Integrating Language Identification to improve Multilingual Speech Recognition

The process of determining the language of a speech utterance is called Language Identification (LID). This task can be very challenging as it has to take into account various language-specific aspects, such as phonetic, phonotactic, vocabulary and grammar-related cues. In multilingual speech recognition we try to find the most likely word sequence that corresponds to an utterance where the language is not known a priori. This is a considerably harder task compared to monolingual speech recognition and it is common to use LID to estimate the current language. In this project we present two general approaches for LID and describe how to integrate them into multilingual speech recognizers. The first approach uses hierarchical multilayer perceptrons to estimate language posterior probabilities given the acoustics in combination with hidden Markov models. The second approach evaluates the output of a multilingual speech recognizer to determine the spoken language. The research is applied to the MediaParl speech corpus that was recorded at the Parliament of the canton of Valais, where people switch from Swiss French to Swiss German or vice versa. Our experiments show that, on that particular data set, LID can be used to significantly improve the performance of multilingual speech recognizers. We will also point out that ASR dependent LID approaches yield the best performance due to higher-level cues and that our systems perform much worse on non-native data

Multilingual Training and Adaptation in Speech Recognition

Sibo Tong

State-of-the-art acoustic models for Automatic Speech Recognition (ASR) are based on Hidden Markov Models (HMM) and Deep Neural Networks (DNN) and often require thousands of hours of transcribed speech data during training. Therefore, building multilingual ASR systems or systems on a language with few resources is a challenging task. Multilingual training and cross-lingual adaptation are potential solutions. However, context-dependent states modeling creates difficulties for multilingual and cross-lingual ASR because of the large increase in context dependent labels arising from the phone set mismatch. The goal of this thesis is to improve current state-of-the-art acoustic modeling techniques in general for ASR, with a particular focus on multilingual ASR and cross-lingual adaptation. We systematically exploited new training frameworks, from Maximum Likelihood Estimation, Connectionist Temporal Classification to Maximum Mutual Information, in the context of phoneme-based multilingual training. In order to minimize the negative effects of data impurity arising from language mismatch, we investigated language adaptive training approaches which help further improve the multilingual ASR performance. Through comprehensive experimental comparison we demonstrated that phoneme-based multilingual models are easily extensible to unseen phonemes of new languages, from which the cross-lingual adaptation yields significant improvement over traditional approaches on limited data. Finally, we proposed a semi-supervised training approach based on dropout to boost the performance in low-resourced languages using untranscribed data. In the other part of the thesis, we conducted more theoretical analysis of techniques found to be useful in sequential multilingual training. More specifically, we revisited the recurrent architecture based on Bayesâs theorem. This leads to a Bayesian recurrent unit dictated by the probabilistic formulation and naturally support a backward recursion. Experiments show that the proposed architecture exceeds the performance of conventional recurrent network. Together, this thesis constitutes a thorough analysis of the current field. Through theoretical and experimental comparisons, the proposed approaches are shown to yield significant improvement over the conventional hybrid systems on multilingual speech recognition.

EPFL2020

Probabilistic Lexical Modeling and Grapheme-based Automatic Speech Recognition

Ramya Rasipuram

Standard hidden Markov model (HMM) based automatic speech recognition (ASR) systems use phonemes as subword units. Thus, development of ASR system for a new language or domain depends upon the availability of a phoneme lexicon in the target language. In this paper, we introduce the notion of probabilistic lexical modeling and present an ASR approach where a) first, the relationship between acoustics and phonemes is learned on available acoustic and lexical resources (not necessarily from the target language or domain), and then b) probabilistic grapheme-to-phoneme relationship is learned using the acoustic data of targeted language or domain. The resulting system is a grapheme-based ASR system. This brings in two potential advantages. First, development of lexicon for target language or domain becomes easy i.e., creation of a grapheme lexicon where each word is transcribed by its orthography. Second, the ASR system can exploit both acoustic and lexical resources of multiple languages and domains. We evaluate and show the potential of the proposed approach through a) an in-domain study, where acoustic and lexical resources of target language or domain are used to build an ASR system, b) a monolingual cross-domain study, where acoustic and lexical resources of another domain are used to build an ASR system for a new domain, and c) a multilingual cross-domain study, where acoustic and lexical resources of multiple languages are used to build multi-accent non-native speech recognition system.

Idiap2013

Towards Weakly Supervised Acoustic Subword Unit Discovery and Lexicon Development Using Hidden Markov Models

Ramya Rasipuram, Marzieh Razavi

Developing a phonetic lexicon for a language requires linguistic knowledge as well as human effort, which may not be available, particularly for under-resourced languages. An alternative to development of a phonetic lexicon is to automatically derive subword units using acoustic information and generate associated pronunciations. In the literature, this has been mostly studied from the pronunciation variation modeling perspective. In this article, we investigate automatic subword unit derivation from the under-resourced language point of view. Towards that, we present a novel hidden Markov model (HMM) formalism for automatic derivation of subword units and pronunciation generation using only transcribed speech data. In this approach, the subword units are derived from the clustered context-dependent units in a grapheme based system using the maximum-likelihood criterion. The subword unit based pronunciations are then generated either by deterministic or probabilistic learning of the relationship between the graphemes and the acoustic subword units (ASWUs). In this article, we first establish the proposed framework on a well resourced language by comparing it against related approaches in the literature and investigating the transferability of the derived subword units to other domains. We then show the scalability of the proposed approach on real under-resourced scenarios by conducting studies on Scottish Gaelic, a genuinely minority and endangered language, and comparing the approach against state-of-the-art grapheme-based approaches in under-resourced scenarios. Our experimental studies on English show that the derived subword units can not only lead to better ASR systems compared to graphemes, but can also be exploited to build out-of-domain ASR systems. The experimental studies on Scottish Gaelic show that the proposed ASWU-based lexicon development approach retains its dominance over grapheme-based lexicon. Alternately, the proposed approach yields significant gains in ASR performance, even when multilingual resources from resource-rich languages are exploited in the development of ASR systems.