Modelling Pronunciation Variation using Multi-Path HMMS for Syllables

Abstract

The following full text is a publisher's version.For additional information about this publication click this link. http://hdl.handle.net/2066/44459Please be advised that this information was generated on 2024-06-02 and may be subject to change. Article 25fa End User AgreementThis publication is distributed under the terms of Article 25fa of the Dutch Copyright Act. This article entitles the maker of a short scientific work funded either wholly or partially by Dutch public funds to make that work publicly available for no consideration following a reasonable period of time after the work was first published, provided that clear reference is made to the source of the first publication of the work.Research outputs of researchers employed by Dutch Universities that comply with the legal requirements of Article 25fa of the Dutch Copyright Act, are distributed online and free of cost or other barriers in institutional repositories. Research outputs are distributed six months after their first online publication in the original published version and with proper attribution to the source of the original publication.

Acoustic reduction in conversational Dutch: A quantitative analysis based on automatically generated segmental transcriptions

“…In general, pronunciation variants can either be extracted from a large corpus that has already been transcribed manually at the segmental level (data-driven approach, e.g., Hämäläinen et al, 2007;Kessens et al, 2003) or they can be generated by applying a set of rules, proposed in the phonological/phonetic literature, to the canonical forms in the lexicon (knowledge-based approach, e.g., Van Bael et al, 2007b). The set of variants derived with a data-driven approach depends on the corpus from which the variants are extracted and tends to contain fewer pronunciation variants for most words than a lexicon created with the knowledgebased approach.…”

Dual stream speech recognition using articulatory syllable models

“…Previous research into speech recognition using syllables has been mostly based on replacing or combining the triphone-HMMs with models of the same type for syllables (Hetjmánek and Pavelka 2008;Hämäläinen et al 2007;Han et al 2006;Sethy et al 2003;Ogata and Ariki 2003;Ganapathiraju et al 2001;Ahadi 2000;Wu et al 1998;Dupont and Bourlard 1997). Where large vocabulary tasks have been attempted, data sparsity has been fixed by backoff to phone sequences (Sethy et al 2003;Ganapathiraju et al 2001).…”

Modelling pronunciation variation with single-path and multi-path syllable models: Issues to consider

“…In fact, some of the retrained parallel paths were still closely related to the MDVs used to initialise them. In Hämäläinen et al (2007b), we carried out a forced alignment of the training data with the multi-path mixed-model recogniser and analysed the training tokens assigned to each path of the syllable models. Our analysis showed that the token-to-path assignment was clearly related to the articulatory similarity -or dissimilarity -between the transcriptions of the training tokens and the MDVs used to initialise the parallel paths.…”

“…Three parallel paths per syllable appeared a good compromise between too little training data and too small a distance between the triphone sequences used to initialise the paths. Our assumption about the optimal number of paths could later be verified by carrying out a forced alignment of the training data with the syllable models; the majority of the paths were frequently entered (Hämäläinen et al, 2007b). In addition, removing the paths that were rarely used during the forced alignment showed that the recognition performance remained virtually unchanged .…”

“…As the retrained parallel paths of the multi-path syllable models are still closely related to the MDVs used to initialise them (Hämäläinen et al, 2007b;Hämäläinen et al, 2007c), one might argue that we could alleviate the problem of lexical confusability by refining our MDV selection approach. Based on discriminative training methods (Lin and Yvon, 2007;Markov and Nakamura, 2007) or existing methods to detect confusable words (Anguita et al, 2005;Roe and Riley, 1994), we could devise ways of avoiding MDVs that would result in overlapping pronunciations with other words in the lexicon.…”