataset show that the SURT model can achieve promising EP detection without significantly degradation of the recognition accuracy.

Streaming recognition of multi-talker conversations has so far been evaluated only for 2-speaker single-turn sessions. In this paper, we investigate it for multi-turn meetings containing multiple speakers using the Streaming Unmixing and Recognition Transducer (SURT) model, and show that naively extending the single-turn model to this harder setting incurs a performance penalty. As a solution, we propose the dual-path (DP) modeling strategy first used for time-domain speech separation. We experiment with LSTM and Transformer based DP models, and show that they improve word error rate (WER) performance while yielding faster convergence. We also explore training strategies such as chunk width randomization and curriculum learning for these models, and demonstrate their importance through ablation studies. Finally, we evaluate our models on the LibriCSS meeting data, where they perform competitively with offline separation-based methods.

Recently, an end-to-end speaker-attributed automatic speech recognition (E2E SA-ASR) model was proposed as a joint model of speaker counting, speech recognition and speaker identification for monaural overlapped speech. In the previous study, the model parameters were trained based on the speaker-attributed maximum mutual information (SA-MMI) criterion, with which the joint posterior probability for multi-talker transcription and speaker identification are maximized over training data. Although SA-MMI training showed promising results for overlapped speech consisting of various numbers of speakers, the training criterion was not directly linked to the final evaluation metric, i.e., speaker-attributed word error rate (SA-WER). In this paper, we propose a speaker-attributed minimum Bayes risk (SA-MBR) training method where the parameters are trained to directly minimize the expected SA-WER over the training data. Experiments using the LibriSpeech corpus show that the proposed SA-MBR training reduces the SA-WER by 9.0 % relative compared with the SA-MMI-trained model.1

The efficacy of external language model (LM) integration with existing end-to-end (E2E) automatic speech recognition (ASR) systems can be improved significantly using the internal language model estimation (ILME) method [1]. In this method, the internal LM score is subtracted from the score obtained by interpolating the E2E score with the external LM score, during inference. To improve the ILME-based inference, we propose an internal LM training (ILMT) method to minimize an additional internal LM loss by updating only the E2E model components that affect the internal LM estimation. ILMT encourages the E2E model to form a standalone LM inside its existing components, without sacrificing ASR accuracy. After ILMT, the more modular E2E model with matched training and inference criteria enables a more thorough elimination of the source-domain internal LM, and therefore leads to a more effective integration of the target-domain external LM. Experimented with 30K-hour trained recurrent neural network transducer and attention-based encoder- decoder models, ILMT with ILME-based inference achieves up to 31.5% and 11.4% relative word error rate reductions from standard E2E training with Shallow Fusion on out-of-domain LibriSpeech and in-domain Microsoft production test sets, respectively.

In our previous research, we proposed a frame-level self-teaching network to regularize the deep neural network during training. In this paper, we extend the previous approach and propose a sequence self-teaching network to regularize the sequence-level information in speech recognition. The idea is to generate the sequence-level soft supervision labels from the top layer of the network to supervise the training of lower layer parameters. The network is trained with an auxiliary criterion in order to reduce the sequence-level Kullback-Leibler (KL) divergence between the top layer and lower layers, where the posterior probabilities in the KL-divergence term is computed from a lattice at the sequence-level. We evaluated the sequence-level self-teaching regularization approach with bidirectional long short-term memory models on LibriSpeech task, and show consistent improvements over the discriminative sequence maximum mutual information trained baseline.

In multi-talker scenarios such as meetings and conversations, speech processing systems are usually required to transcribe the audio as well as identify the speakers for downstream applications. Since overlapped speech is common in this case, conventional approaches usually address this problem in a cascaded fashion that involves speech separation, speech recognition and speaker identification that are trained independently. In this paper, we propose Streaming Unmixing, Recognition and Identification Transducer (SURIT) -- a new framework that deals with this problem in an end-to-end streaming fashion. SURIT employs the recurrent neural network transducer (RNN-T) as the backbone for both speech recognition and speaker identification. We validate our idea on the LibrispeechMix dataset -- a multi-talker dataset derived from Librispeech, and present encouraging results.

