HMM/SVM segmentation and labelling of Arabic speech for speech recognition applications

Building a large vocabulary continuous speech recognition (LVCSR) system requires a lot of hours of segmented and labelled speech data. Arabic language, as many other low-resourced languages, lacks such data, but the use of automatic segmentation proved to be a good alternative to make these resources available. In this paper, we suggest the combination of hidden Markov models (HMMs) and support vector machines (SVMs) to segment and to label the speech waveform into phoneme units. HMMs generate the sequence of phonemes and their frontiers; the SVM refines the frontiers and corrects the labels. The obtained segmented and labelled units may serve as a training set for speech recognition applications. The HMM/SVM segmentation algorithm is assessed using both the hit rate and the word error rate (WER); the resulting scores were compared to those provided by the manual segmentation and to those provided by the well-known embedded learning algorithm. The results show that the speech recognizer built upon the HMM/SVM segmentation outperforms in terms of WER the one built upon the embedded learning segmentation of about 0.05%, even in noisy background.     

Maximum entropy direct models for speech recognition

Traditional statistical models for speech recognition have mostly been based on a Bayesian framework using generative models such as hidden Markov models (HMMs). This paper focuses on a new framework for speech recognition using maximum entropy direct modeling, where the probability of a state or word sequence given an observation sequence is computed directly from the model. In contrast to HMMs, features can be asynchronous and overlapping. This model therefore allows for the potential combination of many different types of features, which need not be statistically independent of each other. In this paper, a specific kind of direct model, the maximum entropy Markov model (MEMM), is studied. Even with conventional acoustic features, the approach already shows promising results for phone level decoding. The MEMM significantly outperforms traditional HMMs in word error rate when used as stand-alone acoustic models. Preliminary results combining the MEMM scores with HMM and language model scores show modest improvements over the best HMM speech recognizer.     

Prosody dependent speech recognition on radio news corpus of American English

Does prosody help word recognition? This paper proposes a novel probabilistic framework in which word and phoneme are dependent on prosody in a way that reduces word error rates (WER) relative to a prosody-independent recognizer with comparable parameter count. In the proposed prosody-dependent speech recognizer, word and phoneme models are conditioned on two important prosodic variables: the intonational phrase boundary and the pitch accent. An information-theoretic analysis is provided to show that prosody dependent acoustic and language modeling can increase the mutual information between the true word hypothesis and the acoustic observation by exciting the interaction between prosody dependent acoustic model and prosody dependent language model. Empirically, results indicate that the influence of these prosodic variables on allophonic models are mainly restricted to a small subset of distributions: the duration PDFs (modeled using an explicit duration hidden Markov model or EDHMM) and the acoustic-prosodic observation PDFs (normalized pitch frequency). Influence of prosody on cepstral features is limited to a subset of phonemes: for example, vowels may be influenced by both accent and phrase position, but phrase-initial and phrase-final consonants are independent of accent. Leveraging these results, effective prosody dependent allophonic models are built with minimal increase in parameter count. These prosody dependent speech recognizers are able to reduce word error rates by up to 11% relative to prosody independent recognizers with comparable parameter count, in experiments based on the prosodically-transcribed Boston Radio News corpus.     

Within-word pronunciation variation modeling for Arabic ASRs: a direct data-driven approach

Pronunciation variation is a major obstacle in improving the performance of Arabic automatic continuous speech recognition systems. This phenomenon alters the pronunciation spelling of words beyond their listed forms in the pronunciation dictionary, leading to a number of out of vocabulary word forms. This paper presents a direct data-driven approach to model within-word pronunciation variations, in which the pronunciation variants are distilled from the training speech corpus. The proposed method consists of performing phoneme recognition, followed by a sequence alignment between the observation phonemes generated by the phoneme recognizer and the reference phonemes obtained from the pronunciation dictionary. The unique collected variants are then added to dictionary as well as to the language model. We started with a Baseline Arabic speech recognition system based on Sphinx3 engine. The Baseline system is based on a 5.4 hours speech corpus of modern standard Arabic broadcast news, with a pronunciation dictionary of 14,234 canonical pronunciations. The Baseline system achieves a word error rate of 13.39%. Our results show that while the expanded dictionary alone did not add appreciable improvements, the word error rate is significantly reduced by 2.22% when the variants are represented within the language model.     

