| Total: 114
On small datasets, discriminatively trained bottleneck features from deep networks commonly outperform more traditional spectral or cepstral features. While these features are typically trained with small, fully-connected networks, recent studies have used more sophisticated networks with great success. We use the recent deep CNN (VGG) network for bottleneck feature extraction — previously used only for low-resource tasks — and apply it to the Switchboard English conversational telephone speech task. Unlike features derived from traditional MLP networks, the VGG features outperform cepstral features even when used with BLSTM acoustic models trained on large amounts of data. We achieve the best BBN single system performance when combining the VGG features with a BLSTM acoustic model. When decoding with an n-gram language model, which are used for deployable systems, we have a realistic production system with a WER of 7.4%. This result is competitive with the current state-of-the-art in the literature. While our focus is on realistic single system performance, we further reduce the WER to 6.1% through system combination and using expensive neural network language model rescoring.
Student-teacher training allows a large teacher model or ensemble of teachers to be compressed into a single student model, for the purpose of efficient decoding. However, current approaches in automatic speech recognition assume that the state clusters, often defined by Phonetic Decision Trees (PDT), are the same across all models. This limits the diversity that can be captured within the ensemble, and also the flexibility when selecting the complexity of the student model output. This paper examines an extension to student-teacher training that allows for the possibility of having different PDTs between teachers, and also for the student to have a different PDT from the teacher. The proposal is to train the student to emulate the logical context dependent state posteriors of the teacher, instead of the frame posteriors. This leads to a method of mapping frame posteriors from one PDT to another. This approach is evaluated on three speech recognition tasks: the Tok Pisin and Javanese low resource conversational telephone speech tasks from the IARPA Babel programme, and the HUB4 English broadcast news task.
An embedding-based speaker adaptive training (SAT) approach is proposed and investigated in this paper for deep neural network acoustic modeling. In this approach, speaker embedding vectors, which are a constant given a particular speaker, are mapped through a control network to layer-dependent element-wise affine transformations to canonicalize the internal feature representations at the output of hidden layers of a main network. The control network for generating the speaker-dependent mappings are jointly estimated with the main network for the overall speaker adaptive acoustic modeling. Experiments on large vocabulary continuous speech recognition (LVCSR) tasks show that the proposed SAT scheme can yield superior performance over the widely-used speaker-aware training using i-vectors with speaker-adapted input features.
This paper reports our recent progress on using multilingual data for improving speech-to-text (STT) systems that can be easily delivered. We continued the work BBN conducted on the use of multilingual data for improving Babel evaluation systems, but focused on training time-delay neural network (TDNN) based chain models. As done for the Babel evaluations, we used multilingual data in two ways: first, to train multilingual deep neural networks (DNN) for extracting bottle-neck (BN) features, and second, for initializing training on target languages. Our results show that TDNN chain models trained on multilingual DNN bottleneck features yield significant gains over their counterparts trained on MFCC plus i-vector features. By initializing from models trained on multilingual data, TDNN chain models can achieve great improvements over random initializations of the network weights on target languages. Two other important findings are: 1) initialization with multilingual TDNN chain models produces larger gains on target languages that have less training data; 2) inclusion of target languages in multilingual training for either BN feature extraction or initialization have limited impact on performance measured on the target languages. Our results also reveal that for TDNN chain models, the combination of multilingual BN features and multilingual initialization achieves the best performance on all target languages.
Word error rates on the Switchboard conversational corpus that just a few years ago were 14% have dropped to 8.0%, then 6.6% and most recently 5.8%, and are now believed to be within striking range of human performance. This then raises two issues: what is human performance, and how far down can we still drive speech recognition error rates? In trying to assess human performance, we performed an independent set of measurements on the Switchboard and CallHome subsets of the Hub5 2000 evaluation and found that human accuracy may be considerably better than what was earlier reported, giving the community a significantly harder goal to achieve. We also report on our own efforts in this area, presenting a set of acoustic and language modeling techniques that lowered the WER of our system to 5.5%/10.3% on these subsets, which is a new performance milestone (albeit not at what we measure to be human performance). On the acoustic side, we use a score fusion of one LSTM with multiple feature inputs, a second LSTM trained with speaker-adversarial multi-task learning and a third convolutional residual net (ResNet). On the language modeling side, we use word and character LSTMs and convolutional WaveNet-style language models.
