Total: 203
In this work, we propose a new end-to-end (E2E) spelling correction method for post-processing of code-switching automatic speech recognition (ASR). Existing E2E spelling correction models take the hypotheses of ASR as inputs and annotated text as the targets. Due to the powerful modeling capabilities of the E2E model, the training of the correction system is extremely prone to over-fitting. It usually requires sufficient data diversity for reliable training. Therefore, it is difficult to apply the E2E correction models to the code-switching ASR task because of the data shortage. In this paper, we introduce the acoustic features into the spelling correction model. Our method can alleviate the problem of over-fitting and has better performance. Meanwhile, because the acoustic features are encode-free, our proposed model can be applied to the ASR model without significantly increasing the computational cost. The experimental results on ASRU 2019 Mandarin-English Code-switching Challenge data set show that the proposed method achieves 11.14% relative error rate reduction compared with baseline.
Models pre-trained on multiple languages have shown significant promise for improving speech recognition, particularly for low-resource languages. In this work, we focus on phoneme recognition using Allosaurus, a method for multilingual recognition based on phonetic annotation, which incorporates phonological knowledge through a language-dependent allophone layer that associates a universal narrow phone-set with the phonemes that appear in each language. To evaluate in a challenging real-world scenario, we curate phone recognition datasets for Bukusu and Saamia, two varieties of the Luhya language cluster of western Kenya and eastern Uganda. To our knowledge, these datasets are the first of their kind. We carry out similar experiments on the dataset of an endangered Tangkhulic language, East Tusom, a Tibeto-Burman language variety spoken mostly in India. We explore both zero-shot and few-shot recognition by fine-tuning using datasets of varying sizes (10 to 1000 utterances). We find that fine-tuning of Allosaurus, even with just 100 utterances, leads to significant improvements in phone error rates.
Source-filter modelling is among the fundamental techniques in speech processing with a wide range of applications. In acoustic modelling, features such as MFCC and PLP which parametrise the filter component are widely employed. In this paper, we investigate the efficacy of building acoustic models from the raw filter and source components. The raw magnitude spectrum, as the primary information stream, is decomposed into the excitation and vocal tract information streams via cepstral liftering. Then, acoustic models are built via multi-head CNNs which, among others, allow for processing each individual stream via a sequence of bespoke transforms and fusing them at an optimal level of abstraction. We discuss the possible advantages of such information factorisation and recombination, investigate the dynamics of these models and explore the optimal fusion level. Furthermore, we illustrate the CNN’s learned filters and provide some interpretation for the captured patterns. The proposed approach with optimal fusion scheme results in up to 14% and 7% relative WER reduction in WSJ and Aurora-4 tasks.
This paper addresses a noise-robust automatic speech recognition (ASR) method under the constraints of real-time, one-pass, and single-channel processing. Under such strong constraints, single-channel speech enhancement becomes a key technology because methods with multiple-passes or batch processing, such as acoustic model adaptation, are not suitable for use. However, single-channel speech enhancement often degrades ASR performance due to speech distortion. To overcome this problem, we propose a noise robust acoustic modeling method based on the stream-wise transformer model. The proposed method accepts multi-stream features obtained by multiple single-channel speech enhancement methods as input and selectively uses an appropriate feature stream according to the noise environment by paying attention to the noteworthy stream on the basis of multi-head attention. The proposed method considers the attention for the stream direction instead of the time series direction, and it is thus capable of real-time and low-latency processing. Comparative evaluations reveal that the proposed method successfully improves the accuracy of ASR in noisy environments and reduces the number of model parameters even under strong constraints.
We present a Generative Adversarial Network (GAN) based room impulse response generator (IR-GAN) for generating realistic synthetic room impulse responses (RIRs). IR-GAN extracts acoustic parameters from captured real-world RIRs and uses these parameters to generate new synthetic RIRs. We use these generated synthetic RIRs to improve far-field automatic speech recognition in new environments that are different from the ones used in training datasets. In particular, we augment the far-field speech training set by convolving our synthesized RIRs with a clean LibriSpeech dataset [1]. We evaluate the quality of our synthetic RIRs on the far-field LibriSpeech test set created using real-world RIRs from the BUT ReverbDB [2] and AIR [3] datasets. Our IR-GAN reports up to an 8.95% lower error rate than Geometric Acoustic Simulator (GAS) in far-field speech recognition benchmarks. We further improve the performance when we combine our synthetic RIRs with synthetic impulse responses generated using GAS. This combination can reduce the word error rate by up to 14.3% in far-field speech recognition benchmarks.
