Language Models For Speech Recognition Research Articles

VioLA: Conditional Language Models for Speech Recognition, Synthesis, and Translation

VioLA: Conditional Language Models for Speech Recognition, Synthesis, and Translation

Tuning Large Language Model for Speech Recognition With Mixed-Scale Re-Tokenization

Tuning Large Language Model for Speech Recognition With Mixed-Scale Re-Tokenization

Bayesian Neural Network Language Modeling for Speech Recognition

Bayesian Neural Network Language Modeling for Speech Recognition

Dynamic out-of-vocabulary word registration to language model for speech recognition

We propose a method of dynamically registering out-of-vocabulary (OOV) words by assigning the pronunciations of these words to pre-inserted OOV tokens, editing the pronunciations of the tokens. To do this, we add OOV tokens to an additional, partial copy of our corpus, either randomly or to part-of-speech (POS) tags in the selected utterances, when training the language model (LM) for speech recognition. This results in an LM containing OOV tokens, to which we can assign pronunciations. We also investigate the impact of acoustic complexity and the “natural” occurrence frequency of OOV words on the recognition of registered OOV words. The proposed OOV word registration method is evaluated using two modern automatic speech recognition (ASR) systems, Julius and Kaldi, using DNN-HMM acoustic models and N-gram language models (plus an additional evaluation using RNN re-scoring with Kaldi). Our experimental results show that when using the proposed OOV registration method, modern ASR systems can recognize OOV words without re-training the language model, that the acoustic complexity of OOV words affects OOV recognition, and that differences between the “natural” and the assigned occurrence frequencies of OOV words have little impact on the final recognition results.

Mixed Precision Low-Bit Quantization of Neural Network Language Models for Speech Recognition

State-of-the-art language models (LMs) represented by long-short term memory recurrent neural networks (LSTM-RNNs) and Transformers are becoming increasingly complex and expensive for practical applications. Low-bit neural network quantization provides a powerful solution to dramatically reduce their model size. Current quantization methods are based on uniform precision and fail to account for the varying performance sensitivity at different parts of LMs to quantization errors. To this end, novel mixed precision neural network LM quantization methods are proposed in this paper. The optimal local precision choices for LSTM-RNN and Transformer based neural LMs are automatically learned using three techniques. The first two approaches are based on quantization sensitivity metrics in the form of either the KL-divergence measured between full precision and quantized LMs, or Hessian trace weighted quantization perturbation that can be approximated efficiently using matrix free techniques. The third approach is based on mixed precision neural architecture search. In order to overcome the difficulty in using gradient descent methods to directly estimate discrete quantized weights, alternating direction methods of multipliers (ADMM) are used to efficiently train quantized LMs. Experiments were conducted on state-of-the-art LF-MMI CNN-TDNN systems featuring speed perturbation, i-Vector and learning hidden unit contribution (LHUC) based speaker adaptation on two tasks: Switchboard telephone speech and AMI meeting transcription. The proposed mixed precision quantization techniques achieved "lossless" quantization on both tasks, by producing model size compression ratios of up to approximately 16 times over the full precision LSTM and Transformer baseline LMs, while incurring no statistically significant word error rate increase.

Neural candidate-aware language models for speech recognition

This paper presents novel neural network based language models that can correct automatic speech recognition (ASR) errors by using speech recognizer outputs as a context. Our proposed models, called neural candidate-aware language models (NCALMs), estimate the generative probability of a target sentence while considering ASR outputs including hypotheses and their posterior probabilities. Recently, neural network language models have achieved great success in ASR field because of their ability to learn long-range contexts and model the word representation in continuous space. However, they estimate a sentence probability without considering other candidates and their posterior probabilities, even though the competing hypotheses are available and include important information to increase the speech recognition accuracy. To overcome this limitation, our idea is to utilize ASR outputs in both the training phase and the inference phase. Our proposed models are conditional generative models consisting of a Transformer encoder and a Transformer decoder. The encoder embeds the candidates as context vectors and the decoder estimates a sentence probability given the context vectors. We evaluate the proposed models in Japanese lecture transcription and English conversational speech recognition tasks. Experimental results show that a NCALM has better ASR performance than a system including a deep neural network-hidden Markov model hybrid system. We further improve ASR performance by using a NCALM and a Transformer language model simultaneously.

