Improving Voice Search Using Forward-Backward LVCSR System Combination

Japanese large-vocabulary continuous-speech recognition using a newspaper corpus and broadcast news

—This paper presents the main improvements brought recently to the SpeeD automatic speech recognition system. Several aspects, such as speech and text resources acquisition, noise-robust speech features and feature transforms are discussed. All the updates in our ASR system are accompanied by experimental results illustrating significant improvements: between 30% and 35% relative WER reductions for various case studies (read/spontaneous speech, noisy/clean speech). In the last part of the paper, our ASR system is also compared with Google’s ASR system and a brief analysis of the results is presented. Keywords -ASR; LVCSR; noise-robust speech recognition; PNCC; LDA. I. I NTRODUCTION Large vocabulary continuous speech recognition (LVCSR) is still an unsolved topic for many languages. The reasons for this are (i) there is a lack of acoustic and linguistic resources needed for development (it is the case of so-called under-resourced languages) and (ii) the scientific research community is not stimulated by any national or international evaluation campaigns (as opposed to languages such as English, French or Chinese). The Romanian language is affected by both the aforementioned problems. In this context, the development of speech and language resources for automatic speech recognition (ASR) is a critical issue that must be addressed to push forward the research in this direction and create LVCSR systems comparable to those available for other languages. This is one of the main goals of the Speech and Dialogue (SpeeD) research group

This paper deals with the use of a stochastic context‐free grammar (SCFG) for large vocabulary continuous speech recognition; in particular, an SCFG with phrase‐level dependency rules is built. Unlike n‐gram models, the SCFG can describe not only local constraints but also global constraints pertaining to the sentence as a whole, thus making possible language models with great expressive power. However, the inside‐outside algorithm must be used for estimation of the SCFG parameters, which involves a great amount of calculation, proportional to the third power of the number of nonterminal symbols and of the input string length. Hence, due to problems in dealing with extensive text corpora, the SCFG has hardly been applied as a language model for very large vocabulary continuous speech recognition. The proposed phrase‐level dependency SCFG allows a significant reduction of the computational load. In experiments with the EDR corpus, the proposed method proved effective. In experiments with the Mainichi corpus, a large‐scale phrase‐level dependency SCFG was built for a very large vocabulary continuous speech recognition system. Speech recognition tests with a vocabulary of about 5000 words showed that the proposed method could not compare with the trigram model in performance; however, when it was used in combination with a trigram model, the error rate was reduced by 14% compared to the trigram model alone. © 2002 Wiley Periodicals, Inc. Syst Comp Jpn, 33(13): 48–59, 2002; Published online in Wiley InterScience (www.interscience.wiley.com). DOI 10.1002/scj.1172

The diachronic nature of broadcast news data leads to the problem of out-of-vocabulary (OOV) words in large vocabulary continuous speech recognition (LVCSR) systems. Analysis of OOV words reveals that a majority of them are proper names (PNs). However, PNs are important for automatic indexing of audio–video content and for obtaining reliable automatic transcriptions. In this paper, we focus on the problem of OOV PNs in diachronic audio documents. To enable the recovery of the PNs missed by the LVCSR system, relevant OOV PNs are retrieved by exploiting the semantic context of the LVCSR transcriptions. For retrieval of OOV PNs, we explore topic and semantic context derived from latent Dirichlet allocation (LDA) topic models, continuous word vector representations and the neural bag-of-words (NBOW) model which is capable of learning task specific word and context representations. We propose a neural bag-of-weighted words (NBOW2) model which learns to assign higher weights to words that are important for retrieval of an OOV PN. With experiments on French broadcast news videos, we show that the NBOW and NBOW2 models outperform the methods based on raw embeddings from LDA and Skip-gram models. Combining the NBOW and NBOW2 models gives a faster convergence during training. Second pass speech recognition experiments, in which the LVCSR vocabulary and language model are updated with the retrieved OOV PNs, demonstrate the effectiveness of the proposed context models.

