Improvement In Word Error Rate Research Articles

Accurate synthesis of dysarthric Speech for ASR data augmentation

Dysarthria is a motor speech disorder often characterized by reduced speech intelligibility through slow, uncoordinated control of speech production muscles. Automatic Speech Recognition (ASR) systems can help dysarthric talkers communicate more effectively. However, robust dysarthria-specific ASR requires a significant amount of training speech, which is not readily available for dysarthric talkers.This paper presents a new dysarthric speech synthesis method for the purpose of ASR training data augmentation. Differences in prosodic and acoustic characteristics of dysarthric spontaneous speech at varying severity levels are important components for dysarthric speech modeling, synthesis, and augmentation. For dysarthric speech synthesis, a modified neural multi-talker TTS is implemented by adding a dysarthria severity level coefficient and a pause insertion model to synthesize dysarthric speech for varying severity levels.To evaluate the effectiveness for synthesis of training data for ASR, dysarthria-specific speech recognition was used. Results show that a DNNHMM model trained on additional synthetic dysarthric speech achieves relative Word Error Rate (WER) improvement of 12.2 % compared to the baseline, and that the addition of the severity level and pause insertion controls decrease WER by 6.5 %, showing the effectiveness of adding these parameters. Overall results on the TORGO database demonstrate that using dysarthric synthetic speech to increase the amount of dysarthric-patterned speech for training has significant impact on the dysarthric ASR systems. In addition, we have conducted a subjective evaluation to evaluate the dysarthricness and similarity of synthesized speech. Our subjective evaluation shows that the perceived dysarthricness of synthesized speech is similar to that of true dysarthric speech, especially for higher levels of dysarthria. Audio samples are available at https://mohammadelc.github.io/SpeechGroupUKY/

Server-side rescoring of spoken entity-centric knowledge queries for virtual assistants

On-device virtual assistants (VAs) powered by automatic speech recognition (ASR) require effective knowledge integration for the challenging entity-rich query recognition. In this paper, we conduct an empirical study of modeling strategies for server-side rescoring of spoken information domain queries using various categories of language models (LMs) (N-gram word LMs, sub-word neural LMs). We investigate the combination of on-device and server-side signals, and demonstrate significant word error rate improvements of 23%-35% relative on various entity-centric query subpopulations by integrating various server-side LMs compared to performing ASR on-device only. We also perform a comparison between LMs trained on domain data and a generative pre-trained (GPT) (a variant GPT-3) offered by OpenAI as a baseline. Furthermore, we also show that model fusion of multiple server-side LMs trained from scratch most effectively combines complementary strengths of each model and integrates knowledge learned from domain-specific data to a VA ASR system.

Improving speech recognition systems for the morphologically complex Malayalam language using subword tokens for language modeling

This article presents the research work on improving speech recognition systems for the morphologically complex Malayalam language using subword tokens for language modeling. The speech recognition system is built using a deep neural network–hidden Markov model (DNN-HMM)-based automatic speech recognition (ASR). We propose a novel method, syllable-byte pair encoding (S-BPE), that combines linguistically informed syllable tokenization with the data-driven tokenization method of byte pair encoding (BPE). The proposed method ensures words are always segmented at valid pronunciation boundaries. On a text corpus that has been divided into tokens using the proposed method, we construct statistical n-gram language models and assess the modeling effectiveness in terms of both information-theoretic and corpus linguistic metrics. A comparative study of the proposed method with other data-driven (BPE, Morfessor, and Unigram), linguistic (Syllable), and baseline (Word) tokenization algorithms is also presented. Pronunciation lexicons of subword tokenized units are built with pronunciation described as graphemes. We develop ASR systems employing the subword tokenized language models and pronunciation lexicons. The resulting ASR models are comprehensively evaluated to answer the research questions regarding the impact of subword tokenization algorithms on language modeling complexity and on ASR performance. Our study highlights the strong performance of the hybrid S-BPE tokens, achieving a notable 10.6% word error rate (WER), which represents a substantial 16.8% improvement over the baseline word-level ASR system. The ablation study has revealed that the performance of S-BPE segmentation, which initially underperformed compared to syllable tokens with lower amounts of textual data for language modeling, exhibited steady improvement with the increase in LM training data. The extensive ablation study indicates that there is a limited advantage in raising the n-gram order of the language model beyond n=3. Such an increase results in considerable model size growth without significant improvements in WER. The implementation of the algorithm and all associated experiments are available under an open license, allowing for reproduction, adaptation, and reuse.

