| Total: 37
The goal of this work is to build a state-of-the-art English conversational telephone speech recognition system. We investigated several techniques to improve acoustic modeling, namely speaker-dependent bottleneck features, deep Bidirectional Long Short-Term Memory (BLSTM) recurrent neural networks, data augmentation and score fusion of DNN and BLSTM models. Training set consisted of the 300 hour Switchboard English speech corpus. We also examined the hypothesis rescoring using language models based on recurrent neural networks. The resulting system achieves a word error rate of 7.8% on the Switchboard part of the HUB5 2000 evaluation set which is the competitive result.
We describe a collection of acoustic and language modeling techniques that lowered the word error rate of our English conversational telephone LVCSR system to a record 6.6% on the Switchboard subset of the Hub5 2000 evaluation testset. On the acoustic side, we use a score fusion of three strong models: recurrent nets with maxout activations, very deep convolutional nets with 3×3 kernels, and bidirectional long short-term memory nets which operate on FMLLR and i-vector features. On the language modeling side, we use an updated model “M” and hierarchical neural network LMs.
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 thinner and 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 80 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.
In this paper, we propose a deep convolutional neural network (CNN) with layer-wise context expansion and location-based attention, for large vocabulary speech recognition. In our model each higher layer uses information from broader contexts, along both the time and frequency dimensions, than its immediate lower layer. We show that both the layer-wise context expansion and the location-based attention can be implemented using the element-wise matrix product and the convolution operation. For this reason, contrary to other CNNs, no pooling operation is used in our model. Experiments on the 309hr Switchboard task and the 375hr short message dictation task indicates that our model outperforms both the DNN and LSTM significantly.
Recently neural network acoustic models trained with Connectionist Temporal Classification (CTC) were proposed as an alternative approach to conventional cross-entropy trained neural network acoustic models which output frame-level decisions every 10ms [1]. As opposed to conventional models, CTC learns an alignment jointly with the acoustic model, and outputs a blank symbol in addition to the regular acoustic state units. This allows the CTC model to run with a lower frame rate, outputting decisions every 30ms rather than 10ms as in conventional models, thus improving overall system speed. In this work, we explore how conventional models behave with lower frame rates. On a large vocabulary Voice Search task, we will show that with conventional models, we can slow the frame rate to 40ms while improving WER by 3% relative over a CTC-based model.
Neural Network (NN) Acoustic Models (AMs) are usually trained using context-dependent Hidden Markov Model (CD-HMM) states as independent targets. For example, the CD-HMM states of A-b-2 (second variant of beginning state of A) and A-m-1 (first variant of middle state of A) both correspond to the phone A, and A-b-1 and A-b-2 both correspond to the Context-independent HMM (CI-HMM) state A-b, but this relationship is not explicitly modeled. We propose a method that treats some neurons in the final hidden layer just below the output layer as dedicated neurons for phones or CI-HMM states by initializing connections between the dedicated neurons and the corresponding CD-HMM outputs with stronger weights than to other outputs. We obtained 6.5% and 3.6% relative error reductions with a DNN AM and a CNN AM, respectively, on a 50-hour English broadcast news task and 4.6% reduction with a CNN AM on a 500-hour Japanese task, in all cases after Hessian-free sequence training. Our proposed method only changes the NN parameter initialization and requires no additional computation in NN training or speech recognition run-time.
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. Essentially, 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.
Convolutional and bidirectional recurrent neural networks have achieved considerable performance gains as acoustic models in automatic speech recognition in recent years. Latest architectures unify long short-term memory, gated recurrent unit and convolutional neural networks by stacking these different neural network types on each other, and providing short and long-term features to different depth levels of the network. For the first time, we propose a unified layer for acoustic modeling which is simultaneously recurrent and convolutional, and which operates only on short-term features. Our unified model introduces a bidirectional gated recurrent unit that uses convolutional operations for the gating units. We analyze the performance behavior of the proposed layer, compare and combine it with bidirectional gated recurrent units, deep neural networks and frequency-domain convolutional neural networks on a 50 hour English broadcast news task. The analysis indicates that the proposed layer in combination with stacked bidirectional gated recurrent units outperforms other architectures.