Subspace Gaussian mixture based language modeling for large vocabulary continuous speech recognition

In large vocabulary continuous speech recognition (LVCSR) for agglutinative and inflectional languages, we encounter problems due to theoretically infinite full-word lexicon size. Sub-word lexicon units may be utilized to dramatically reduce the out-of-vocabulary rate in test data. One can develop language models based on sub-word units to perform LVCSR. However, it has not always been beneficial to use sub-word lexicon units, since shorter units have higher acoustic confusability among them and language model history is effectively shorter as compared to the history in full-word language models. To reduce the aforementioned problems, we propose using the longest possible sub-word units in our lexicon, namely half-words and full-words only. We also incorporate linguistic rules of half-word combination into our statistical language model. The language constraints are represented with a rule-based WFSM which can be combined with an N-gram language model to yield a better and smaller language model. We study the performance of the proposed system for Turkish LVCSR, when the language constraint takes the form of enforcing vowel harmony between stems and endings. We also introduce novel error-rate metrics that are more appropriate than word-error-rate for agglutinative languages. Using half-words with a bi-gram model yields a significant reduction in word-error-rate as compared to a bi-gram full-word model. In addition, combining a tri-gram half-word language model with the vowel-harmony WFSM improves the accuracy further when rescoring the bi-gram lattices.

We like to conversate with other people using both sounds and visuals, as our perception of speech is bimodal. Essentially echoing the same speech structure, we manage to integrate the two modalities and often understand the message better than with the eyes closed. In this work we would like to learn more about the visual nature of speech, coined lip-reading, and to make use of it towards better automatic speech recognition systems. Recent developments in the Machine Learning area, together with the release of suitable audio-visual datasets aimed at large vocabulary continuous speech recognition, have led to a renewal of the lip-reading topic, and allow us to address the recurring question of how to better integrate visual and acoustic speech.

A fast and memory-efficient N-gram language model lookup method for large vocabulary continuous speech recognition

This paper describes a reading quality scoring system based on large vocabulary continuous speech recognition (LVCSR). Our previous scoring system was based on forced alignment. A disadvantage of forced alignment based system is it can hardly catch huge kinds of reading miscues, while LVCSR based system avoids this disadvantage. The most challenge was that the LVCSR recognition rate was low on our corpus. To improve the recognition rate, we optimized our LVCSR engine for passage scoring tasks by presenting a novel dynamic language model (LM) constructing algorithm. The optimized LVCSR's recognition rate on test speech data was 70.2%, while the recognition rate of the original LVCSR on the same database was 37.9%. Our scoring method was to align the text reference and the confusion network generated from the LVCSR decoding result. The LVCSR based system reduced the scoring error rate of the baseline system by 14.5% relatively.

In this dissertation, a novel approach of integrating template matching with statistical modeling is proposed to improve continuous speech recognition. Hidden Markov Modeling (HMMs) has been the dominant approach in statistical speech recognition since it provides a principled way of jointly modeling speech spectral variations and time dynamics. However, HMMs have the shortcoming of assuming the observations being independent within each state, which makes it ineffective in modeling the details of speech temporal evolutions that are important for characterizing nonstationary speech sounds. Template-based methods make comparisons between a test pattern and the templates derived from training data, and therefore they are able to capture speech dynamics and time correlation of speech frames better than HMM based methods. However, template matching requires large memory space and computational time since feature vectors of training data need to be stored in computer memory for access at the recognition stage, which is difficult in large vocabulary continuous speech recognition (LVCSR). Our proposed approach takes advantages of both statistical modeling and template matching, which overcomes the weakness of conventional template-based method and is feasible for LVCSR. We use multiple Gaussian Mixture Model (GMM) indices to represent each frame of speech templates, and define the template unit to be context-dependent phone segments (triphone context). We also use phonetic decision trees borrowed from those commonly used in HMMs to tie triphone templates and predict triphones unseen in training data. Two local distances, log likelihood ratio (LLR) and Kullback-Leibler (KL) divergence, are proposed for dynamic time warping (DTW) based template matching. In order to reduce computational complexity and storage space, we propose methods of minimum distance template selection (MDTS) and maximum log-likelihood template selection (MLTS), and investigate a template compression method on top of template selection to further improve recognition performance. The template based methods were used to rescore lattices generated by baseline HMMs on the tasks of TIMIT continuous phone recognition and teleheath LVCSR and experimental results demonstrated that the proposed approach of integrating template matching with statistical modeling significantly improved recognition performances over the HMM baselines. The template selection methods also provided significant recognition accuracy improvements over the HMM baseline while largely reducing the computation and storage complexities. When all templates or MDTS were used, using the LLR local distance obtained better recognition performance than the KL divergence local distance. For MLTS and template compression, KL divergence local distance provided better performance than the LLR local distance, and the template compression method made further improvements over KL based MLTS. Since the templates were constructed based on the GMM indices extracted from HMM baselines, we also validated the effectiveness of the proposed template methods based on enhanced HMM baselines. Experimental results showed that LLR based all template method was able to consistently improve TIMIT phone recognition accuracies based on four enhanced HMM baselines. Prosodic features such as duration, energy, and pitch can reflect longer span information of speech than conventional single frame vectors but they have commonly been ignored by HMMs. Template based methods provide possibilities to conveniently integrate prosodic features into speech recognition, which has not been well studied in the past. In this dissertation, we investigate combining template based methods with the speech prosodic features of duration, energy and pitch to further improve speech recognition accuracy. The scores of prosodic information were computed by a GMM based method and a non-parametric method, and the prosodic scores were combined with the acoustic scores in triphone template matching. Experimental results obtained on the telehealth task showed that prosodic information had positive effects on vowel sound recognition.