The XMUSPEECH System for Accented English Automatic Speech Recognition

In this paper, we present the XMUSPEECH systems for Track 2 of the Interspeech 2020 Accented English Speech Recognition Challenge (AESRC2020). Track 2 is an Automatic Speech Recognition (ASR) task where the non-native English speakers have various accents, which reduces the accuracy of the ASR system. To solve this problem, we experimented with acoustic models and input features. Furthermore, we trained a TDNN-LSTM language model for lattice rescoring to obtain better results. Compared with our baseline system, we achieved relative word error rate (WER) improvements of 40.7% and 35.7% on the development set and evaluation set, respectively.

Prosodic Feature-Based Discriminatively Trained Low Resource Speech Recognition System

Speech recognition has been an active field of research in the last few decades since it facilitates better human–computer interaction. Native language automatic speech recognition (ASR) systems are still underdeveloped. Punjabi ASR systems are in their infancy stage because most research has been conducted only on adult speech systems; however, less work has been performed on Punjabi children’s ASR systems. This research aimed to build a prosodic feature-based automatic children speech recognition system using discriminative modeling techniques. The corpus of Punjabi children’s speech has various runtime challenges, such as acoustic variations with varying speakers’ ages. Efforts were made to implement out-domain data augmentation to overcome such issues using Tacotron-based text to a speech synthesizer. The prosodic features were extracted from Punjabi children’s speech corpus, then particular prosodic features were coupled with Mel Frequency Cepstral Coefficient (MFCC) features before being submitted to an ASR framework. The system modeling process investigated various approaches, which included Maximum Mutual Information (MMI), Boosted Maximum Mutual Information (bMMI), and feature-based Maximum Mutual Information (fMMI). The out-domain data augmentation was performed to enhance the corpus. After that, prosodic features were also extracted from the extended corpus, and experiments were conducted on both individual and integrated prosodic-based acoustic features. It was observed that the fMMI technique exhibited 20% to 25% relative improvement in word error rate compared with MMI and bMMI techniques. Further, it was enhanced using an augmented dataset and hybrid front-end features (MFCC + POV + Fo + Voice quality) with a relative improvement of 13% compared with the earlier baseline system.

Real-time pre-processing for improved feature extraction of noisy speech

Several improvements of algorithms for the front-end feature extraction of real-time speech decoding in noisy ambiance have been proposed with their demonstration on the TIMIT speech corpus. Real-Time Voice Activity Detection (RT-VAD) is used to separate the voiced–unvoiced part of input from silence in the streaming speech input. Novel techniques for RT-Zero Crossing Detection and RT-Pitch Detection are presented as part of RT-VAD. Real-Time approximate Kalman filter is then applied to de-noise the incoming signal. All these are applied across a collection of frames of speech called context. Frame-based Linear Discriminant Analysis (LDA)-feature extraction is done by RT-Cepstral Mean and Variance Normalization (RT-CMVN) and RT-Splicing. The algorithms are tested on the TIMIT database for various noise levels. It is observed that we obtain a word-error rate (WER) improvement of 5% for 30 dB and 7% for 10 dB SNR, thus validating the proposed algorithms. Also, the comparison with other works shows a superior Speech Hit Rate (SHR) of 90.6% and Noise Hit Rate (NHR) of 86.2%.

A comparative case study of neural network training by using frame-level cost functions for automatic speech recognition purposes in Spanish

Training procedures of a deep neural network are still an area with ample research possibilities and constant improvement either to increase its efficiency or its time performance. One of the lesser-addressed components is its objective function, which is an underlying aspect to consider when there is the necessity to achieve better error rates in the area of automatic speech recognition. The aim of this paper is to present two new variations of the frame-level cost function for training a deep neural network with the purpose of obtaining superior word error rates in speech recognition applied to a case study in Spanish. The first proposed function is a fusion between the boosted cross-entropy and the so called cross-entropy/log-posterior-ratio. The main idea is to jointly emphasize the prediction of difficult/crucial frames provided by a boosting factor and at the same time enlarge the distance between the target senone and its closest competitor. The second proposal is a fusion between the non-uniform mapped cross-entropy and the cross-entropy/log-posterior-ratio. This function utilizes both the mapped function to enhance the frames that have ambiguity in their belonging to specific senones and the log-posterior-ratio with the purpose of separating the target senone against the most competing tied tri-phone state. The proposed approaches are compared against those frame-level cost functions discussed in the state of the art. This comparative has been made by using a personalized mid-vocabulary speaker-independent voice corpus. This dataset is employed for the recognition of digit strings and personal name lists in Spanish from the northern central part of Mexico on a connected-words phone dialing task. A relative word error rate improvement of 15.14% and 12.30% is obtained with the two proposed approaches, respectively, against the plain well-established cross-entropy loss function.