Hybrid Autoregressive Transducer (HAT) is a recently proposed end-to-end acoustic model that extends the standard Recurrent Neural Network Transducer (RNN-T) for the purpose of the external language model (LM) fusion.In HAT, the blank probability and the label probability are estimated using two separate probability distributions, which provides a more accurate solution for internal LM score estimation, and thus works better when combining with an external LM.Previous work mainly focuses on HAT model training with the negative log-likelihood loss, while in this paper, we study the minimum word error rate (MWER) training of HAT -a criterion that is closer to the evaluation metric for speech recognition, and has been successfully applied to other types of end-to-end models such as sequenceto-sequence (S2S) and RNN-T models.From experiments with around 30,000 hours of training data, we show that MWER training can improve the accuracy of HAT models, while at the same time, improving the robustness of the model against the decoding hyper-parameters such as length normalization and decoding beam during inference.

Integrating external language models (LMs) into end-to-end (E2E) models remains a challenging task for domain-adaptive speech recognition.Recently, internal language model estimation (ILME)-based LM fusion has shown significant word error rate (WER) reduction from Shallow Fusion by subtracting a weighted internal LM score from an interpolation of E2E model and external LM scores during beam search.However, on different test sets, the optimal LM interpolation weights vary over a wide range and have to be tuned extensively on well-matched validation sets.In this work, we perform LM fusion in the minimum WER (MWER) training of an E2E model to obviate the need for LM weights tuning during inference.Besides MWER training with Shallow Fusion (MWER-SF), we propose a novel MWER training with ILME (MWER-ILME) where the ILME-based fusion is conducted to generate N-best hypotheses and their posteriors.Additional gradient is induced when internal LM is engaged in MWER-ILME loss computation.During inference, LM weights pre-determined in MWER training enable robust LM integrations on test sets from different domains.Experimented with 30K-hour trained transformer transducers, MWER-ILME achieves on average 8.8% and 5.8% relative WER reductions from MWER and MWER-SF training, respectively, on 6 different test sets.

The external language models (LM) integration remains a challenging task for end-to-end (E2E) automatic speech recognition (ASR) which has no clear division between acoustic and language models. In this work, we propose an internal LM estimation (ILME) method to facilitate a more effective integration of the external LM with all pre-existing E2E models with no additional model training, including the most popular recurrent neural network transducer (RNN-T) and attention-based encoder-decoder (AED) models. Trained with audio-transcript pairs, an E2E model implicitly learns an internal LM that characterizes the training data in the source domain. With ILME, the internal LM scores of an E2E model are estimated and subtracted from the log-linear interpolation between the scores of the E2E model and the external LM. The internal LM scores are approximated as the output of an E2E model when eliminating its acoustic components. ILME can alleviate the domain mismatch between training and testing, or improve the multi-domain E2E ASR. Experimented with 30K-hour trained RNN-T and AED models, ILME achieves up to 15.5% and 6.8% relative word error rate reductions from Shallow Fusion on out-of-domain LibriSpeech and in-domain Microsoft production test sets, respectively.

End-to-end multi-talker speech recognition is an emerging research trend in the speech community due to its vast potential in applications such as conversation and meeting transcriptions. To the best of our knowledge, all existing research works are constrained in the offline scenario. In this work, we propose the Streaming Unmixing and Recognition Transducer (SURT) for end-to-end multi-talker speech recognition. Our model employs the Recurrent Neural Network Transducer (RNN-T) as the backbone that can meet various latency constraints. We study two different model architectures that are based on a speaker-differentiator encoder and a mask encoder respectively. To train this model, we investigate the widely used Permutation Invariant Training (PIT) approach and the Heuristic Error Assignment Training (HEAT) approach. Based on experiments on the publicly available LibriSpeechMix dataset, we show that HEAT can achieve better accuracy compared with PIT, and the SURT model with 150 milliseconds algorithmic latency constraint compares favorably with the offline sequence-to-sequence based baseline model in terms of accuracy.

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.

Recently, the recurrent neural network transducer (RNN-T) architecture has become an emerging trend in end-to-end automatic speech recognition research due to its advantages of being capable for online streaming speech recognition. However, RNN-T training is made difficult by the huge memory requirements, and complicated neural structure. A common solution to ease the RNN-T training is to employ connectionist temporal classification (CTC) model along with RNN language model (RNNLM) to initialize the RNN-T parameters. In this work, we conversely leverage external alignments to seed the RNN-T model. Two different pre-training solutions are explored, referred to as encoder pre-training, and whole-network pre-training respectively. Evaluated on Microsoft 65,000 hours anonymized production data with personally identifiable information removed, our proposed methods can obtain significant improvement. In particular, the encoder pre-training solution achieved a 10% and a 8% relative word error rate reduction when compared with random initialization and the widely used CTC+RNNLM initialization strategy, respectively. Our solutions also significantly reduce the RNN-T model latency from the baseline.