Semantic and phonetic automatic reconstruction of medical dictations

The Tigrinya language is agglutinative and has a large number of inflected and derived forms of words. Therefore a Tigrinya large vocabulary continuous speech recognition system often has a large number of different units and a high out-of-vocabulary (OOV) rate if a word is used as a recognition unit of a language model (LM) and lexicon. Therefore a morpheme-based approach has often been used and a morpheme is used as the recognition unit to reduce the high OOV rate. This paper presents an automatic speech recognition experiment conducted to see the effect of OOV words on the performance speech recognition system for Tigrinya. We tried to solve the OOV problem by using morphemes as lexicon and language model units. It has been found that the morpheme-based recognition system is better lexical and language modeling units than words. An absolute improvement (in word recognition accuracy) of 3.45 token and 8.36 types has been obtained as a result of using a morph-based vocabulary.

Masked Language Models (MLM) are self-supervised neural networks trained to fill in the blanks in a given sentence with masked tokens. Despite the tremendous success of MLMs for various text based tasks, they are not robust for spoken language understanding, especially for spontaneous conversational speech recognition noise. In this work we introduce Warped Language Models (WLM) in which input sentences at training time go through the same modifications as in MLM, plus two additional modifications, namely inserting and dropping random tokens. These two modifications extend and contract the sentence in addition to the modifications in MLMs, hence the word "warped" in the name. The insertion and drop modification of the input text during training of WLM resemble the types of noise due to Automatic Speech Recognition (ASR) errors, and as a result WLMs are likely to be more robust to ASR noise. Through computational results we show that natural language understanding systems built on top of WLMs perform better compared to those built based on MLMs, especially in the presence of ASR errors.

This paper reports the development of a multi-genre Urdu Broadcast (BC) corpus and a Large Vocabulary Continuous Speech Recognition (LVCSR) system. BC speech corpus of 98 hours from 453 speakers is collected and annotated. For acoustic modeling, Time-delay Neural Network (TDNN) is developed with prior Gaussian Mixture Model-Hidden Markov Model (GMM-HMM) training and alignments. For the language model, 3-gram, 4-gram and Recurrent Neural Network (RNN) based models are developed on a text corpus of 188 million words. The developed models are tested on 4.3 hours of unseen BC multi-genre speech dataset and the best Word Error Rate (WER) 18.59% is achieved using RNN based Language Model (LM). Moreover, a detailed word error analysis is carried out to compare the errors made by humans and the Automatic Speech Recognition (ASR) System. The results showed a similar behavior of word misrecognitions by both humans and ASR.

In this paper, we present an approach to Spoken Language Understanding (SLU) where we perform a combination of multiple hypotheses from several Automatic Speech Recognizers (ASRs) in order to reduce the impact of recognition errors in the SLU module. This combination is performed using a Grammatical Inference algorithm that provides a generalization of the input sentences by means of a weighted graph of words. We have also developed a specific SLU algorithm that is able to process these graphs of words according to a stochastic semantic modelling.The results show that the combinations of several hypotheses from the ASR module outperform the results obtained by taking just the 1-best transcription.