Recently, speech recognition with ad-hoc microphone arrays has received much attention. It is known that channel selection is an important problem of ad-hoc microphone arrays, however, this topic seems far from explored in speech recognition yet, particularly with a large-scale ad-hoc microphone array. To address this problem, we propose a Scaling Sparsemax algorithm for the channel selection problem of the speech recognition with large-scale ad-hoc microphone arrays. Specifically, we first replace the conventional Softmax operator in the stream attention mechanism of a multichannel end-to-end speech recognition system with Sparsemax, which conducts channel selection by forcing the channel weights of noisy channels to zero. Because Sparsemax punishes the weights of many channels to zero harshly, we propose Scaling Sparsemax which punishes the channels mildly by setting the weights of very noisy channels to zero only. Experimental results with ad-hoc microphone arrays of over 30 channels under the conformer speech recognition architecture show that the proposed Scaling Sparsemax yields a word error rate of over 30% lower than Softmax on simulation data sets, and over 20% lower on semi-real data sets, in test scenarios with both matched and mismatched channel numbers.
Multi-channel inputs offer several advantages over single-channel, to improve the robustness of on-device speech recognition systems. Recent work on multi-channel transformer, has proposed a way to incorporate such inputs into end-to-end ASR for improved accuracy. However, this approach is characterized by a high computational complexity, which prevents it from being deployed in on-device systems. In this paper, we present a novel speech recognition model, Multi-Channel Transformer Transducer (MCTT), which features end-to-end multi-channel training, low computation cost, and low latency so that it is suitable for streaming decoding in on-device speech recognition. In a far-field in-house dataset, our MCTT outperforms stagewise multi-channel models with transformer-transducer up to 6.01% relative WER improvement (WERR). In addition, MCTT outperforms the multi-channel transformer up to 11.62% WERR, and is 15.8 times faster in terms of inference speed. We further show that we can improve the computational cost of MCTT by constraining the future and previous context in attention computations.
Although end-to-end automatic speech recognition (E2E ASR) has achieved great performance in tasks that have numerous paired data, it is still challenging to make E2E ASR robust against noisy and low-resource conditions. In this study, we investigated data augmentation methods for E2E ASR in distant-talk scenarios. E2E ASR models are trained on the series of CHiME challenge datasets, which are suitable tasks for studying robustness against noisy and spontaneous speech. We propose to use three augmentation methods and their combinations: 1) data augmentation using text-to-speech (TTS) data, 2) cycle-consistent generative adversarial network (Cycle-GAN) augmentation trained to map two different audio characteristics, the one of clean speech and of noisy recordings, to match the testing condition, and 3) pseudo-label augmentation provided by the pretrained ASR module for smoothing label distributions. Experimental results using the CHiME-6/CHiME-4 datasets show that each augmentation method individually improves the accuracy on top of the conventional SpecAugment; further improvements are obtained by combining these approaches. We achieved 4.3% word error rate (WER) reduction, which was more significant than that of the SpecAugment, when we combine all three augmentations for the CHiME-6 task.
In Uyghur speech, consonant and vowel reduction are often encountered, especially in spontaneous speech with high speech rate, which will cause a degradation of speech recognition performance. To solve this problem, we propose an effective phone mask training method for Conformer-based Uyghur end-to-end (E2E) speech recognition. The idea is to randomly mask off a certain percentage features of phones during model training, which simulates the above verbal phenomena and facilitates E2E model to learn more contextual information. According to experiments, the above issues can be greatly alleviated. In addition, deep investigations are carried out into different units in masking, which shows the effectiveness of our proposed masking unit. We also further study the masking method and optimize filling strategy of phone mask. Finally, compared with Conformer-based E2E baseline without mask training, our model demonstrates about 5.51% relative Word Error Rate (WER) reduction on reading speech and 12.92% on spontaneous speech, respectively. The above approach has also been verified on test-set of open-source data THUYG-20, which shows 20% relative improvements.