Recently, a few novel streaming attention-based sequence-to-sequence (S2S) models have been proposed to perform online speech recognition with linear-time decoding complexity. However, in these models, the decisions to generate tokens are delayed compared to the actual acoustic boundaries since their unidirectional encoders lack future information. This leads to an inevitable latency during inference. To alleviate this issue and reduce latency, we propose several strategies during training by leveraging external hard alignments extracted from the hybrid model. We investigate to utilize the alignments in both the encoder and the decoder. On the encoder side, (1) multi-task learning and (2) pre-training with the framewise classification task are studied. On the decoder side, we (3) remove inappropriate alignment paths beyond an acceptable latency during the alignment marginalization, and (4) directly min-imize the differentiable expected latency loss. Experiments on the Cortana voice search task demonstrate that our proposed methods can significantly reduce the latency, and even improve the recognition accuracy in certain cases on the decoder side. We also present some analysis to understand the behaviors of streaming S2S models.

Attention-based encoder-decoder model has achieved impressive results for both automatic speech recognition (ASR) and text-to-speech (TTS) tasks.This approach takes advantage of the memorization capacity of neural networks to learn the mapping from the input sequence to the output sequence from scratch, without the assumption of prior knowledge such as the alignments.However, this model is prone to overfitting, especially when the amount of training data is limited.Inspired by SpecAugment and BERT, in this paper, we propose a semantic mask based regularization for training such kind of end-toend (E2E) model.The idea is to mask the input features corresponding to a particular output token, e.g., a word or a wordpiece, in order to encourage the model to fill the token based on the contextual information.While this approach is applicable to the encoder-decoder framework with any type of neural network architecture, we study the transformer-based model for ASR in this work.We perform experiments on Librispeech 960h and TedLium2 data sets, and achieve the state-of-the-art performance on the test set in the scope of E2E models.

The attention-based Transformer model has achieved promising results for speech recognition (SR) in the offline mode.However, in the streaming mode, the Transformer model usually incurs significant latency to maintain its recognition accuracy when applying a fixed-length look-ahead window in each encoder layer.In this paper, we propose a novel low-latency streaming approach for Transformer models, which consists of a scout network and a recognition network.The scout network detects the whole word boundary without seeing any future frames, while the recognition network predicts the next subword by utilizing the information from all the frames before the predicted boundary.Our model achieves the best performance (2.7/6.4WER) with only 639 ms latency on the test-clean and test-other data sets of Librispeech.

While recurrent neural networks still largely define state-of-theart speech recognition systems, the Transformer network has been proven to be a competitive alternative, especially in the offline condition.Most studies with Transformers have been constrained in a relatively small scale setting, and some forms of data argumentation approaches are usually applied to combat the data sparsity issue.In this paper, we aim at understanding the behaviors of Transformers in the large-scale speech recognition setting, where we have used around 65,000 hours of training data.We investigated various aspects on scaling up Transformers, including model initialization, warmup training as well as different Layer Normalization strategies.In the streaming condition, we compared the widely used attention mask based future context lookahead approach to the Transformer-XL network.From our experiments, we show that Transformers can achieve around 6% relative word error rate (WER) reduction compared to the BLSTM baseline in the offline fashion, while in the streaming fashion, Transformer-XL is comparable to LC-BLSTM with 800 millisecond latency constraint.

Recent studies suggest that it may now be possible to construct end-to-end Neural Network (NN) models that perform on-par with, or even outperform, hybrid models in speech recognition. These models differ in their designs, and as such, may exhibit diverse and complementary error patterns. A combination between the predictions of these models may therefore yield significant gains. This paper studies the feasibility of performing hypothesis-level combination between hybrid and end-to-end NN models. The end-to-end NN models often exhibit a bias in their posteriors toward short hypotheses, and this may adversely affect Minimum Bayes’ Risk (MBR) combination methods. MBR training and length normalisation can be used to reduce this bias. Models are trained on Microsoft’s 75 thousand hours of anonymised data and evaluated on test sets with 1.8 million words. The results show that significant gains can be obtained by combining the hypotheses of hybrid and end-to-end NN models together.

