Joint Learning of Phonetic Units and Word Pronunciations for ASR

Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing, pages 182–192, Seattle, Washington, USA, 18-21 October 2013. c⃝2013 Association for Computational Linguistics Joint Learning of Phonetic Units and Word Pronunciations for ASR Chia-ying Lee, Yu Zhang, James Glass Computer Science and Artificial Intelligence Laboratory Massachusetts Institute of Technology Cambridge, MA 02139, USA {chiaying,yzhang87,jrg}@csail.mit.edu Abstract The creation of a pronunciation lexicon re- mains the most inefficient process in develop- ing an Automatic Speech Recognizer (ASR). In this paper, we propose an unsupervised alternative – requiring no language-specific knowledge – to the conventional manual ap- proach for creating pronunciation dictionar- ies. We present a hierarchical Bayesian model, which jointly discovers the phonetic inven- tory and the Letter-to-Sound (L2S) mapping rules in a language using only transcribed data. When tested on a corpus of spontaneous queries, the results demonstrate the superior- ity of the proposed joint learning scheme over its sequential counterpart, in which the la- tent phonetic inventory and L2S mappings are learned separately. Furthermore, the recogniz- ers built with the automatically induced lexi- con consistently outperform grapheme-based recognizers and even approach the perfor- mance of recognition systems trained using conventional supervised procedures. 1 Introduction Modern automatic speech recognizers require a few essential ingredients such as a signal representation of the speech signal, a search component, and typ- ically a set of stochastic models that capture 1) the acoustic realizations of the basic sounds of a lan- guage, for example, phonemes, 2) the realization of words in terms of these sounds, and 3) how words are combined in spoken language. When creating a speech recognizer for a new language the usual requirements are: first, a large speech corpus with word-level annotations; second, a pronunciation dic- tionary that essentially defines a phonetic inventory for the language as well as word-level pronuncia- tions, and third, optional additional text data that can be used to train the language model. Given these data and some decision about the signal rep- resentation, e.g., centi-second Mel-Frequency Cep- stral Coefficients (MFCCs) (Davis and Mermelstein, 1980) with various derivatives, as well as the nature of the acoustic and language model such as 3-state HMMs and n-grams, iterative training methods can be used to effectively learn the model parameters for the acoustic and language models. Although the de- tails of the components have changed through the years, this basic ASR formulation was well estab- lished by the late 1980’s, and has not really changed much since then. One of the interesting aspects of this formulation is the inherent dependence on the dictionary, which defines both the phonetic inventory of a language, and the pronunciations of all the words in the vo- cabulary. The dictionary is arguably the cornerstone of a speech recognizer as it provides the essential transduction from sounds to words. Unfortunately, the dependency on this resource is a significant im- pediment to the creation of speech recognizers for new languages, since they are typically created by experts, whereas annotated corpora can be relatively more easily created …

Content truncated to 500 words