Automatic phonetic transcription of large speech corpora

Abstract

Most large speech corpora are delivered with a lexicon that contains a canonical transcription of every word in the orthographic transcription. Such a lexicon can be used for generating a hypothetical 'canonical' phonetic transcription from the orthography. In addition, time and money permitting, some speech corpora are provided with a manually verified broad phonetic transcription of at least part of the material. Since the manual verification of phonetic transcriptions is time-consuming and expensive, we investigated whether existing automatic transcription procedures and combinations of such procedures can offer a quick and cheap alternative for the generation of phonetic transcriptions like the manually verified transcriptions delivered with large speech corpora. In our study, we used ten automatic transcription procedures to generate a broad phonetic transcription of well-prepared speech (readaloud texts) and spontaneous speech (telephone dialogues) from the Spoken Dutch Corpus. The performance was assessed in terms of the number and the nature of the discrepancies between the emerging phonetic transcriptions and the corresponding manually verified phonetic transcriptions delivered with the Spoken Dutch Corpus. The resulting automatic transcriptions appeared to be comparable to the manually verified transcriptions.

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

“…We used a lexicon with many pronunciation variants for each word, which we generated by means of rules applied to the canonical pronunications. Contrary to Van Bael et al (2007b) and Cucchiarini and Binnenpoorte (2002), whose rules were insensitive to the stress pattern and syllable structure of the word, our rules are sensitive to this information. As a result, we obtained a larger number of probable variants.…”

“…The models were trained at a frame shift of 5 ms and a window length of 25 ms, where for each frame 13 MFCCs (i.e., the mel-scaled cepstral coefficients C0-C12) and their first and second order derivatives (39 features) were calculated. We used a shorter frame shift than the default of 10 ms used in earlier studies of segmental reductions (e.g., Van Bael et al, 2007b;Adda-Decker et al, 2005;Schuppler et al, 2009) tin order to achieve more accurate positions of the segment boundaries and in order to be able to identify very short segments. With a frame shift of 5ms and acoustic models consisting of three emitting states (no skips), segments will be assigned a minimum length of 15ms.…”

“…Only the columns 'Type' and 'Order' are relevant for this section, the other columns will be discussed in Section 4. Some of the rules are well-studied for Dutch and have been used before in the automatic generation of phonemic transcriptions (Van Bael et al, 2007b;Kessens et al, 2003); these are: 'schwa-insertion' (1.0), '[n]-deletion after schwa' (1.1), 'regressive assimilation of voice for obstruents before voiced plosives' (1.2), 'devoicing of plosives following voiceless plosives' (1.3), 'devoicing of fricatives in all word-positions' (1.4), '[t]-deletion in word-final position, preceded by consonant' (4.8) and '[r]-deletion after schwa' (4.5). The other rules were formulated on the basis of the research by Ernestus (2000) on voice assimilation and segment reduction in casual Dutch.…”

“…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.…”

“…A more recently available method is to create phonetic transcriptions by using an automatic speech recognition (ASR) system to determine the most likely pronunciation variant for each word in a spoken corpus (e.g., Binnenpoorte, 2006;Cucchiarini and Binnenpoorte, 2002;Van Bael et al, 2007b). With this method large amounts of speech material can be transcribed in a relatively short period of time.…”