Overlapped speech is one of the main challenges in conversational speech applications such as meeting transcription. Blind speech separation and speech extraction are two common approaches to this problem. Both of them, however, suffer from limitations resulting from the lack of abilities to either leverage additional information or process multiple speakers simultaneously. In this work, we propose a novel method called speech separation using speaker inventory (SSUSI), which combines the advantages of both approaches and thus solves their problems. SSUSI makes use of a speaker inventory, i.e. a pool of pre-enrolled speaker signals, and jointly separates all participating speakers. This is achieved by a specially designed attention mechanism, eliminating the need for accurate speaker identities. Experimental results show that SSUSI outperforms permutation invariant training based blind speech separation by up to 48% relatively in word error rate (WER). Compared with speech extraction, SSUSI reduces computation time by up to 70% and improves the WER by more than 13% relatively.

In our recent work, we proposed a layer trajectory long short-term memory (ltLSTM) model which decouples the tasks of temporal modeling and senone classification with time-LSTMs and depth-LSTMs. The ltLSTM model achieved significant accuracy improvement over the traditional multi-layer LSTM models from our previous study. Considering the future context frames carrying valuable information for predicting the target label evidenced by the success of bi-directional LSTMs, in this work we investigate how to incorporate this kind of information with hidden vectors from either time-LSTM or depth-LSTM. Trained with 30 thousand hours of EN-US Microsoft internal data, the best ltLSTM model with future context frames can improve the baseline ltLSTM with up to 11.5% relative word error rate (WER) reduction and improve the baseline LSTM with up to 24.6% relative WER reduction across different tasks.

We propose self-teaching networks to improve the generalization capacity of deep neural networks. The idea is to generate soft supervision labels using the output layer for training the lower layers of the network. During the network training, we seek an auxiliary loss that drives the lower layer to mimic the behavior of the output layer. The connection between the two network layers through the auxiliary loss can help the gradient flow, which works similar to the residual networks. Furthermore, the auxiliary loss also works as a regularizer, which improves the generalization capacity of the network. We evaluated the self-teaching network with deep recurrent neural networks on speech recognition tasks, where we trained the acoustic model using 30 thousand hours of data. We tested the acoustic model using data collected from 4 scenarios. We show that the self-teaching network can achieve consistent improvements and outperform existing methods such as label smoothing and confidence penalization.

Connectionist temporal classification (CTC) is a popular sequence prediction approach for automatic speech recognition that is typically used with models based on recurrent neural networks (RNNs). We explore whether deep convolutional neural networks (CNNs) can be used effectively instead of RNNs as the “encoder” in CTC. CNNs lack an explicit representation of the entire sequence, but have the advantage that they are much faster to train. We present an exploration of CNN s as encoders for CTC models, in the context of character-based (lexicon-free) automatic speech recognition. In particular, we explore a range of one-dimensional convolutionallayers, which are particularly efficient. We compare the performance of our CNN-based models against typical RNN-based models in terms of training time, decoding time, model size and word error rate (WER) on the Switchboard Eva12000 corpus. We find that our CNN-based models are close in performance to LSTMs, while not matching them, and are much faster to train and decode.

Traditional LSTM model and its variants normally work in a frame-by-frame and layer-by-layer fashion, which deals with the temporal modeling and target classification problems at the same time. In this paper, we extend our recently proposed layer trajectory LSTM (ltLSTM) and present a generalized framework, which is equipped with a depth processing block that scans the hidden states of each time-LSTM layer, and uses the summarized layer trajectory information for final senone classification. We explore different modeling units used in the depth processing block to have a good tradeoff between accuracy and runtime cost. Furthermore, we integrate an attention module into this framework to explore wide context information, which is especially beneficial for uni-directional LSTMs. Trained with 30 thousand hours of EN-US Microsoft internal data and cross entropy criterion, the proposed generalized ltLSTM performed significantly better than the standard multi-layer time-LSTM, with up to 12.8% relative word error rate (WER) reduction across different tasks. With attention modeling, the relative WER reduction can be up to 17.9%. We observed similar gain when the models were trained with sequence discriminative training criterion.

Deep learning has significantly advanced state-of-the-art of speech recognition in the past few years. However, compared to conventional Gaussian mixture acoustic models, neural network models are usually much larger, and are therefore not very deployable in embedded devices. Previously, we investigated a compact highway deep neural network (HDNN) for acoustic modelling, which is a type of depth-gated feedforward neural network. We have shown that HDNN-based acoustic models can achieve comparable recognition accuracy with much smaller number of model parameters compared to plain deep neural network (DNN) acoustic models. In this paper, we push the boundary further by leveraging on the knowledge distillation technique that is also known as teacher-student training, i.e., we train the compact HDNN model with the supervision of a high accuracy cumbersome model. Furthermore, we also investigate sequence training and adaptation in the context of teacher-student training. Our experiments were performed on the AMI meeting speech recognition corpus. With this technique, we significantly improved the recognition accuracy of the HDNN acoustic model with less than 0.8 million parameters, and narrowed the gap between this model and the plain DNN with 30 million parameters.