Language modeling in speech recognition for grammatical error detection based on neural machine translation

Language modeling in speech recognition for grammatical error detection based on neural machine translation

Efficient Search Mechanism from Large Scale Corpora for Domain-Specific Language Modeling in Speech Recognition

With the Internet and the World Wide Web revolution, large corpora in variety of forms are germinating ceaselessly that can be manifested as big data. One obligatory area for the usage of such large corpora is language modeling for large vocabulary continuous speech recognition. Language modeling is an indispensable module in speech recognition architecture, which plays a vital role in reducing the search space during the recognition process. Additionally, the language model that is contiguous to the domain of the speech can dwindle the search space and escalate the recognition accuracy. In this paper, an efficient searching mechanism for domain-specific document retrieval from the large corpora has been elucidated using Elasticsearch which is a distributed and an efficient search engine for big data. This assisted us in tuning the language model in accordance with the domain and also by reducing the search time by more than 90% in comparison to conventional search and retrieval mechanism used in our earlier work. A word level and a phrase level retrieval process for creating domain-specific language model has been implemented. The evaluation of the system is performed on the basis of word error rate (WER) and perplexity (PPL) of the speech recognition system. The results shows nearly 10% decrease on WER and a major reduction in the PPL that helped in boosting the performance of the speech recognition process. From the results, it can be consummated that Elasticsearch is an efficient mechanism for domain specific document retrieval from large corpora rather than using topic modeling toolkits

Automatic estimation of extra-linguistic information in speech and its integration into recurrent neural network-based language models for speech recognition

When one talks with others, he/she often changes the way of lexical choice and the speaking style depending on various contextual factors. For example in Japanese, we often use gender-dependent expressions and, when we speak to elderly people, we use polite expressions. It can be said that in any language, formality or informality of speaking changes drastically depending on the situation where speakers are involved. The aim of automatic speech recognition (ASR) is to convert speech signals into a sequence of words and the above fact indicates that the performance of ASR can be improved by taking the various contextual factors into account in the ASR modules. In this study, at first, we attempt to estimate those contextual factors as para-linguistic or extra-linguistic information and then, to integrate the results into language models based on Recurrent Neural Network (RNN) for speech recognition. In experiments, from an input utterance, i-vector and openSMILE features were extracted to represent speaker identity and speaking style. These acoustically driven features were integrated into the reranking process of RNN-based language models. Reductions of perplexity of the language models were shown to be 1 to 2% relative.

N-Best Re-scoring Approaches for Mandarin Speech Recognition

The predominant language model for speech recognition is n-gram language model, which is locally learned and usually lacks global linguistic information such as long-distance syntactic constraints. We first explore two n-best re-scoring approaches for Mandarin speech recognition to overcome this problem. The first approach is linear re-scoring that can combine several language models from various perspectives. The weights of these models are optimized using minimum error rate learning method. Discriminative approach can also be used for re-scoring with rich syntactic features. To overcome the speech text insufficiency problem for discriminative model, we propose a domain adaptation method that trains the model using Chinese pinyin-to-character conversion dataset. Then we present a cascaded approach to combine the two re-scoring models in pipeline that takes the probability output of linear re-scoring model as the initial weight of the discriminative model. Experimental results show that both re-scoring approaches outperform the baseline system, and the cascaded approach achieves the best performance.

Leveraging topical and positional cues for language modeling in speech recognition

This paper investigates language modeling with topical and positional information for large vocabulary continuous speech recognition. We first compare among a few topic models both theoretically and empirically, including document topic models and word topic models. On the other hand, since for some spoken documents such as broadcast news stories, the composition and the word usage of documents of the same style are usually similar, the documents hence can be separated into partitions consisting of identical rhetoric or topic styles by the literary structures, like introductory remarks, elucidations of methodology or affairs, conclusions of the articles, references or footnotes of reporters, etc. We hence present two position-dependent language models for speech recognition by integrating word positional information into the exiting n-gram and topic models. The experiments conducted on broadcast news transcription seem to indicate that such position-dependent models obtain comparable results to the existing n-gram and topic models.