Deep neural network models have achieved considerable success in a wide range of fields. Several architectures have been proposed to alleviate the vanishing gradient problem, and hence enable training of very deep networks. In the speech recognition area, convolutional neural networks, recurrent neural networks, and fully connected deep neural networks have been shown to be complimentary in their modeling capabilities. Combining all three components, called CLDNN, yields the best performance to date. In this paper, we extend the CLDNN model by introducing a highway connection between LSTM layers, which enables direct information flow from cells of lower layers to cells of upper layers. With this design, we are able to better exploit the advantages of a deeper structure. Experiments on the GALE Chinese Broadcast Conversation/News Speech dataset indicate that our model outperforms all previous models and achieves a new benchmark, which is 22.41% character error rate on the dataset.
Deep neural networks (DNNs) and deep learning approaches yield state-of-the-art performance in a range of tasks, including speech recognition. However, the parameters of the network are hard to analyze, making network regularization and robust adaptation challenging. Stimulated training has recently been proposed to address this problem by encouraging the node activation outputs in regions of the network to be related. This kind of information aids visualization of the network, but also has the potential to improve regularization and adaptation. This paper investigates stimulated training of DNNs for both of these options. These schemes take advantage of the smoothness constraints that stimulated training offers. The approaches are evaluated on two large vocabulary speech recognition tasks: a U.S. English broadcast news (BN) task and a Javanese conversational telephone speech task from the IARPA Babel program. Stimulated DNN training acquires consistent performance gains on both tasks over unstimulated baselines. On the BN task, the proposed smoothing approach is also applied to rapid adaptation, again outperforming the standard adaptation scheme.
This paper proposes an approach, named phonetic context embedding, to model phonetic context effects for deep neural network - hidden Markov model (DNN-HMM) phone recognition. Phonetic context embeddings can be regarded as continuous and distributed vector representations of context-dependent phonetic units (e.g., triphones). In this work they are computed using neural networks. First, all phone labels are mapped into vectors of binary distinctive features (DFs, e.g., nasal/not-nasal). Then for each speech frame the corresponding DF vector is concatenated with DF vectors of previous and next frames and fed into a neural network that is trained to estimate the acoustic coefficients (e.g., MFCCs) of that frame. The values of the first hidden layer represent the embedding of the input DF vectors. Finally, the resulting embeddings are used as secondary task targets in a multi-task learning (MTL) setting when training the DNN that computes phone state posteriors. The approach allows to easily encode a much larger context than alternative MTL-based approaches. Results on TIMIT with a fully connected DNN shows phone error rate (PER) reductions from 22.4% to 21.0% and from 21.3% to 19.8% on the test core and the validation set respectively and lower PER than an alternative strong MTL approach.
Convolutional Neural Networks (CNNs) are effective models for reducing spectral variations and modeling spectral correlations in acoustic features for automatic speech recognition (ASR). Hybrid speech recognition systems incorporating CNNs with Hidden Markov Models/Gaussian Mixture Models (HMMs/GMMs) have achieved the state-of-the-art in various benchmarks. Meanwhile, Connectionist Temporal Classification (CTC) with Recurrent Neural Networks (RNNs), which is proposed for labeling unsegmented sequences, makes it feasible to train an ‘end-to-end’ speech recognition system instead of hybrid settings. However, RNNs are computationally expensive and sometimes difficult to train. In this paper, inspired by the advantages of both CNNs and the CTC approach, we propose an end-to-end speech framework for sequence labeling, by combining hierarchical CNNs with CTC directly without recurrent connections. By evaluating the approach on the TIMIT phoneme recognition task, we show that the proposed model is not only computationally efficient, but also competitive with the existing baseline systems. Moreover, we argue that CNNs have the capability to model temporal correlations with appropriate context information.
We assume that only word pairs identified by human are available in a low-resource target language. The word pairs are parameterized by a bottleneck feature (BNF) extractor that is trained using transcribed data in a high-resource language. The cross-lingual BNFs of the word pairs are used for training another neural network to generate a new feature representation in the target language. Pairwise learning of frame-level and word-level feature representations are investigated. Our proposed feature representations were evaluated in a word discrimination task on the Switchboard telephone speech corpus. Our learned features could bring 27.5% relative improvement over the previously best reported result on the task.
In this paper we investigate neural graph embeddings as front-end features for various deep neural network (DNN) architectures for speech recognition. Neural graph embedding features are produced by an autoencoder that maps graph structures defined over speech samples to a continuous vector space. The resulting feature representation is then used to augment the standard acoustic features at the input level of a DNN classifier. We compare two different neural graph embedding methods, one based on a local neighborhood graph encoding, and another based on a global similarity graph encoding. They are evaluated in DNN-HMM-based and LSTM-CTC-based ASR systems on a 110-hour Switchboard conversational speech recognition task. Significant improvements in word error rates are achieved by both methods in the DNN-HMM system, and by global graph embeddings in the LSTM-CTC system.