Is pushing numbers on a single benchmark valuable in automatic speech recognition? Research results in acoustic modeling are typically evaluated based on performance on a single dataset. While the research community has coalesced around various benchmarks, we set out to understand generalization performance in acoustic modeling across datasets — in particular, if models trained on a single dataset transfer to other (possibly out-of-domain) datasets. Further, we demonstrate that when a large enough set of benchmarks is used, average word error rate (WER) performance over them provides a good proxy for performance on real-world data. Finally, we show that training a single acoustic model on the most widely-used datasets — combined — reaches competitive performance on both research and real-world benchmarks.
End-to-end speech recognition generally uses hand-engineered acoustic features as input and excludes the feature extraction module from its joint optimization. To extract learnable and adaptive features and mitigate information loss, we propose a new encoder that adopts globally attentive locally recurrent (GALR) networks and directly takes raw waveform as input. We observe improved ASR performance and robustness by applying GALR on different window lengths to aggregate fine-grain temporal information into multi-scale acoustic features. Experiments are conducted on a benchmark dataset AISHELL-2 and two large-scale Mandarin speech corpus of 5,000 hours and 21,000 hours. With faster speed and comparable model size, our proposed multi-scale GALR waveform encoder achieved consistent character error rate reductions (CERRs) from 7.9% to 28.1% relative over strong baselines, including Conformer and TDNN-Conformer. In particular, our approach demonstrated notable robustness than the traditional handcrafted features and outperformed the baseline MFCC-based TDNN-Conformer model by a 15.2% CERR on a music-mixed real-world speech test set.
Recently self-supervised learning has emerged as an effective approach to improve the performance of automatic speech recognition (ASR). Under such a framework, the neural network is usually pre-trained with massive unlabeled data and then fine-tuned with limited labeled data. However, the non-streaming architecture like bidirectional transformer is usually adopted by the neural network to achieve competitive results, which cannot be used in streaming scenarios. In this paper, we mainly focus on improving the performance of streaming transformer under the self-supervised learning framework. Specifically, we propose a novel two-stage training method during fine-tuning, which combines knowledge distilling and self-training. The proposed training method achieves 16.3% relative word error rate (WER) reduction on Librispeech noisy test set. Finally, by only using the 100h clean subset of Librispeech as the labeled data and the rest (860h) as the unlabeled data, our streaming transformer based model obtains competitive WERs 3.5/8.7 on Librispeech clean/noisy test sets.
wav2vec-C introduces a novel representation learning technique combining elements from wav2vec 2.0 and VQ-VAE. Our model learns to reproduce quantized representations from partially masked speech encoding using a contrastive loss in a way similar to wav2vec 2.0. However, the quantization process is regularized by an additional consistency network that learns to reconstruct the input features to the wav2vec 2.0 network from the quantized representations in a way similar to a VQ-VAE model. The proposed self-supervised model is trained on 10k hours of unlabeled data and subsequently used as the speech encoder in a RNN-T ASR model and fine-tuned with 1k hours of labeled data. This work is one of the very few studies of self-supervised learning on speech tasks with a large volume of real far-field labeled data. The wav2vec-C encoded representations achieve, on average, twice the error reduction over baseline and a higher codebook utilization in comparison to wav2vec 2.0.
The use of semi-supervised training (SST) has become an increasingly popular way of increasing the performance of ASR acoustic models without the need for further transcribed speech data. However, the performance of the technique can be very sensitive to the quality of the initial ASR system. This paper undertakes a comprehensive study of the improvements gained with respect to variation in the initial systems, the quantity of untranscribed data used, and the learning schedules. We postulate that the reason SST can be effective even when the initial model is poor is because it enables utterance-level information to be propagated to the frame level, and hence hypothesise that the quality of the language model plays a much larger role than the quality of the acoustic model. In experiments on Tagalog data from the IARPA MATERIAL programme, we find that indeed this is the case, and show that with an appropriately chosen recipe it is possible to achieve over 50% relative WER reductions from SST, even when the WER of the initial system is more than 80%.