Continuous space language models

This paper describes the use of a neural network language model for large vocabulary continuous speech recognition. The underlying idea of this approach is to attack the data sparseness problem by performing the language model probability estimation in a continuous space. Highly efficient learning algorithms are described that enable the use of training corpora of several hundred million words. It is also shown that this approach can be incorporated into a large vocabulary continuous speech recognizer using a lattice rescoring framework at a very low additional processing time. The neural network language model was thoroughly evaluated in a state-of-the-art large vocabulary continuous speech recognizer for several international benchmark tasks, in particular the Nist evaluations on broadcast news and conversational speech recognition. The new approach is compared to four-gram back-off language models trained with modified Kneser–Ney smoothing which has often been reported to be the best known smoothing method. Usually the neural network language model is interpolated with the back-off language model. In that way, consistent word error rate reductions for all considered tasks and languages were achieved, ranging from 0.4% to almost 1% absolute.     

Study of sub-word acoustical models for <Emphasis Type="Italic">Kannada</Emphasis> isolated word recognition system

The speech recognition system basically extracts the textual information present in the speech. In the present work, speaker independent isolated word recognition system for one of the south Indian language—Kannada has been developed. For European languages such as English, large amount of research has been carried out in the context of speech recognition. But, speech recognition in Indian languages such as Kannada reported significantly less amount of work and there are no standard speech corpus readily available. In the present study, speech database has been developed by recording the speech utterances of regional Kannada news corpus of different speakers. The speech recognition system has been implemented using the Hidden Markov Tool Kit. Two separate pronunciation dictionaries namely phone based and syllable based dictionaries are built in-order to design and evaluate the performances of phone-level and syllable-level sub-word acoustical models. Experiments have been carried out and results are analyzed by varying the number of Gaussian mixtures in each state of monophone Hidden Markov Model (HMM). Also, context dependent triphone HMM models have been built for the same Kannada speech corpus and the recognition accuracies are comparatively analyzed. Mel frequency cepstral coefficients along with their first and second derivative coefficients are used as feature vectors and are computed in acoustic front-end processing. The overall word recognition accuracy of 60.2 and 74.35 % respectively for monophone and triphone models have been obtained. The study shows a good improvement in the accuracy of isolated-word Kannada speech recognition system using triphone HMM models compared to that of monophone HMM models.     

Thai spelling analysis for automatic spelling speech recognition

Spelling speech recognition can be applied for several purposes including enhancement of speech recognition systems and implementation of name retrieval systems. This paper presents a Thai spelling analysis to develop a Thai spelling speech recognizer. The Thai phonetic characteristics, alphabet system and spelling methods have been analyzed. As a training resource, two alternative corpora, a small spelling speech corpus and an existing large continuous speech corpus, are used to train hidden Markov models (HMMs). Then their recognition results are compared to each other. To solve the problem of utterance speed difference between spelling utterances and continuous speech utterances, the adjustment of utterance speed has been taken into account. Two alternative language models, bigram and trigram, are used for investigating performance of spelling speech recognition. Our approach achieves up to 98.0% letter correction rate, 97.9% letter accuracy and 82.8% utterance correction rate when the language model is trained based on trigram and the acoustic model is trained from the small spelling speech corpus with eight Gaussian mixtures.     