Transfer learning for domain and environment adaptation in Serbian ASR

In automatic speech recognition systems, the training data used for system development and the data actually obtained from the users of the system sometimes significantly differ in practice. However, other, more similar data may be available. Transfer learning can help to exploit such similar data for training in order to boost the automatic speech recognizer's performance for a certain domain. This paper presents a few applications of transfer learning in the context of speech recognition, specifically for the Serbian language. Several methods are proposed, with the goal of optimizing system performance on a specific part of the existing speech database for Serbian, or in a noisy environment. The experimental results evaluated on a test set from the desired domain show significant improvement in both word error rate and character error rate.

Training of reduced-rank linear transformations for multi-layer polynomial acoustic features for speech recognition

The use of higher-order polynomial acoustic features can improve the performance of automatic speech recognition (ASR). However, dimensionality of polynomial representation can be prohibitively large, making acoustic model training using polynomial features infeasible for large vocabulary ASR systems. This paper presents a multi-layer polynomial training framework for acoustic modeling, which recursively expands the acoustic features into their second-order polynomial feature space. After each expansion the dimensionality of resultant features is reduced by a linear transformation. Experimental results obtained for two large-vocabulary continuous speech recognition tasks show that the proposed method outperforms conventional mixture models. More recently the acoustic modelling community has shifted its focus to deep neural networks. We also train multi-layer polynomial features in a similar way: allowing backpropagation and using mean-normalized stochastic gradient descent algorithm. This has led to encouraging results. Specifically, appending a sigmoid-based feed-forward deep neural network with a final polynomial layer has resulted in significant word error rate improvement.

Recurrent Neural Network Language Model Adaptation for Multi-Genre Broadcast Speech Recognition and Alignment

Recurrent neural network language models (RNNLMs) generally outperform $n$ -gram language models when used in automatic speech recognition (ASR). Adapting RNNLMs to new domains is an open problem and current approaches can be categorised as either feature-based or model based. In feature-based adaptation, the input to the RNNLM is augmented with auxiliary features whilst model-based adaptation includes model fine-tuning and the introduction of adaptation layer(s) in the network. In this paper, the properties of both types of adaptation are investigated on multi-genre broadcast speech recognition. Existing techniques for both types of adaptation are reviewed and the proposed techniques for model-based adaptation, namely the linear hidden network adaptation layer and the $K$ -component adaptive the RNNLM, are investigated. Moreover, new features derived from the acoustic domain are investigated for the RNNLM adaptation. The contributions of this paper include two hybrid adaptation techniques: the fine-tuning of feature-based RNNLMs and a feature-based adaptation layer. Moreover, the semi-supervised adaptation of RNNLMs using genre information is also proposed. The ASR systems were trained using 700 h of multi-genre broadcast speech. The gains obtained when using the RNNLM adaptation techniques proposed in this paper are consistent when using RNNLMs trained on an in-domain set of 10M words and on a combination of in-domain and out-of-domain sets of 660 M words, with approx. $\text{10}{\%}$ perplexity and $\text{2}{\%}$ relative word error rate improvements on a 28.3 h. test set. The best RNNLM adaptation techniques for ASR are also evaluated on a lightly supervised alignment of subtitles task for the same data, where the use of RNNLM adaptation leads to an absolute increase in the F–measure of $\text{0.5}{\%}$ .

Emerging trends: APIs for speech and machine translation and more

AbstractLots of companies are offering lots of APIs. Reviews are not always as constructive as they could be. Some reviews encourage unproductive work on checkbox features that no one wants. It makes no sense to do the wrong thing badly. Constructive reviews should help focus priorities on what matters. Users care more about a great box opening experience than small improvements in word error rate and BLEU, popular metrics for speech and translation. In 15 minutes or less, can we teach potential users something new (and fun) that does something useful, such as how to translate PowerPoint between English and Chinese, preserving many of the features that are important to PowerPoint such as graphics and animations? See Appendix for the solution.