Segmental conditional random fields (SCRFs) and connectionist temporal classification (CTC) are two sequence labeling methods used for end-to-end training of speech recognition models. Both models define a transcription probability by marginalizing decisions about latent segmentation alternatives to derive a sequence probability: the former uses a globally normalized joint model of segment labels and durations, and the latter classifies each frame as either an output symbol or a "continuation" of the previous label. In this paper, we train a recognition model by optimizing an interpolation between the SCRF and CTC losses, where the same recurrent neural network (RNN) encoder is used for feature extraction for both outputs. We find that this multitask objective improves recognition accuracy when decoding with either the SCRF or CTC models. Additionally, we show that CTC can also be used to pretrain the RNN encoder, which improves the convergence rate when learning the joint model.

End-to-end training of deep learning-based models allows for implicit learning of intermediate representations based on the final task loss. However, the end-to-end approach ignores the useful domain knowledge encoded in explicit intermediate-level supervision. We hypothesize that using intermediate representations as auxiliary supervision at lower levels of deep networks may be a good way of combining the advantages of end-to-end training and more traditional pipeline approaches. We present experiments on conversational speech recognition where we use lower-level tasks, such as phoneme recognition, in a multitask training approach with an encoder-decoder model for direct character transcription. We compare multiple types of lower-level tasks and analyze the effects of the auxiliary tasks. Our results on the Switchboard corpus show that this approach improves recognition accuracy over a standard encoder-decoder model on the Eval2000 test set.

State-of-the-art speech recognition systems typically employ neural network acoustic models. However, compared to Gaussian mixture models, deep neural network (DNN) based acoustic models often have many more model parameters, making it challenging for them to be deployed on resource-constrained platforms, such as mobile devices. In this paper, we study the application of the recently proposed highway deep neural network (HDNN) for training small-footprint acoustic models. HDNNs are a depth-gated feedforward neural network, which include two types of gate functions to facilitate the information flow through different layers. Our study demonstrates that HDNNs are more compact than regular DNNs for acoustic modeling, i.e., they can achieve comparable recognition accuracy with many fewer model parameters. Furthermore, HDNNs are more controllable than DNNs: The gate functions of an HDNN can control the behavior of the whole network using a very small number of model parameters. Finally, we show that HDNNs are more adaptable than DNNs. For example, simply updating the gate functions using adaptation data can result in considerable gains in accuracy. We demonstrate these aspects by experiments using the publicly available AMI corpus, which has around 80 h of training data.

Recently, there has been an increasing interest in end-to-end speech recognition using neural networks, with no reliance on hidden Markov models (HMMs) for sequence modelling as in the standard hybrid framework. The recurrent neural network (RNN) encoderdecoder is such a model, performing sequence to sequence mapping without any predefined alignment. This model first transforms the input sequence into a fixed length vector representation, from which the decoder recovers the output sequence. In this paper, we extend our previous work on this model for large vocabulary end-to-end speech recognition. We first present a more effective stochastic gradient decent (SGD) learning rate schedule that can significantly improve the recognition accuracy. We then extend the decoder with long memory by introducing another recurrent layer that performs implicit language modelling. Finally, we demonstrate that using multiple recurrent layers in the encoder can reduce the word error rate. Our experiments were carried out on the Switchboard corpus using a training set of around 300 hours of transcribed audio data, and we have achieved significantly higher recognition accuracy, thereby reduced the gap compared to the hybrid baseline.

Long Short-Term Memory (LSTM) is a particular type of recurrent neural network (RNN) that can model long term temporal dynamics. Recently it has been shown that LSTM-RNNs can achieve higher recognition accuracy than deep feed-forword neural networks (DNNs) in acoustic modelling. However, speaker adaption for LSTM-RNN based acoustic models has not been well investigated. In this paper, we study the LSTM-RNN speaker-aware training that incorporates the speaker information during model training to normalise the speaker variability. We first present several speaker-aware training architectures, and then empirically evaluate three types of speaker representation: I-vectors, bottleneck speaker vectors and speaking rate. Furthermore, to factorize the variability in the acoustic signals caused by speakers and phonemes respectively, we investigate the speaker-aware and phone-aware joint training under the framework of multi-task learning. In AMI meeting speech transcription task, speaker-aware training of LSTM-RNNs reduces word error rates by 6.5% relative to a very strong LSTM-RNN baseline, which uses FMLLR features.