Recent work in automatic recognition of conversational telephone speech (CTS) has achieved accuracy levels comparable to human transcribers, although there is some debate how to precisely quantify human performance on this task, using the NIST 2000 CTS evaluation set. This raises the question what systematic differences, if any, may be found differentiating human from machine transcription errors. In this paper we approach this question by comparing the output of our most accurate CTS recognition system to that of a standard speech transcription vendor pipeline. We find that the most frequent substitution, deletion and insertion error types of both outputs show a high degree of overlap. The only notable exception is that the automatic recognizer tends to confuse filled pauses (“uh”) and backchannel acknowledgments (“uhhuh”). Human tend not to make this error, presumably due to the distinctive and opposing pragmatic functions attached to these words. Furthermore, we quantify the correlation between human and machine errors at the speaker level, and investigate the effect of speaker overlap between training and test data. Finally, we report on an informal “Turing test” asking humans to discriminate between automatic and human transcription error cases.
We describe the structure and application of an acoustic room simulator to generate large-scale simulated data for training deep neural networks for far-field speech recognition. The system simulates millions of different room dimensions, a wide distribution of reverberation time and signal-to-noise ratios, and a range of microphone and sound source locations. We start with a relatively clean training set as the source and artificially create simulated data by randomly sampling a noise configuration for every new training example. As a result, the acoustic model is trained using examples that are virtually never repeated. We evaluate performance of this approach based on room simulation using a factored complex Fast Fourier Transform (CFFT) acoustic model introduced in our earlier work, which uses CFFT layers and LSTM AMs for joint multichannel processing and acoustic modeling. Results show that the simulator-driven approach is quite effective in obtaining large improvements not only in simulated test conditions, but also in real / rerecorded conditions. This room simulation system has been employed in training acoustic models including the ones for the recently released Google Home.
In this paper, we propose a novel speech dereverberation framework that utilizes deep neural network (DNN)-based spectrum estimation to construct linear inverse filters. The proposed dereverberation framework is based on the state-of-the-art inverse filter estimation algorithm called weighted prediction error (WPE) algorithm, which is known to effectively reduce reverberation and greatly boost the ASR performance in various conditions. In WPE, the accuracy of the inverse filter estimation, and thus the dereverberation performance, is largely dependent on the estimation of the power spectral density (PSD) of the target signal. Therefore, the conventional WPE iteratively performs the inverse filter estimation, actual dereverberation and the PSD estimation to gradually improve the PSD estimate. However, while such iterative procedure works well when sufficiently long acoustically-stationary observed signals are available, WPE’s performance degrades when the duration of observed/accessible data is short, which typically is the case for real-time applications using online block-batch processing with small batches. To solve this problem, we incorporate the DNN-based spectrum estimator into the framework of WPE, because a DNN can estimate the PSD robustly even from very short observed data. We experimentally show that the proposed framework outperforms the conventional WPE, and improves the ASR performance in real noisy reverberant environments in both single-channel and multichannel cases.
The assumed scenario is transcription of a face-to-face conversation, such as in the financial industry when an agent and a customer talk over a desk with microphones placed between the speakers and then it is transcribed. From the automatic speech recognition (ASR) perspective, one of the speakers is the target speaker, and the other speaker is a directional noise source. When the number of microphones is small, we often accept microphone intervals that are larger than the spatial aliasing limit because the performance of the beamformer is better. Unfortunately, such a configuration results in significant leakage of directional noise in certain frequency bands because the spatial aliasing makes the beamformer and post-filter inaccurate there. Thus, we introduce a factorial model to compensate only the degraded bands with information from the reliable bands in a probabilistic framework integrating our proposed metrics and speech model. In our experiments, the proposed method reduced the errors from 29.8% to 24.9%.