Leveraging relevance cues for language modeling in speech recognition

Language modeling (LM), providing a principled mechanism to associate quantitative scores to sequences of words or tokens, has long been an interesting yet challenging problem in the field of speech and language processing. The n-gram model is still the predominant method, while a number of disparate LM methods, exploring either lexical co-occurrence or topic cues, have been developed to complement the n-gram model with some success. In this paper, we explore a novel language modeling framework built on top of the notion of relevance for speech recognition, where the relationship between a search history and the word being predicted is discovered through different granularities of semantic context for relevance modeling. Empirical experiments on a large vocabulary continuous speech recognition (LVCSR) task seem to demonstrate that the various language models deduced from our framework are very comparable to existing language models both in terms of perplexity and recognition error rate reductions.

Structured Output Layer Neural Network Language Models for Speech Recognition

This paper extends a novel neural network language model (NNLM) which relies on word clustering to structure the output vocabulary: Structured OUtput Layer (SOUL) NNLM. This model is able to handle arbitrarily-sized vocabularies, hence dispensing with the need for shortlists that are commonly used in NNLMs. Several softmax layers replace the standard output layer in this model. The output structure depends on the word clustering which is based on the continuous word representation determined by the NNLM. Mandarin and Arabic data are used to evaluate the SOUL NNLM accuracy via speech-to-text experiments. Well tuned speech-to-text systems (with error rates around 10%) serve as the baselines. The SOUL model achieves consistent improvements over a classical shortlist NNLM both in terms of perplexity and recognition accuracy for these two languages that are quite different in terms of their internal structure and recognition vocabulary size. An enhanced training scheme is proposed that allows more data to be used at each training iteration of the neural network.

Evaluation of Topic Identification Methods for Arabic Texts and their Combination by using a Corpus Extracted from the Omani Newspaper Alwatan

Topic identification is used in several applications, as adapting language models for speech recognition and machine translation, focusing on a specific use for search engines, etc. Topic identification consists to assign one or several topic labels to a flow of textual data. Labels are chosen from a set of topics fixed a priori. In this paper, we present a study about identifying topics of Arabic texts. For this, a considerable amount of data is needed. Thus, we started by collecting texts from the website of the Omani newspaper “Alwatan”. The result is an Arabic corpus composed of more than 9000 articles corresponding to nearly 10 millions words. The considered topics in our experiments are: Culture, Religion, Economy, Local news, International news and sports. Some of the methods presented in this study, are well known in the text categorization community, as TFIDF classifier and kNN “k Nearest Neighbors”. The objective to use these methods is to compare them to TR-classifier “TRiggers-based classifier”, a new method that we have proposed, which is based on computing triggers or the Average Mutual Information of each couple of words. In order to enhance performances, we have combined results of the three methods by using three approaches: Majority Vote, Enhanced Majority Vote and Linear Combination.

Dirichlet Class Language Models for Speech Recognition

Latent Dirichlet allocation (LDA) was successfully developed for document modeling due to its generalization to unseen documents through the latent topic modeling. LDA calculates the probability of a document based on the bag-of-words scheme without considering the order of words. Accordingly, LDA cannot be directly adopted to predict words in speech recognition systems. This work presents a new Dirichlet class language model (DCLM), which projects the sequence of history words onto a latent class space and calculates a marginal likelihood over the uncertainties of classes, which are expressed by Dirichlet priors. A Bayesian class-based language model is established and a variational Bayesian procedure is presented for estimating DCLM parameters. Furthermore, the long-distance class information is continuously updated using the large-span history words and is dynamically incorporated into class mixtures for a cache DCLM. Different language models are experimentally evaluated using the Wall Street Journal (WSJ) corpus. The amount of training data and the size of vocabulary are evaluated. We find that the cache DCLM effectively characterizes the unseen -gram events and stores the class information for long-distance language modeling. This approach outperforms the other class-based and topic-based language models in terms of perplexity and recognition accuracy. The DCLM and cache DCLM achieved relative gain of word error rate by 3% to 5% over the LDA topic-based language model with different sizes of training data .

