A Cost-Benefit Analysis of Hybrid Phone-Manner Representations for ASR

Proceedings of Human Language Technology Conference and Conference on Empirical Methods in Natural Language Processing (HLT/EMNLP), pages 571–578, Vancouver, October 2005. c⃝2005 Association for Computational Linguistics A Cost-Benefit Analysis of Hybrid Phone-Manner Representations for ASR Eric Fosler-Lussier Department of Computer Science and Engineering Department of Linguistics Ohio State University Columbus, OH 43210 fosler@cse.ohio-state.edu C. Anton Rytting Department of Linguistics Ohio State University Columbus, OH 43210 rytting@ling.ohio-state.edu Abstract In the past decade, several researchers have started reinvestigating the use of sub-phonetic models for lexical represen- tations within automatic speech recogni- tion systems. Lest history repeat itself, it may be instructive to mine the further past for models of lexical representations in the lexical access literature. In this work, we re-evaluate the model of Briscoe (1989), in which a hybrid strategy of lex- ical representation between phones and manner classes is promoted. While many of Briscoe’s assumptions do not match up with current ASR processing models, we show that his conclusions are essentially correct, and that reconsidering this struc- ture for ASR lexica is an appropriate av- enue for future ASR research. 1 Introduction Almost every state-of-the-art large vocabulary au- tomatic speech recognition (ASR) system requires the sharing of sub-word units in order to achieve the desired vocabulary coverage. Traditionally, these sub-word units are determined by the phones or phonemes of a language (depending on desired detail of representation). However, phonetic (or phonemic) representation has its pitfalls (cf. (Os- tendorf, 1999)). Among the problems cited in the literature are that (1) segments are often dif- ficult for machines to recognize from the acoustic cues alone, because the acoustic cues to a particu- lar phoneme are multi-faceted, and (2) the intended words and phrases are not always recoverable even from correctly recognized segments, because speak- ers themselves will also fail to articulate words with the dictionary-listed phonemes. The first of these problems refers to the discriminability of phonemes within an inventory; the second to the reliability of (actual) phone sequences mapping to the canonical phonemic representations of words. This is partic- ularly true in conversational speech (such as that found in the Switchboard corpus), where pragmatic context and conversational conventions assist human comprehension (but not current ASR systems). A common approach for handling pronunciation variation is to introduce alternative entries into the lexicon. However, phones that are perceived as non- canonical (for example, when an /eh/ is heard as an /ih/ by linguistic transcribers) often are closer in acoustic space to the Gaussian means of the canonical phones, rather than the perceived phones (Sarac¸lar et al., 2000). This insight suggests that acoustic models need to be cognizant of potential pronunciation changes. Thus the lexical and acous- tic models should work hand in hand. Another way to model this type of pronunciation variation is to find the commonalities that the canon- ical and perceived phone share in terms of a sub- phonetic representation. In the past decade, a signif- icant community in acoustic-phonetic ASR research has been turning to distinctive features (Jakobson et al., …

Content truncated to 500 words