We design a novel deep learning framework for multi-channel speech recognition in two aspects. First, for the front-end, an iterative mask estimation (IME) approach based on deep learning is presented to improve the beamforming approach based on the conventional complex Gaussian mixture model (CGMM). Second, for the back-end, deep convolutional neural networks (DCNNs), with augmentation of both noisy and beamformed training data, are adopted for acoustic modeling while the forward and backward long short-term memory recurrent neural networks (LSTM-RNNs) are used for language modeling. The proposed framework can be quite effective to multi-channel speech recognition with random combinations of fixed microphones. Testing on the CHiME-4 Challenge speech recognition task with a single set of acoustic and language models, our approach achieves the best performance of all three tracks (1-channel, 2-channel, and 6-channel) among submitted systems.
This paper describes the technical and system building advances made to the Google Home multichannel speech recognition system, which was launched in November 2016. Technical advances include an adaptive dereverberation frontend, the use of neural network models that do multichannel processing jointly with acoustic modeling, and Grid-LSTMs to model frequency variations. On the system level, improvements include adapting the model using Google Home specific data. We present results on a variety of multichannel sets. The combination of technical and system advances result in a reduction of WER of 8–28% relative compared to the current production system.
Recognition of distant (far-field) speech is a challenge for ASR due to mismatch in recording conditions resulting from room reverberation and environment noise. Given the remarkable learning capacity of deep neural networks, there is increasing interest to address this problem by using a large corpus of reverberant far-field speech to train robust models. In this study, we explore how an end-to-end RNN acoustic model trained on speech from different rooms and acoustic conditions (different domains) achieves robustness to environmental variations. It is shown that the first hidden layer acts as a domain separator, projecting the data from different domains into different subspaces. The subsequent layers then use this encoded domain knowledge to map these features to final representations that are invariant to domain change. This mechanism is closely related to noise-aware or room-aware approaches which append manually-extracted domain signatures to the input features. Additionally, we demonstrate how this understanding of the learning procedure provides useful guidance for model adaptation to new acoustic conditions. We present results based on AMI corpus to demonstrate the propagation of domain information in a deep RNN, and perform recognition experiments which indicate the role of encoded domain knowledge on training and adaptation of RNN acoustic models.
The shared-hidden-layer multilingual deep neural network (SHL-MDNN), in which the hidden layers of feed-forward deep neural network (DNN) are shared across multiple languages while the softmax layers are language dependent, has been shown to be effective on acoustic modeling of multilingual low-resource speech recognition. In this paper, we propose that the shared-hidden-layer with Long Short-Term Memory (LSTM) recurrent neural networks can achieve further performance improvement considering LSTM has outperformed DNN as the acoustic model of automatic speech recognition (ASR). Moreover, we reveal that shared-hidden-layer multilingual LSTM (SHL-MLSTM) with residual learning can yield additional moderate but consistent gain from multilingual tasks given the fact that residual learning can alleviate the degradation problem of deep LSTMs. Experimental results demonstrate that SHL-MLSTM can relatively reduce word error rate (WER) by 2.1–6.8% over SHL-MDNN trained using six languages and 2.6–7.3% over monolingual LSTM trained using the language specific data on CALLHOME datasets. Additional WER reduction, about relatively 2% over SHL-MLSTM, can be obtained through residual learning on CALLHOME datasets, which demonstrates residual learning is useful for SHL-MLSTM on multilingual low-resource ASR.
Phone-sized acoustic units such as triphones cannot properly capture the long-term co-articulation effects that occur in spontaneous speech. For that reason, it is interesting to construct acoustic units covering a longer time-span such as syllables or words. Unfortunately, the frequency distribution of those units is such that a few high frequency units account for most of the tokens, while many units rarely occur. As a result, those units suffer from data sparsity and can be difficult to train. In this paper we propose a scalable data-driven approach to construct a set of salient units made of sequences of phones called M-phones. We illustrate that since the decomposition of a word sequence into a sequence of M-phones is ambiguous, those units are well suited to be used with a connectionist temporal classification (CTC) approach which does not rely on an explicit frame-level segmentation of the word sequence into a sequence of acoustic units. Experiments are presented on a Voice Search task using 12,500 hours of training data.