Despite the significant progress in speech recognition enabled by deep neural networks, poor performance persists in some scenarios. In this work, we focus on far-field speech recognition which remains challenging due to high levels of noise and reverberation in the captured speech signals. We propose to represent the stages of acoustic processing including beamforming, feature extraction, and acoustic modeling, as three components of a single unified computational network. The parameters of a frequency-domain beam-former are first estimated by a network based on features derived from the microphone channels. These filter coefficients are then applied to the array signals to form an enhanced signal. Conventional features are then extracted from this signal and passed to a second network that performs acoustic modeling for classification. The parameters of both the beamforming and acoustic modeling networks are trained jointly using back-propagation with a common cross-entropy objective function. In experiments on the AMI meeting corpus, we observed improvements by pre-training each sub-network with a network-specific objective function before joint training of both networks. The proposed method obtained a 3.2% absolute word error rate reduction compared to a conventional pipeline of independent processing stages.

We study the segmental recurrent neural network for end-to-end acoustic modelling. This model connects the segmental conditional random field (CRF) with a recurrent neural network (RNN) used for feature extraction. Compared to most previous CRF-based acoustic models, it does not rely on an external system to provide features or segmentation boundaries. Instead, this model marginalises out all the possible segmentations, and features are extracted from the RNN trained together with the segmental CRF. In essence, this model is self-contained and can be trained end-to-end. In this paper, we discuss practical training and decoding issues as well as the method to speed up the training in the context of speech recognition. We performed experiments on the TIMIT dataset. We achieved 17.3 phone error rate (PER) from the first-pass decoding --- the best reported result using CRFs, despite the fact that we only used a zeroth-order CRF and without using any language model.

For speech recognition, deep neural networks (DNNs) have significantly improved the recognition accuracy in most of benchmark datasets and application domains. However, compared to the conventional Gaussian mixture models, DNN-based acoustic models usually have much larger number of model parameters, making it challenging for their applications in resource constrained platforms, e.g., mobile devices. In this paper, we study the application of the recently proposed highway network to train small-footprint DNNs, which are {\it thinner} and {\it deeper}, and have significantly smaller number of model parameters compared to conventional DNNs. We investigated this approach on the AMI meeting speech transcription corpus which has around 70 hours of audio data. The highway neural networks constantly outperformed their plain DNN counterparts, and the number of model parameters can be reduced significantly without sacrificing the recognition accuracy.

Long Short-Term Memory (LSTM) networks, a type of recurrent neural network with a more complex computational unit, have been successfully applied to a variety of sequence modeling tasks.In this paper we develop Tree Long Short-Term Memory (TREELSTM), a neural network model based on LSTM, which is designed to predict a tree rather than a linear sequence.TREELSTM defines the probability of a sentence by estimating the generation probability of its dependency tree.At each time step, a node is generated based on the representation of the generated subtree.We further enhance the modeling power of TREELSTM by explicitly representing the correlations between left and right dependents.Application of our model to the MSR sentence completion challenge achieves results beyond the current state of the art.We also report results on dependency parsing reranking achieving competitive performance.

Acoustic models based on Gaussian mixture models (GMMs) typically use short span acoustic feature inputs. This does not capture long-term temporal information from speech owing to the conditional independence assumption of hidden Markov models. In this paper, we present an implicit approach that approximates the joint distribution of long span features by product of factorized models, in contrast to deep neural networks (DNNs) that model feature correlations directly. The approach is applicable to a broad range of acoustic models. We present experiments using GMM and probabilistic linear discriminant analysis (PLDA) based models on Switchboard, observing consistent word error rate reductions.