Articulatory features provide robustness to speaker and environment variability by incorporating speech production knowledge. Pseudo articulatory features are a way of extracting articulatory features using articulatory classifiers trained from speech data. One of the major problems faced in building articulatory classifiers is the requirement of speech data aligned in terms of articulatory feature values at frame level. Manually aligning data at frame level is a tedious task and alignments obtained from the phone alignments using phone-to-articulatory feature mapping are prone to errors. In this paper, a technique using connectionist temporal classification (CTC) criterion to train an articulatory classifier using bidirectional long short-term memory (BLSTM) recurrent neural network (RNN) is proposed. The CTC criterion eliminates the need for forced frame level alignments. Articulatory classifiers were also built using different neural network architectures like deep neural networks (DNN), convolutional neural network (CNN) and BLSTM with frame level alignments and were compared to the proposed approach of using CTC. Among the different architectures, articulatory features extracted using articulatory classifiers built with BLSTM gave better recognition performance. Further, the proposed approach of BLSTM with CTC gave the best overall performance on both SVitchboard (6 hours) and Switchboard 33 hours data set.
Deep neural networks (DNNs) are widely utilized for acoustic modeling in speech recognition systems. Through training, DNNs used for phoneme recognition nonlinearly transform the time-frequency representation of a speech signal into a sequence of invariant phonemic categories. However, little is known about how this nonlinear mapping is performed and what its implications are for the classification of individual phones and phonemic categories. In this paper, we analyze a sigmoid DNN trained for a phoneme recognition task and characterized several aspects of the nonlinear transformations that occur in hidden layers. We show that the function learned by deeper hidden layers becomes increasingly nonlinear, and that network selectively warps the feature space so as to increase the discriminability of acoustically similar phones, aiding in their classification. We also demonstrate that the nonlinear transformation of the feature space in deeper layers is more dedicated to the phone instances that are more difficult to discriminate, while the more separable phones are dealt with in the superficial layers of the network. This study describes how successive nonlinear transformations are applied to the feature space non-uniformly when a deep neural network model learns categorical boundaries, which may partly explain their superior performance in pattern classification applications.
State-of-the-art automatic speech recognition (ASR) systems typically rely on pre-processed features. This paper studies the time-frequency duality in ASR feature extraction methods and proposes extending the standard acoustic model with a complex-valued linear projection layer to learn and optimize features that minimize standard cost functions such as cross-entropy. The proposed Complex Linear Projection (CLP) features achieve superior performance compared to pre-processed Log Mel features.
Various neural network architectures have been proposed in the literature to model 2D correlations in the input signal, including convolutional layers, frequency LSTMs and 2D LSTMs such as time-frequency LSTMs, grid LSTMs and ReNet LSTMs. It has been argued that frequency LSTMs can model translational variations similar to CNNs, and 2D LSTMs can model even more variations [1], but no proper comparison has been done for speech tasks. While convolutional layers have been a popular technique in speech tasks, this paper compares convolutional and LSTM architectures to model time-frequency patterns as the first layer in an LDNN [2] architecture. This comparison is particularly interesting when the convolutional layer degrades performance, such as in noisy conditions or when the learned filterbank is not constant-Q [3]. We find that grid-LDNNs offer the best performance of all techniques, and provide between a 1–4% relative improvement over an LDNN and CLDNN on 3 different large vocabulary Voice Search tasks.
While DNN-HMM acoustic models have replaced GMM-HMMs in the standard ASR pipeline due to performance improvements, one unrealistic assumption that remains in these models is the conditional independence assumption of the Hidden Markov Model (HMM). In this work, we explore the extent to which depth of neural networks helps compensate for these poor conditional independence assumptions. Using a bootstrap resampling framework that allows us to control the amount of data dependence in the test set while still using real observations from the data, we can determine how robust neural networks, and particularly deeper models, are to data dependence. Our conclusions are that if the data were to match the conditional independence assumptions of the HMM, there would be little benefit from using deeper models. It is only when data become more dependent that depth improves ASR performance. That performance substantially degrades, however, as the data becomes more realistic suggests that better temporal modeling is still needed for ASR.
