Data Driven Approaches to Phonetic Transcription with Integration of Automatic Speech Recognition and Grapheme-to-Phoneme for Spoken Buddhist Sutra

We propose a new approach for performing phonetic transcription of text that utilizes automatic speech recognition (ASR) to help traditional grapheme-to-phoneme (G2P) techniques. This approach was applied to transcribe Chinese text into Taiwanese phonetic symbols. By augmenting the text with speech and using automatic speech recognition with a sausage searching net constructed from multiple pronunciations of text, we are able to reduce the error rate of phonetic transcription. Using a pronunciation lexicon with multiple pronunciations for each item, a transcription error rate of 12.74% was achieved. Further improvement can be achieved by adapting the pronunciation lexicon with pronunciation variation (PV) rules derived manually from corrected transcription in a speech corpus. The PV rules can be categorized into two kinds: knowledge-based and data-driven rules. By incorporating the PV rules, an error rate of 10.56% could be achieved. Although this technique was developed for Taiwanese speech, it could easily be adapted to other Chinese spoken languages or dialects.