Probabilistic linear discriminant analysis (PLDA) acoustic models extend Gaussian mixture models by factorizing the acoustic variability using state-dependent and observation-dependent variables. This enables the use of higher dimensional acoustic features, and the capture of intra-frame feature corre-lations. In this paper, we investigate the estimation of speaker adaptive feature-space (constrained) maximum likelihood lin-ear regression transforms from PLDA-based acoustic models. This feature-space speaker transformation estimation approach is potentially very useful due to the ability of PLDA acoustic models to use different types of acoustic features, for example applying these transforms to deep neural network (DNN) acous-tic models for cross adaptation. We evaluated the approach on the Switchboard corpus, and observe significant word error re-duction by using both the mel-frequency cepstral coefficients and DNN bottleneck features. Index Terms: speech recognition, probabilistic linear discrim-inant analysis, speaker adaptation, fMLLR, PLDA

Deep neural networks have advanced the state-of-the-art in automatic speech recognition, when combined with hidden Markov models (HMMs). Recently there has been interest in using systems based on recurrent neural networks (RNNs) to perform sequence modelling directly, without the requirement of an HMM superstructure. In this paper, we study the RNN encoder-decoder approach for large vocabulary end-toend speech recognition, whereby an encoder transforms a sequence of acoustic vectors into a sequence of feature representations, from which a decoder recovers a sequence of words. We investigated this approach on the Switchboard corpus using a training set of around 300 hours of transcribed audio data. Without the use of an explicit language model or pronunciation lexicon, we achieved promising recognition accuracy, demonstrating that this approach warrants further investigation. Index Terms: end-to-end speech recognition, deep neural networks, recurrent neural networks, encoder-decoder.

We have recently proposed a new acoustic model based on probabilistic linear discriminant analysis (PLDA) which enjoys the flexibility of using higher dimensional acoustic features, and is more capable to capture the intra-frame feature correlations. In this paper, we investigate the use of bottleneck features obtained from a deep neural network (DNN) for the PLDA-based acoustic model. Experiments were performed on the Switchboard dataset — a large vocabulary conversational telephone speech corpus. We observe significant word error reduction by using the bottleneck features. In addition, we have also compared the PLDA-based acoustic model to three others using Gaussian mixture models (GMMs), subspace GMMs and hybrid deep neural networks (DNNs), and PLDA can achieve comparable or slightly higher recognition accuracy from our experiments.

In this letter, we propose a new acoustic modeling approach for automatic speech recognition based on probabilistic linear discriminant analysis (PLDA), which is used to model the state density function for the standard hidden Markov models (HMMs). Unlike the conventional Gaussian mixture models (GMMs) where the correlations are weakly modelled by using the diagonal covariance matrices, PLDA captures the correlations of feature vector in subspaces without vastly expanding the model. It also allows the usage of high dimensional feature input, and therefore is more flexible to make use of different type of acoustic features. We performed the preliminary experiments on the Switchboard corpus, and demonstrated the feasibility of this acoustic model.

This paper studies cross-lingual acoustic modeling in the context of subspace Gaussian mixture models (SGMMs). SGMMs factorize the acoustic model parameters into a set that is globally shared between all the states of a hidden Markov model (HMM) and another that is specific to the HMM states. We demonstrate that the SGMM global parameters are transferable between languages, particularly when the parameters are trained multilingually. As a result, acoustic models may be trained using limited amounts of transcribed audio by borrowing the SGMM global parameters from one or more source languages, and only training the state-specific parameters on the target language audio. Model regularization using ℓ1-norm penalty is shown to be particularly effective at avoiding overtraining and leading to lower word error rates. We investigate maximum a posteriori (MAP) adaptation of subspace parameters in order to reduce the mismatch between the SGMM global parameters of the source and target languages. In addition, monolingual and cross-lingual speaker adaptive training is used to reduce the model variance introduced by speakers. We have systematically evaluated these techniques by experiments on the GlobalPhone corpus.

Speech recognition systems normally use handcrafted pronunciation lexicons designed by linguistic experts. Building and maintaining such a lexicon is expensive and time consuming. This paper concerns automatically learning a pronunciation lexicon for speech recognition. We assume the availability of a small seed lexicon and then learn the pronunciations of new words directly from speech that is transcribed at word-level. We present two implementations for refining the putative pronunciations of new words based on acoustic evidence. The first one is an expectation maximization (EM) algorithm based on weighted finite state transducers (WFSTs) and the other is its Viterbi approximation. We carried out experiments on the Switchboard corpus of conversational telephone speech. The expert lexicon has a size of more than 30,000 words, from which we randomly selected 5,000 words to form the seed lexicon. By using the proposed lexicon learning method, we have significantly improved the accuracy compared with a lexicon learned using a grapheme-to-phoneme transformation, and have obtained a word error rate that approaches that achieved using a fully handcrafted lexicon.