Retrospective Analysis of Clinical Performance of an Estonian Speech Recognition System for Radiology: Effects of Different Acoustic and Language Models

The aim of this study was to analyze retrospectively the influence of different acoustic and language models in order to determine the most important effects to the clinical performance of an Estonian language-based non-commercial radiology-oriented automatic speech recognition (ASR) system. An ASR system was developed for Estonian language in radiology domain by utilizing open-source software components (Kaldi toolkit, Thrax). The ASR system was trained with the real radiology text reports and dictations collected during development phases. The final version of the ASR system was tested by 11 radiologists who dictated 219 reports in total, in spontaneous manner in a real clinical environment. The audio files collected in the final phase were used to measure the performance of different versions of the ASR system retrospectively. ASR system versions were evaluated by word error rate (WER) for each speaker and modality and by WER difference for the first and the last version of the ASR system. Total average WER for the final version throughout all material was improved from 18.4% of the first version (v1) to 5.8% of the last (v8) version which corresponds to relative improvement of 68.5%. WER improvement was strongly related to modality and radiologist. In summary, the performance of the final ASR system version was close to optimal, delivering similar results to all modalities and being independent on user, the complexity of the radiology reports, user experience, and speech characteristics.

Training deep neural networks with non-uniform frame-level cost function for automatic speech recognition

The aim of this paper is to exhibit two new variations of the frame-level cost function for training a deep neural network in order to achieve better word error rates in speech recognition. Optimization methods and their minimization functions are underlying aspects to consider when someone is working on neural nets, and hence their improvement is one of the salient objectives of researchers, and this paper deals in part with such a situation. The first proposed framework is based on the concept of extropy, the complementary dual function of an uncertainty measure. The conventional cross-entropy function can be mapped to a non-uniform loss function based on its corresponding extropy, enhancing the frames that have ambiguity in their belonging to specific senones. The second proposal makes a fusion of the presented mapped cross-entropy function and the idea of boosted cross-entropy, which emphasizes those frames with low target posterior probability. The proposed approaches have been performed by using a personalized mid-vocabulary speaker-independent voice corpus. This dataset is employed for recognition of digit strings and personal name lists in Spanish from the northern central part of Mexico on a connected-words phone dialing task. A relative word error rate improvement of $12.3\%$ and $10.7\%$ is obtained with the two proposed approaches, respectively, with regard to the conventional well-established cross-entropy objective function.

Multichannel Signal Processing With Deep Neural Networks for Automatic Speech Recognition

Multichannel automatic speech recognition (ASR) systems commonly separate speech enhancement, including localization, beamforming, and postfiltering, from acoustic modeling. In this paper, we perform multichannel enhancement jointly with acoustic modeling in a deep neural network framework. Inspired by beamforming, which leverages differences in the fine time structure of the signal at different microphones to filter energy arriving from different directions, we explore modeling the raw time-domain waveform directly. We introduce a neural network architecture, which performs multichannel filtering in the first layer of the network, and show that this network learns to be robust to varying target speaker direction of arrival, performing as well as a model that is given oracle knowledge of the true target speaker direction. Next, we show how performance can be improved by factoring the first layer to separate the multichannel spatial filtering operation from a single channel filterbank which computes a frequency decomposition. We also introduce an adaptive variant, which updates the spatial filter coefficients at each time frame based on the previous inputs. Finally, we demonstrate that these approaches can be implemented more efficiently in the frequency domain. Overall, we find that such multichannel neural networks give a relative word error rate improvement of more than 5% compared to a traditional beamforming-based multichannel ASR system and more than 10% compared to a single channel waveform model.

Vector Taylor series based model adaptation using noisy speech trained hidden Markov models

Conventionally, in vector Taylor series (VTS) based compensation for noise-robust speech recognition, hidden Markov models (HMMs) are usually trained with clean speech. However, it is known that better performance is generally obtained by training the HMM with noisy speech rather than clean speech. From this viewpoint, we propose a novel VTS-based HMM adaptation method for the noisy speech trained HMM. We derive a mathematical relation between the training and test noisy speech in the cepstrum-domain using VTS and the mean and covariance of the noisy speech trained HMM are adapted to the test noisy speech in an iterative expectation-maximization (EM) algorithm. In the experimental results on the Aurora 2 database, we could obtain about 10–25% relative improvements in word error rates (WERs) over multi-condition training (MTR) method depending on speech front-ends and the HMM complexity.

A SYSTEM COMBINATION FOR MALAY BROADCAST NEWS TRANSCRIPTION

In this paper, we propose a post decoding system combination approach for automatic transcribing Malay broadcast news. This approach combines the hypotheses produced by parallel automatic speech recognition (ASR) systems. Each ASR system uses different language models, one which is generic domain model and another is domain specific model. The main idea is to take advantage of different ASR knowledge to improve ASR decoding result. It uses the language score and time information to produce a 1-best lattice, and then rescore the 1-best lattice to get the most likely word sequence as the final output. The proposed approach was compared with conventional combination approach, the recognizer output voting error reduction (ROVER). Our proposed approach improved the word error rate (WER) from 33.9% to 30.6% with an average relative WER improvement of 9.74%, and it is better than the conventional ROVER approach.