Different training and adaptation techniques for multilingual Automatic Speech Recognition (ASR) are explored in the context of hybrid systems, exploiting Deep Neural Networks (DNN) and Hidden Markov Models (HMM). In multilingual DNN training, the hidden layers (possibly extracting bottleneck features) are usually shared across languages, and the output layer can either model multiple sets of language-specific senones or one single universal IPA-based multilingual senone set. Both architectures are investigated, exploiting and comparing different language adaptive training (LAT) techniques originating from successful DNN-based speaker-adaptation. More specifically, speaker adaptive training methods such as Cluster Adaptive Training (CAT) and Learning Hidden Unit Contribution (LHUC) are considered. In addition, a language adaptive output architecture for IPA-based universal DNN is also studied and tested. Experiments show that LAT improves the performance and adaptation on the top layer further improves the accuracy. By combining state-level minimum Bayes risk (sMBR) sequence training with LAT, we show that a language adaptively trained IPA-based universal DNN outperforms a monolingually sequence trained model.
The paper provides an analysis of BUT automatic speech recognition systems (ASR) built for the 2016 IARPA Babel evaluation. The IARPA Babel program concentrates on building ASR system for many low resource languages, where only a limited amount of transcribed speech is available for each language. In such scenario, we found essential to train the ASR systems in a multilingual fashion. In this work, we report superior results obtained with pre-trained multilingual BLSTM acoustic models, where we used multi-task training with separate classification layer for each language. The results reported on three Babel Year 4 languages show over 3% absolute WER reductions obtained from such multilingual pre-training. Experiments with different input features show that the multilingual BLSTM performs the best with simple log-Mel-filter-bank outputs, which makes our previously successful multilingual stack bottleneck features with CMLLR adaptation obsolete. Finally, we experiment with different configurations of i-vector based speaker adaptation in the mono- and multi-lingual BLSTM architectures. This results in additional WER reductions over 1% absolute.
Speech enhancement directly using deep neural network (DNN) is of major interest due to the capability of DNN to tangibly reduce the impact of noisy conditions in speech recognition tasks. Similarly, DNN based acoustic model adaptation to new environmental conditions is another challenging topic. In this paper we present an analysis of acoustic model adaptation in presence of a disjoint speech enhancement component, identifying an optimal setting for improving the speech recognition performance. Adaptation is derived from a consolidated technique that introduces in the training process a regularization term to prevent overfitting. We propose to optimize the adaptation of the clean acoustic models towards the enhanced speech by tuning the regularization term based on the degree of enhancement. Experiments on a popular noisy dataset (e.g., AURORA-4) demonstrate the validity of the proposed approach.
Recently, speaker adaptation methods in deep neural networks (DNNs) have been widely studied for automatic speech recognition. However, almost all adaptation methods for DNNs have to consider various heuristic conditions such as mini-batch sizes, learning rate scheduling, stopping criteria, and initialization conditions because of the inherent property of a stochastic gradient descent (SGD)-based training process. Unfortunately, those heuristic conditions are hard to be properly tuned. To alleviate those difficulties, in this paper, we propose a least squares regression-based speaker adaptation method in a DNN framework utilizing posterior mean of each class. Also, we show how the proposed method can provide a unique solution which is quite easy and fast to calculate without SGD. The proposed method was evaluated in the TED-LIUM corpus. Experimental results showed that the proposed method achieved up to a 4.6% relative improvement against a speaker independent DNN. In addition, we report further performance improvement of the proposed method with speaker-adapted features.
It is challenging to obtain large amounts of native (matched) labels for audio in under-resourced languages. This could be due to a lack of literate speakers of the language or a lack of universally acknowledged orthography. One solution is to increase the amount of labeled data by using mismatched transcription, which employs transcribers who do not speak the language (in place of native speakers), to transcribe what they hear as nonsense speech in their own language (e.g., Mandarin). This paper presents a multi-task learning framework where the DNN acoustic model is simultaneously trained using both a limited amount of native (matched) transcription and a larger set of mismatched transcription. We find that by using a multi-task learning framework, we achieve improvements over monolingual baselines and previously proposed mismatched transcription adaptation techniques. In addition, we show that using alignments provided by a GMM adapted by mismatched transcription further improves acoustic modeling performance. Our experiments on Georgian data from the IARPA Babel program show the effectiveness of the proposed method.