Self-supervised learning of speech representations has been a very active research area but most work is focused on a single domain such as read audio books for which there exist large quantities of labeled and unlabeled data. In this paper, we explore more general setups where the domain of the unlabeled data for pre-training data differs from the domain of the labeled data for fine-tuning, which in turn may differ from the test data domain. Our experiments show that using target domain data during pre-training leads to large performance improvements across a variety of setups. With no access to in-domain labeled data, pre-training on unlabeled in-domain data closes 66–73% of the performance gap between the ideal setting of in-domain labeled data and a competitive supervised out-of-domain model. This has obvious practical implications since it is much easier to obtain unlabeled target domain data than labeled data. Moreover, we find that pre-training on multiple domains improves generalization performance on domains not seen during training. We will release pre-trained models.
Pseudo-labeling (PL) has been shown to be effective in semi-supervised automatic speech recognition (ASR), where a base model is self-trained with pseudo-labels generated from unlabeled data. While PL can be further improved by iteratively updating pseudo-labels as the model evolves, most of the previous approaches involve inefficient retraining of the model or intricate control of the label update. We present momentum pseudo-labeling (MPL), a simple yet effective strategy for semi-supervised ASR. MPL consists of a pair of online and offline models that interact and learn from each other, inspired by the mean teacher method. The online model is trained to predict pseudo-labels generated on the fly by the offline model. The offline model maintains a momentum-based moving average of the online model. MPL is performed in a single training process and the interaction between the two models effectively helps them reinforce each other to improve the ASR performance. We apply MPL to an end-to-end ASR model based on the connectionist temporal classification. The experimental results demonstrate that MPL effectively improves over the base model and is scalable to different semi-supervised scenarios with varying amounts of data or domain mismatch.
In the absence of large-scale in-domain supervised training data, ASR models can achieve reasonable performance through pre-training on additional data that is unlabeled, mismatched or both. Given such data constraints, we compare pre-training end-to-end models on matched but unlabeled data (unsupervised) and on labeled but mismatched data (supervised), where the labeled data is mismatched in either domain or language. Across encoder architectures, pre-training methods and languages, our experiments indicate that both types of pre-training improve performance, with relative WER reductions of 15–30% in the domain mismatch case and up to 15% in the language mismatch condition. We further find that the advantage from unsupervised pre-training is most prominent when there is no matched and labeled fine-tuning data, provided that a sufficient amount of mismatched data is still available for supervised fine-tuning.
Semi and self-supervised training techniques have the potential to improve performance of speech recognition systems without additional transcribed speech data. In this work, we demonstrate the efficacy of two approaches to semi-supervision for automated speech recognition. The two approaches leverage vast amounts of available unspoken text and untranscribed audio. First, we present factorized multilingual speech synthesis to improve data augmentation on unspoken text. Next, we propose the Sequential MixMatch algorithm with iterative learning to learn from untranscribed speech. The algorithm is built on top of our online implementation of Noisy Student Training. We demonstrate the compatibility of these techniques yielding an overall relative reduction of word error rate of up to 14.4% on the voice search tasks on 4 Indic languages.
Recent results in end-to-end automatic speech recognition have demonstrated the efficacy of pseudo-labeling for semi-supervised models trained both with Connectionist Temporal Classification (CTC) and Sequence-to-Sequence (seq2seq) losses. Iterative Pseudo-Labeling (IPL), which continuously trains a single model using pseudo-labels iteratively re-generated as the model learns, has been shown to further improve performance in ASR. We improve upon the IPL algorithm: as the model learns, we propose to iteratively re-generate transcriptions with hard labels (the most probable tokens), that is, without a language model. We call this approach Language-Model-Free IPL (slimIPL) and give a resultant training setup for low-resource settings with CTC-based models. slimIPL features a dynamic cache for pseudo-labels which reduces sensitivity to changes in relabeling hyperparameters and results in improved training stability. slimIPL is also highly-efficient and requires 3.5–4× fewer computational resources to converge than other state-of-the-art semi/self-supervised approaches. With only 10 hours of labeled audio, slimIPL is competitive with self-supervised approaches, and is state-of-the-art with 100 hours of labeled audio without the use of a language model both at test time and during pseudo-label generation.