In this paper, we propose a novel approach for weakly-supervised word recognition. Most state of the art automatic speech recognition systems are based on frame-level labels obtained through forced alignments or through a sequential loss. Recently, weakly-supervised trained models have been proposed in vision, that can learn which part of the input is relevant for classifying a given pattern [1]. Our system is composed of a convolutional neural network and a temporal score aggregation mechanism. For each sentence, it is trained using as supervision only some of the words (most frequent) that are present in a given sentence, without knowing their order nor quantity. We show that our proposed system is able to jointly classify and localize words. We also evaluate the system on a keyword spotting task, and show that it can yield similar performance to strong supervised HMM/GMM baseline.
In this paper we present a simple and computationally efficient quantization scheme that enables us to reduce the resolution of the parameters of a neural network from 32-bit floating point values to 8-bit integer values. The proposed quantization scheme leads to significant memory savings and enables the use of optimized hardware instructions for integer arithmetic, thus significantly reducing the cost of inference. Finally, we propose a ‘quantization aware’ training process that applies the proposed scheme during network training and find that it allows us to recover most of the loss in accuracy introduced by quantization. We validate the proposed techniques by applying them to a long short-term memory-based acoustic model on an open-ended large vocabulary speech recognition task.
In this paper we describe a method to perform sequence-discriminative training of neural network acoustic models without the need for frame-level cross-entropy pre-training. We use the lattice-free version of the maximum mutual information (MMI) criterion: LF-MMI. To make its computation feasible we use a phone n-gram language model, in place of the word language model. To further reduce its space and time complexity we compute the objective function using neural network outputs at one third the standard frame rate. These changes enable us to perform the computation for the forward-backward algorithm on GPUs. Further the reduced output frame-rate also provides a significant speed-up during decoding. We present results on 5 different LVCSR tasks with training data ranging from 100 to 2100 hours. Models trained with LF-MMI provide a relative word error rate reduction of ~11.5%, over those trained with cross-entropy objective function, and ~8%, over those trained with cross-entropy and sMBR objective functions. A further reduction of ~2.5%, relative, can be obtained by fine tuning these models with the word-lattice based sMBR objective function.
We explore virtual adversarial training (VAT) applied to neural higher-order conditional random fields for sequence labeling. VAT is a recently introduced regularization method promoting local distributional smoothness: It counteracts the problem that predictions of many state-of-the-art classifiers are unstable to adversarial perturbations. Unlike random noise, adversarial perturbations are minimal and bounded perturbations that flip the predicted label. We utilize VAT to regularize neural higher-order factors in conditional random fields. These factors are for example important for phone classification where phone representations strongly depend on the context phones. However, without using VAT for regularization, the use of such factors was limited as they were prone to overfitting. In extensive experiments, we successfully apply VAT to improve performance on the TIMIT phone classification task. In particular, we achieve a phone error rate of 13.0%, exceeding the state-of-the-art performance by a wide margin.
The performance of automatic speech recognition can often be significantly improved by combining multiple systems together. Though beneficial, ensemble methods can be computationally expensive, often requiring multiple decoding runs. An alternative approach, appropriate for deep learning schemes, is to adopt student-teacher training. Here, a student model is trained to reproduce the outputs of a teacher model, or ensemble of teachers. The standard approach is to train the student model on the frame posterior outputs of the teacher. This paper examines the interaction between student-teacher training schemes and sequence training criteria, which have been shown to yield significant performance gains over frame-level criteria. There are several possible options for integrating sequence training, including training of the ensemble and further training of the student. This paper also proposes an extension to the student-teacher framework, where the student is trained to emulate the hypothesis posterior distribution of the teacher, or ensemble of teachers. This sequence student-teacher training approach allows the benefit of student-teacher training to be directly combined with sequence training schemes. These approaches are evaluated on two speech recognition tasks: a Wall Street Journal based task and a low-resource Tok Pisin conversational telephone speech task from the IARPA Babel programme.
We investigate speech-to-speech translation where one language does not have a well-defined written form. We use English-Spanish and Mandarin-English bitext corpora in order to provide both gold-standard text-based translations and experimental results for different levels of automatically derived symbolic representations from speech. We constrain our experiments such that the methods developed can be extended to low-resource languages. We derive different phonetic representations of the source texts in order to model the kinds of transcriptions that can be learned from low-resource-language speech data. We experiment with different methods of clustering the elements of the phonetic representations together into word-like units. We train MT models on the resulting texts, and report BLEU scores for the different representations and clustering methods in order to compare their effectiveness. Finally, we discuss our findings and suggest avenues for future research.