HMM based online handwriting recognition

Hidden Markov model (HMM) based recognition of handwriting is now quite common, but the incorporation of HMM's into a complex stochastic language model for handwriting recognition is still in its infancy. We have taken advantage of developments in the speech processing field to build a more sophisticated handwriting recognition system. The pattern elements of the handwriting model are subcharacter stroke types modeled by HMMs. These HMMs are concatenated to form letter models, which are further embedded in a stochastic language model. In addition to better language modeling, we introduce new handwriting recognition features of various kinds. Some of these features have invariance properties, and some are segmental, covering a larger region of the input pattern. We have achieved a writer independent recognition rate of 94.5% on 3,823 unconstrained handwritten word samples from 18 writers covering a 32 word vocabulary     

Effect of acoustic and linguistic contexts on human and machine speech recognition

We compared the performance of an automatic speech recognition system using n-gram language models, HMM acoustic models, as well as combinations of the two, with the word recognition performance of human subjects who either had access to only acoustic information, had information only about local linguistic context, or had access to a combination of both. All speech recordings used were taken from Japanese narration and spontaneous speech corpora.Humans have difficulty recognizing isolated words taken out of context, especially when taken from spontaneous speech, partly due to word-boundary coarticulation. Our recognition performance improves dramatically when one or two preceding words are added. Short words in Japanese mainly consist of post-positional particles (i.e. wa, ga, wo, ni, etc.), which are function words located just after content words such as nouns and verbs. So the predictability of short words is very high within the context of the one or two preceding words, and thus recognition of short words is drastically improved. Providing even more context further improves human prediction performance under text-only conditions (without acoustic signals). It also improves speech recognition, but the improvement is relatively small.Recognition experiments using an automatic speech recognizer were conducted under conditions almost identical to the experiments with humans. The performance of the acoustic models without any language model, or with only a unigram language model, were greatly inferior to human recognition performance with no context. In contrast, prediction performance using a trigram language model was superior or comparable to human performance when given a preceding and a succeeding word. These results suggest that we must improve our acoustic models rather than our language models to make automatic speech recognizers comparable to humans in recognition performance under conditions where the recognizer has limited linguistic context.     

Syllable modeling in continuous speech recognition for Tamil language

In automatic speech recognition, the phone has probably been a dominating sub-word unit for more than one decade. Context Dependent phone or triphone modeling accounts for contextual variations between adjacent phones and state tying addresses modeling of triphones that are not seen during training. Recently, syllable is gaining momentum as a new sub-word unit. Syllable being a larger unit than a phone addresses the severe contextual variations between phones within it. Therefore, it is more stable than a phone and models pronunciation variability in a systematic way. Tamil language has challenging features like agglutination and morpho-phonology. In this paper, attempts have been made to provide solutions to these issues by using the syllable as a sub-word unit in an acoustic model. Initially, a small vocabulary context independent word models and a medium vocabulary context dependent phone models are developed. Subsequently, an algorithm based on prosodic syllable is proposed and two experiments have been conducted. First, syllable based context independent models have been trained and tested. Despite large number of syllables, this system has performed reasonably well compared to context independent word models in terms of word error rate and out of vocabulary words. Subsequently, in the second experiment, syllable information is integrated in conventional triphone modeling wherein cross-syllable triphones are replaced with monophones and the number of context dependent phone models is reduced by 22.76% in untied units. In spite of reduction in the number of models, the accuracy of the proposed system is comparable to that of the baseline triphone system.     

Large vocabulary word recognition using context-dependent allophonic hidden Markov models

In this paper, we report our development of context-dependent allophonic hidden Markov models (HMMs) implemented in a 75 000-word speaker-dependent Gaussian-HMM recognizer. The context explored is the immediate left and/or right adjacent phoneme. To achieve reliable estimation of the model parameters, phonemes are grouped into classes based on their expected co-articulatory effects on neighboring phonemes. Only five separate preceding and following contexts are identified explicitly for each phoneme. By grouping the contexts we ensure that they occur frequently enough in the training data to allow reliable estimation of the parameters of the HMM representing the context-dependent units. Further improvement in the estimation reliability is obtained by tying the covariance matrices in the HMM output distributions across all contexts. Speech recognition experiments show that when a large amount of data (e.g. over 2500 words) is used to train context-dependent HMMs, the word recognition error rate is reduced by 33%, compared with the context-independent HMMs. For smaller amounts of training data the error reduction becomes less significant.     