Self-supervised representation learning has seen remarkable success in encoding high-level semantic information from unlabelled speech data. The studies have been focused on exploring new pretext tasks to improve the learned speech representation and various masking schemes with reference to speech frames. We consider effective latent speech representation should be phonetically informed. In this work, we propose a novel phonetically motivated masking scheme. Specifically, we select the masked speech frames according to the phonetic segmentation in an utterance. The phonetically motivated self-supervised representation learns the speech representation that benefits downstream speech processing tasks. We evaluate the proposed learning algorithm on phoneme classification, speech recognition, and speaker recognition, and show that it consistently outperforms competitive baselines.
Recurrent neural network transducer (RNN-T) has shown to be comparable with conventional hybrid model for speech recognition. However, there is still a challenge in out-of-domain scenarios with context or words different from training data. In this paper, we explore the semi-supervised training which optimizes RNN-T jointly with neural text-to-speech (TTS) to better generalize to new domains using domain-specific text data. We apply the method to two tasks: one with out-of-domain context and the other with significant out-of-vocabulary (OOV) words. The results show that the proposed method significantly improves the recognition accuracy in both tasks, resulting in 61.4% and 53.8% relative word error rate (WER) reductions respectively, from a well-trained RNN-T with 65 thousand hours of training data. We do further study on the semi-supervised training methodology: 1) which modules of RNN-T model to be updated; 2) the impact of using different neural TTS models; 3) the performance of using text with different relevancy to target domain. Finally, we compare several RNN-T customization methods, and conclude that semi-supervised training with neural TTS is comparable and complementary with Internal Language Model Estimation (ILME) or biasing.
We present a number of low-resource approaches to the tasks of the Zero Resource Speech Challenge 2021. We build on the unsupervised representations of speech proposed by the organizers as a baseline, derived from CPC and clustered with the k-means algorithm. We demonstrate that simple methods of refining those representations can narrow the gap, or even improve upon the solutions which use a high computational budget. The results lead to the conclusion that the CPC-derived representations are still too noisy for training language models, but stable enough for simpler forms of pattern matching and retrieval.
We investigate the possibility of forcing a self-supervised model trained using a contrastive predictive loss, to extract slowly varying latent representations. Rather than producing individual predictions for each of the future representations, the model emits a sequence of predictions shorter than the sequence of upcoming representations to which they will be aligned. In this way, the prediction network solves a simpler task of predicting the next symbols, but not their exact timing, while the encoding network is trained to produce piece-wise constant latent codes. We evaluate the model on a speech coding task and demonstrate that the proposed Aligned Contrastive Predictive Coding (ACPC) leads to higher linear phone prediction accuracy and lower ABX error rates, while being slightly faster to train due to the reduced number of prediction heads.
This paper presents a simple sequence-to-sequence approach to restore standard orthography in raw, normalized speech transcripts, including insertion of punctuation marks, prediction of capitalization, restoration of numeric forms, formatting of dates and times, and other, fully data-driven adjustments. We further describe our method to generate synthetic parallel training data, and explore suitable performance metrics, which we align with human judgment through subjective MOS-like evaluations. Our models for English, Russian, and German have a word error rate of 6.36%, 4.88%, and 5.23%, respectively. We focus on simplicity and reproducibility, make our framework available under a BSD license, and share our base models for English and Russian.
The Fearless Steps Challenge (FSC) initiative was designed to host a series of progressively complex tasks to promote advanced speech research across naturalistic “Big Data” corpora. The Center for Robust Speech Systems at UT-Dallas in collaboration with the National Institute of Standards and Technology (NIST) and Linguistic Data Consortium (LDC) conducted Phase-3 of the FSC series (FSC P3), with a focus on motivating speech and language technology (SLT) system generalizability across channel and mission diversity under the same training conditions as in Phase-2. The FSC P3 introduced 10 hours of previously unseen channel audio from Apollo-11 and 5 hours of novel audio from Apollo-13 to be evaluated over both previously established and newly introduced SLT tasks with streamlined tracks. This paper presents an overview of the newly introduced conversational analysis tracks, Apollo-13 data, and analysis of system performance for matched and mismatched challenge conditions. We also discuss the Phase-3 challenge results, evolution of system performance across the three Phases, and next steps in the Challenge Series.