Generalized distillation framework has been shown to be effective in speech enhancement in the past. We extend this idea to speaker normalization without any explicit adaptation data in this paper. In the generalized distillation framework, we assume the presence of some “privileged” information to guide the training process in addition to the training data. In the proposed approach, the privileged information is obtained from a “teacher” model, trained on speaker-normalized FMLLR features. The “student” model is trained on un-normalized filterbank features and uses teacher’s supervision for cross-entropy training. The proposed distillation method does not need first pass decode information during testing and imposes no constraints on the duration of the test data for computing speaker-specific transforms unlike in FMLLR or i-vector. Experiments done on Switchboard and AMI corpus show that the generalized distillation framework shows improvement over un-normalized features with or without i-vectors.
Factorized Hidden Layer (FHL) adaptation has been proposed for speaker adaptation of deep neural network (DNN) based acoustic models. In FHL adaptation, a speaker-dependent (SD) transformation matrix and an SD bias are included in addition to the standard affine transformation. The SD transformation is a linear combination of rank-1 matrices whereas the SD bias is a linear combination of vectors. Recently, the Long Short-Term Memory (LSTM) Recurrent Neural Networks (RNNs) have shown to outperform DNN acoustic models in many Automatic Speech Recognition (ASR) tasks. In this work, we investigate the effectiveness of SD transformations for LSTM-RNN acoustic models. Experimental results show that when combined with scaling of LSTM cell states’ outputs, SD transformations achieve 2.3% and 2.1% absolute improvements over the baseline LSTM systems for the AMI IHM and AMI SDM tasks respectively.
Adapting acoustic models jointly to both speaker and environment has been shown to be effective. In many realistic scenarios, however, either the speaker or environment at test time might be unknown, or there may be insufficient data to learn a joint transform. Generating independent speaker and environment transforms improves the match of an acoustic model to unseen combinations. Using i-vectors, we demonstrate that it is possible to factorise speaker or environment information using multi-condition training with neural networks. Specifically, we extract bottleneck features from networks trained to classify either speakers or environments. We perform experiments on the Wall Street Journal corpus combined with environment noise from the Diverse Environments Multichannel Acoustic Noise Database. Using the factorised i-vectors we show improvements in word error rates on perturbed versions of the eval92 and dev93 test sets, both when one factor is missing and when the factors are seen but not in the desired combination.
In this work, we conduct a detailed evaluation of various all-neural, end-to-end trained, sequence-to-sequence models applied to the task of speech recognition. Notably, each of these systems directly predicts graphemes in the written domain, without using an external pronunciation lexicon, or a separate language model. We examine several sequence-to-sequence models including connectionist temporal classification (CTC), the recurrent neural network (RNN) transducer, an attention-based model, and a model which augments the RNN transducer with an attention mechanism. We find that the sequence-to-sequence models are competitive with traditional state-of-the-art approaches on dictation test sets, although the baseline, which uses a separate pronunciation and language model, outperforms these models on voice-search test sets.
We formulate a generalized hybrid HMM-NN training procedure using the full-sum over the hidden state-sequence and identify CTC as a special case of it. We present an analysis of the alignment behavior of such a training procedure and explain the strong localization of label output behavior of full-sum training (also referred to as peaky or spiky behavior). We show how to avoid that behavior by using a state prior. We discuss the temporal decoupling between output label position/time-frame, and the corresponding evidence in the input observations when this is trained with BLSTM models. We also show a way how to overcome this by jointly training a FFNN. We implemented the Baum-Welch alignment algorithm in CUDA to be able to do fast soft realignments on GPU. We have published this code along with some of our experiments as part of RETURNN, RWTH’s extensible training framework for universal recurrent neural networks. We finish with experimental validation of our study on WSJ and Switchboard.
We present a state-of-the-art end-to-end Automatic Speech Recognition (ASR) model. We learn to listen and write characters with a joint Connectionist Temporal Classification (CTC) and attention-based encoder-decoder network. The encoder is a deep Convolutional Neural Network (CNN) based on the VGG network. The CTC network sits on top of the encoder and is jointly trained with the attention-based decoder. During the beam search process, we combine the CTC predictions, the attention-based decoder predictions and a separately trained LSTM language model. We achieve a 5–10% error reduction compared to prior systems on spontaneous Japanese and Chinese speech, and our end-to-end model beats out traditional hybrid ASR systems.