Adaptation of Morph-Based Speech Recognition for Foreign Names and Acronyms

In this paper, we improve morph-based speech recognition system by focusing adaptation efforts on acronyms (ACRs) and foreign proper names (FPNs). An unsupervised language model (LM) adaptation framework based on two-pass decoding is used. Vocabulary adaptation is applied alongside unsupervised LM adaptation. The aim is to improve both language and pronunciation modeling for FPNs and ACRs. A smart selection algorithm is used to find the most likely topically related foreign words and acronyms from in-domain text. New pronunciation rules are generated for the selected words. Different kinds of morpheme adaptation operations are also evaluated on the ACR and FPN candidate words, to ensure optimal results are gained from pronunciation adaptation. Statistically significant improvements in average word error rate (WER), and term error rate (TER), are achieved using a combination of unsupervised LM adaptation with vocabulary adaptation focused on ACRs and FPNs.

Combination of SPLICE and Feature Normalization for Noise Robust Speech Recognition

It is well-known that the performance of automatic speech recognition (ASR) systems are easily affected by acoustic mismatch between training and testing conditions. This mismatch is often caused by various kinds of environmental noise or distortion. To reduce the effect of mismatch, feature normalization, feature enhancement, model adaptation, etc. have been studied intensively. Cepstral mean normalization (CMN), mean and variance normalization (MVN) and histogram equalization (HEQ) are well-known methods of feature normalization. Stereo-based piecewise linear compensation for environments (SPLICE) is one of the feature enhancement methods. In this paper, we describe how to combine these methods to effectively improve the robustness of ASR systems. In the experiments performed on the Aurora-2 database, a good combination showed a 41% improvement in word error rate over SPLICE only, and a 25% improvement over the conventional combination of SPLICE and CMN.

A Novel Mask Estimation Method Employing Posterior-Based Representative Mean Estimate for Missing-Feature Speech Recognition

This paper proposes a novel mask estimation method for missing-feature reconstruction to improve speech recognition performance in various types of background noise conditions. A conventional mask estimation method based on spectral subtraction degrades performance, due to incorrect estimation of the noise signal which fails to accurately represent the variations of background noise during the incoming speech utterance. The proposed mask estimation method utilizes a Posterior-based Representative Mean (PRM) estimate for determining the reliability of the input speech spectral components, which is obtained as a weighted sum of the mean parameters of the speech model using the posterior probability. To obtain the noise-corrupted speech model, a model combination method is employed, which was proposed in our previous study for a feature compensation method. Experimental results demonstrate that the proposed mask estimation method provides more separable distributions for the reliable/unreliable component classifier compared to the conventional mask estimation method. The recognition performance is evaluated using the Aurora 2.0 framework over various types of background noise conditions and the CU-Move real-life in-vehicle corpus. The performance evaluation shows that the proposed mask estimation method is considerably more effective at increasing speech recognition performance in various types of background noise conditions, compared to the conventional mask estimation method which is based on spectral subtraction. By employing the proposed PRM-based mask estimation for missing-feature reconstruction, we obtain +23.41% and +9.45% average relative improvements in word error rate for all four types of noise conditions and CU-Move corpus, respectively, compared to conventional mask estimation methods.

Variational noise model composition through model perturbation for robust speech recognition with time-varying background noise

This study proposes a novel model composition method to improve speech recognition performance in time-varying background noise conditions. It is suggested that each element of the cepstral coefficients represents the frequency degree of the changing components in the envelope of the log-spectrum. With this motivation, in the proposed method, variational noise models are formulated by selectively applying perturbation factors to the mean parameters of a basis model, resulting in a collection of noise models that more accurately reflect the natural range of spectral patterns seen in the log-spectral domain. The basis noise model is obtained from the silence segments of the input speech. The perturbation factors are designed separately for changes in the energy level and spectral envelope. The proposed variational model composition (VMC) method is employed to generate multiple environmental models for our previously proposed parallel combined gaussian mixture model (PCGMM) based feature compensation algorithm. The mixture sharing technique is integrated to reduce computational expenses, caused by employing the variational models. Experimental results prove that the proposed method is considerably more effective at increasing speech recognition performance in time-varying background noise conditions, with +31.31%, +10.65%, and +20.54% average relative improvements in word error rate for speech babble, background music, and real-life in-vehicle noise conditions respectively, compared to the original basic PCGMM method.