Keyword detection in conversational speech utterances using hidden Markov model based continuous speech recognition

This paper describes a set of modeling techniques for detecting a small vocabulary of keywords in running conversational speech. The techniques are applied in the context of a hidden Markov model (HMM) based continuous speech recognition (CSR) approach to keyword spotting. The word spotting task is derived from the Switchboard conversational speech corpus, and involves unconstrained conversational speech utterances spoken over the public switched telephone network. The utterances in this task contain many of the artifacts that are characteristic of unconstrained speech as it appears in many telecommunications based automatic speech recognition (ASR) applications. Results are presented for an experimental study that was performed on this task. Performance was measured by computing the percentage correct keyword detection over a range of false alarm rates evaluated over 2·2 h of speech for a 20 keyword vocabulary. The results of the study demonstrate the importance of several techniques. These techniques include the use of decision tree based allophone clustering for defining acoustic subword units, different representations for non-vocabulary words appearing in the input utterance, and the definition of simple language models for keyword detection. Decision tree based allophone clustering resulted in a significant increase in keyword detection performance over that obtained using tri-phone based subword units while at the same time reducing the size of the inventory of subword acoustic models by 40%. More complex representations of non-vocabulary speech were also found to significantly improve keyword detection performance; however, these representations also resulted in a significant increase in computational complexity.     

Airwriting: a wearable handwriting recognition system

We present a wearable input system which enables interaction through 3D handwriting recognition. Users can write text in the air as if they were using an imaginary blackboard. The handwriting gestures are captured wirelessly by motion sensors applying accelerometers and gyroscopes which are attached to the back of the hand. We propose a two-stage approach for spotting and recognition of handwriting gestures. The spotting stage uses a support vector machine to identify those data segments which contain handwriting. The recognition stage uses hidden Markov models (HMMs) to generate a text representation from the motion sensor data. Individual characters are modeled by HMMs and concatenated to word models. Our system can continuously recognize arbitrary sentences, based on a freely definable vocabulary. A statistical language model is used to enhance recognition performance and to restrict the search space. We show that continuous gesture recognition with inertial sensors is feasible for gesture vocabularies that are several orders of magnitude larger than traditional vocabularies for known systems. In a first experiment, we evaluate the spotting algorithm on a realistic data set including everyday activities. In a second experiment, we report the results from a nine-user experiment on handwritten sentence recognition. Finally, we evaluate the end-to-end system on a small but realistic data set.     

基于上下文三音素DBN模型的连续语音识别

考虑连续语音中的协同发音问题,提出基于词内扩展的单流上下文相关三音素动态贝叶斯网络（SS-DBN-TRI）模型和词间扩展的单流上下文相关三音素DBN（SS-DBN-TRI-CON）模型。SS-DBN-TRI模型是Bilmes提出单流DBN（SS-DBN）模型的改进,采用词内上下文相关三音素节点替代单音素节点,每个词由它的对应三音素单元构成,而三音素单元和观测向量相联系;SS-DBN-TRI-CON模型基于SS-DBN模型,通过增加当前音素的前音素节点和后音素节点,构成一个新的词间扩展的三音素变量节点,新的三音素节点和观测向量相联系,采用高斯混合模型来描述,采用数字连续语音数据库的实验结果表明：SS-DBN-TRI-CON具备最好的语音识别性能。     

基于HMM的维吾尔语连续语音识别系统

维吾尔语是黏着性语言,利用丰富的词缀可以用同样的词干产生超大词汇,给维吾尔语语音识别的研究工作带来了很大困难。结合维吾尔语自身特点,建立了维吾尔语连续语音语料库,利用HTK（HMMToolKit）工具实现了基于隐马尔可夫模型（HMM）的维吾尔语连续语音识别系统。在声学层,选取三音子作为基本的识别单元,建立了维吾尔语的三音子声学模型,并使用决策树、三音子绑定、修补哑音、增加高斯混合分量等方法提高模型的识别精度。在语言层,使用了适合于维吾尔语语音特征的基于统计的二元文法语言模型。最后,利用该系统进行了维吾尔语连续语音识别实验。     