Noise adaptive training (NAT) is an effective approach to normalise the environmental distortions in the training data. This paper investigates the model-based NAT scheme using joint uncertainty decoding (JUD) for subspace Gaussian mixture models (SGMMs). A typical SGMM acoustic model has much larger number of surface Gaussian components, which makes it computationally infeasible to compensate each Gaussian explicitly. JUD tackles the problem by sharing the compensation parameters among the Gaussians and hence reduces the computational and memory demands. For noise adaptive training, JUD is reformulated into a generative model, which leads to an efficient expectation-maximisation (EM) based algorithm to update the SGMM acoustic model parameters. We evaluated the SGMMs with NAT on the Aurora 4 database, and obtained higher recognition accuracy compared to systems without adaptive training. Index Terms: adaptive training, noise robustness, joint uncertainty decoding, subspace Gaussian mixture models

Joint uncertainty decoding (JUD) is a model-based noise compensation technique for conventional Gaussian Mixture Model (GMM) based speech recognition systems. Unlike vector Taylor series (VTS) compensation which operates on the individual Gaussian components in an acoustic model, JUD clusters the Gaussian components into a smaller number of classes, sharing the compensation parameters for the set of Gaussians in a given class. This significantly reduces the computational cost. In this paper, we investigate noise compensation for subspace Gaussian mixture model (SGMM) based speech recognition systems using JUD. The total number of Gaussian components in an SGMM is typically very large. Therefore direct compensation of the individual Gaussian components, as performed by VTS, is computationally expensive. In this paper we show that JUD-based noise compensation can be successfully applied to SGMMs in a computationally efficient way. We evaluate the JUD/SGMM technique on the standard Aurora 4 corpus. Our experimental results indicate that the JUD/SGMM system results in lower word error rates compared with a conventional GMM system with either VTS-based or JUD-based noise compensation.

This paper concerns cross-lingual acoustic modeling in the case when there are limited target language resources. We build on an approach in which a subspace Gaussian mixture model (SGMM) is adapted to the target language by reusing the globally shared parameters estimated from out-of-language training data. In current cross-lingual systems, these parameters are fixed when training the target system, which can give rise to a mismatch between the source and target systems. We investigate a maximum a posteriori (MAP) adaptation approach to alleviate the potential mismatch. In particular, we focus on the adaptation of phonetic subspace parameters using a matrix variate Gaussian prior distribution. Experiments on the GlobalPhone corpus using the MAP adaptation approach results in word error rate reductions, compared with the cross-lingual base-line systems and systems updated using maximum likelihood, for training conditions with 1 hour and 5 hours of target language data.

Joint uncertainty decoding (JUD) is an effective model-based noise compensation technique for conventional Gaussian mix-ture model (GMM) based speech recognition systems. In this paper, we apply JUD to subspace Gaussian mixture model (SGMM) based acoustic models. The total number of Gaus-sians in the SGMM acoustic model is usually much larger than for conventional GMMs, which limits the application of approaches which explicitly compensate each Gaussian, such as vector Taylor series (VTS). However, by clustering the Gaussian components into a number of regression classes, JUD-based noise compensation can be successfully applied to SGMM systems. We evaluate the JUD/SGMM technique us-ing the Aurora 4 corpus, and the experimental results indicated that it is more accurate than conventional GMM-based systems using either VTS or JUD noise compensation. 1.

We investigate cross-lingual acoustic modelling for low resource languages using the subspace Gaussian mixture model (SGMM). We assume the presence of acoustic models trained on multiple source languages, and use the global subspace parameters from those models for improved modelling in a target language with limited amounts of transcribed speech. Experiments on the GlobalPhone corpus using Spanish, Portuguese, and Swedish as source languages and German as target language (with 1 hour and 5 hours of transcribed audio) show that multilingually trained SGMM shared parameters result in lower word error rates (WERs) than using those from a single source language. We also show that regularizing the estimation of the SGMM state vectors by penalizing their ℓ1-norm help to overcome numerical instabilities and lead to lower WER.

Subspace Gaussian mixture models (SGMMs) provide a compact representation of the Gaussian parameters in an acoustic model, but may still suffer from over-fitting with insufficient training data. In this letter, the SGMM state parameters are estimated using a penalized maximum-likelihood objective, based on l1and l2regularization, as well as their combination, referred to as the elastic net, for robust model estimation. Experiments on the 5000-word Wall Street Journal transcription task show word error rate reduction and improved model robustness with regularization.