In the past, phonetic study was mostly based on limited data. In current trends, it is expected that phonetic models have to be built on the acoustic analysis of large quantity of speech data with valid statistical analyses. The first step of most acoustic analyses unavoidably involves the alignment of recorded speech sounds with its phonetic annotation. This step is very labor-intensive and cost-ineffective since it has to be performed manually by experienced phoneticians with many hours of work. For speech engineers, this labor-intensive task can be assisted by some computer programs. However, those programs are not user-friendly to most linguists. In recent years, a tool named Speech Phonetization Alignment and Syllabification (SPPAS) has been developed by Laboratoire Parole et Language to produce ‘automatically annotations which include utterance, word, syllabic and phonemic segmentations from a recorded speech sound and its transcription.’ (Bigi 2012). In other words, this tool can automize the phonetic transcription task for speech materials, as well as the alignment task of transcription and speech recordings for further acoustic analyses. Currently, SPPAS can be implemented for French, English, Italian, Spanish, Mandarin Chinese and Taiwanese. It is distributed under a GPL license. The project aims to include Cantonese into SPPAS so that it can do automatic phonetic transcription and alignment for Cantonese speech. Upon completion, the tools will definitely facilitate the phonetic study of Cantonese.
Research plan and methodology
- Inputting a Cantonese pronunciation dictionary in the SPPAS
- This is an required step in SPPAS.
- Though there are some pronunciation dictionaries in Cantonese, most of them include the so-called standard pronunciation of a character only.
- The research team will develop a word-based dictionary with as many as variations possible for each entry based on existing Cantonese dictionaries.
2. Building an acoustic model of Cantonese in SPPAS
- The dictionary built provides the ideal articulatory and acoustic properties of the phones only. The realization of a particular phone can depend on the phonetic environment. Hence, an acoustic model of Cantonese has to be created for SPPAS.
- The effectiveness of acoustic model depends heavily on the availability of accurately annotated data and on good initialization.
- The training of the acoustic model will make use of the eight-hour speech recordings of spontaneous spoken Cantonese with detailed phonetic transcription from Hong Kong Corpus for Adult Cantonese (HKCAC) will be adopted to build the acoustic model.
- HKCAC is the only Cantonese speech corpus that provides detailed phonetic transcription capturing phone assimilation, phone dissimilation, phone deletion, syllable fusion, and tone variation.
3. Producing training materials for promoting Cantonese SPPAS
- Upon completion, the team will develop tutorials for using the tool and the description of Cantonese acoustic models in order to promote Cantonese SPPAS to the communities of Cantonese linguists.
- The promotion is also an important step in improving the performance of Cantonese SPPAS.
- When more linguists are interested in using SPPAS, more audio data will be inputted to the tool. Hence, a better acoustic model can be built.