Use of minimum duration and energy contour for phonemes to improve large vocabulary isolated-word recognition

Many acoustic misrecognitions in our 86 000-word speaker-trained isolated-word recognizer are due to phonemic hidden Markov models (phoneme models) mapping to short segments of speech. When we force these models to map to longer segments corresponding to the observed minimum durations for the phonemes, then the likelihood of the incorrect phoneme sequences drops dramatically. This drop in the likelihood of the incorrect words results in significant reduction in the acoustic recognition¹ error rate. Even in cases where acoustic recognition performance is unchanged, the likelihood of the correct word choice improves relative to the incorrect word choices, resulting in significant reduction in recognition error rate with the language model. On nine speakers, the error rate for acoustic recognition reduces from 18·6 to 17·3%, while the error rate with the language model reduces from 9·2 to 7·2%.We have also improved the phoneme models by correcting the segmentation of the phonemes in the training set. During training, the boundaries between phonemes are not marked accurately. We use energy to correct these boundaries. Application of an energy threshold improves the segment boundaries between stops and sonorants (vowels, liquids and glides), between fricatives and sonorants, between affricates and sonorants and between breath noise and sonorants. Training the phoneme models with these segmented phonemes results in models which increase recognition accuracy significantly. On two speakers, the error rate for acoustic recognition reduces from 26·5 to 23·1%, while the error rate with the language model reduces from 11·3 to 8·8%. This reduction in error rate is in addition to the error rate reductions obtained by imposing minimum duration constraints. The overall reduction in errors for these two speakers using minimum durations and energy thresholds is from 27·3 to 23·1% for acoustic recognition, and from 14·3 to 8·8% with the language model.     

维吾尔语音素的声学特征分析

该文对不同语速下,人工标注的维吾尔语连续语音语料中各音素进行共振峰频率、音长、音强的统计分析,并完成辅-元结构下的塞音、塞擦音的声学特征分析。该文通过美尔频率倒谱系数与共振峰频率等声学特征的融合及模型状态数的修改,对维吾尔语音素识别的声学模型进行了改进,并验证了不同声学特征对音素识别的影响。相比于基线系统,改进后声学模型的识别率取得一定提升。同时,利用语音学知识分析维吾尔语易混淆音素产生原因,为音素识别声学模型的进一步改进提供参考依据。     

Combining language models in the input interface of a spoken dialogue system

This paper presents a new technique to enhance the performance of the input interface of spoken dialogue systems based on a procedure that combines during speech recognition the advantages of using prompt-dependent language models with those of using a language model independent of the prompts generated by the dialogue system. The technique proposes to create a new speech recognizer, termed contextual speech recognizer, that uses a prompt-independent language model to allow recognizing any kind of sentence permitted in the application domain, and at the same time, uses contextual information (in the form of prompt-dependent language models) to take into account that some sentences are more likely to be uttered than others at a particular moment of the dialogue. The experiments show the technique allows enhancing clearly the performance of the input interface of a previously developed dialogue system based exclusively on prompt-dependent language models. But most important, in comparison with a standard speech recognizer that uses just one prompt-independent language model without contextual information, the proposed recognizer allows increasing the word accuracy and sentence understanding rates by 4.09% and 4.19% absolute, respectively. These scores are slightly better than those obtained using linear interpolation of the prompt-independent and prompt-dependent language models used in the experiments.     

Look-ahead techniques for fast beam search

This paper presents two look-ahead techniques for speeding up large vocabulary continuous speech recognition. These two techniques are the language model look-ahead and the phoneme look-ahead; both are incorporated into the pruning process of the time-synchronous one-pass beam search algorithm. The search algorithm is based on a tree-organized pronunciation lexicon in connection with a bigram language model. Both look-ahead techniques have been tested on the 20 000-word NAB’94 task (ARPA North American Business Corpus). The recognition experiments show that the combination of bigram language model look-ahead and phoneme look-ahead reduces the size of search space by a factor of about 30 and the computational effort by a factor of 5 without affecting the word recognition accuracy in comparison with no look-ahead pruning technique.