Pronunciation Variants and ASR of Colloquial Speech: A Case Study on Czech

Pronunciation Variants and ASR of Colloquial Speech: A Case Study on Czech David Lukeš, Marie Kopˇrivová, Zuzana Komrsková, Petra Poukarová Institute of the Czech National Corpus, Charles University, Faculty of Arts, Czech Republic {david.lukes, marie.koprivova, zuzana.komrskova, petra.poukarova}@ff.cuni.cz Abstract A standard ASR system is built using three types of mutually related language resources: apart from speech recordings and orthographic transcripts, a pronunciation component maps tokens in the transcripts to their phonetic representations. Its implementation is either lexicon-based (whether by way of simple lookup or of a stochastic grapheme-to-phoneme converter trained on the source lexicon) or rule-based, or a hybrid thereof. Whichever approach ends up being taken (as determined primarily by the writing system of the language in question), little attention is usually paid to pronunciation variants stemming from connected speech processes, hypoarticulation, and other phenomena typical for colloquial speech, mostly because the resource is seldom directly empirically derived. This paper presents a case study on the automatic recognition of colloquial Czech, using a pronunciation dictionary extracted from the ORTOFON corpus of informal spontaneous Czech, which is manually phonetically transcribed. The performance of the dictionary is compared to a standard rule-based pronunciation component, as evaluated against a subset of the ORTOFON corpus (mul- tiple speakers recorded on a single compact device) and the Vystadial telephone speech corpus, for which prior benchmarks are available. Keywords: speech recognition, pronunciation variants, spontaneous speech 1. Introduction One of the components of an automatic speech recognition (ASR) system is a pronunciation dictionary, which provides a mapping between a conventional symbolic transcript of speech, which can exhibit varying degrees of arbitrariness, and an acoustically / phonetically motivated one. On the one hand, during the training phase (phone-level alignment, creation of acoustic models, henceforth AMs), the phonetic transcription provides information about where to expect recurring patterns in the acoustic signal, which will be ab- stracted away and generalized in the acoustic models for the individual (tri)phones. On the other hand, during decoding, it mediates between (to borrow a neurolinguistic analogy) the top-down (predictive) processes based on the linguistic knowledge (the language models, henceforth LMs), and the bottom-up (recognitive) processes based on spotting acous- tic patterns in the incoming sound data. The complexity of the mapping between the standardized orthographic representation of a word and an encoding of its pronunciation as a string of phones depends on the char- acteristics of the writing system of the language in ques- tion. The possibilities range from more or less complete arbitrariness to varying degrees of systematicity in the re- lationship between graphemes and phonemes (see Samp- son (2015) for a book-length account). The variety present in this respect in the world’s languages is best attested by sketching a few illustrative points along this continuum: • Chinese logographic writing bears little to no relation to the phonetic form of words; phonetically speak- ing, two Mandarin Chinese words such as mother and hemp may have the same segmental content ([ma]) and differ “only” in their suprasegmental features (1st vs. 2nd tone), yet their traditional …

Content truncated to 500 words