Topic tracking language model for speech recognition

In a real environment, acoustic and language features often vary depending on the speakers, speaking styles and topic changes. To accommodate these changes, speech recognition approaches that include the incremental tracking of changing environments have attracted attention. This paper proposes a topic tracking language model that can adaptively track changes in topics based on current text information and previously estimated topic models in an on-line manner. The proposed model is applied to language model adaptation in speech recognition. We use the MIT OpenCourseWare corpus and Corpus of Spontaneous Japanese in speech recognition experiments, and show the effectiveness of the proposed method.

Using knowledge of misunderstandings to increase the robustness of spoken dialogue systems

This paper proposes a new technique to increase the robustness of spoken dialogue systems employing an automatic procedure that aims to correct frames incorrectly generated by the system’s component that deals with spoken language understanding. To do this the technique carries out a training that takes into account knowledge of previous system misunderstandings. The correction is transparent for the user as he is not aware of some mistakes made by the speech recogniser and thus interaction with the system can proceed more naturally. Experiments have been carried out using two spoken dialogue systems previously developed in our lab: Saplen and Viajero, which employ prompt-dependent and prompt-independent language models for speech recognition. The results obtained from 10,000 simulated dialogues show that the technique improves the performance of the two systems for both kinds of language modelling, especially for the prompt-independent language model. Using this type of model the Saplen system increases sentence understanding by 19.54%, task completion by 26.25%, word accuracy by 7.53%, and implicit recovery of speech recognition errors by 20.3%, whereas for the Viajero system these figures increase by 14.93%, 18.06%, 6.98% and 15.63%, respectively.

Joint acoustic and language modeling for speech recognition

In a traditional model of speech recognition, acoustic and linguistic information sources are assumed independent of each other. Parameters of hidden Markov model (HMM) and n-gram are separately estimated for maximum a posteriori classification. However, the speech features and lexical words are inherently correlated in natural language. Lacking combination of these models leads to some inefficiencies. This paper reports on the joint acoustic and linguistic modeling for speech recognition by using the acoustic evidence in estimation of the linguistic model parameters, and vice versa, according to the maximum entropy (ME) principle. The discriminative ME (DME) models are exploited by using features from competing sentences. Moreover, a mutual ME (MME) model is built for sentence posterior probability, which is maximized to estimate the model parameters by characterizing the dependence between acoustic and linguistic features. The N-best Viterbi approximation is presented in implementing DME and MME models. Additionally, the new models are incorporated with the high-order feature statistics and word regularities. In the experiments, the proposed methods increase the sentence posterior probability or model separation. Recognition errors are significantly reduced in comparison with separate HMM and n-gram model estimations from 32.2% to 27.4% using the MATBN corpus and from 5.4% to 4.8% using the WSJ corpus (5K condition).

Robust Understanding in Multimodal Interfaces

Multimodal grammars provide an effective mechanism for quickly creating integration and understanding capabilities for interactive systems supporting simultaneous use of multiple input modalities. However, like other approaches based on hand-crafted grammars, multimodal grammars can be brittle with respect to unexpected, erroneous, or disfluent input. In this article, we show how the finite-state approach to multimodal language processing can be extended to support multimodal applications combining speech with complex freehand pen input, and evaluate the approach in the context of a multimodal conversational system (MATCH). We explore a range of different techniques for improving the robustness of multimodal integration and understanding. These include techniques for building effective language models for speech recognition when little or no multimodal training data is available, and techniques for robust multimodal understanding that draw on classification, machine translation, and sequence edit methods. We also explore the use of edit-based methods to overcome mismatches between the gesture stream and the speech stream.

Review of language modeling for speech recognition(<Special issue>recent progress and future prospects of automatic speech recognition research)

Review of language modeling for speech recognition(<Special issue>recent progress and future prospects of automatic speech recognition research)