| Summary: | 碩士 === 國立臺灣大學 === 電信工程學研究所 === 102 === Most digital speech processing techniques are often designed for certain languages. Therefore, for works that contain cross-language data, information of all languages is required. Yet, when faced upon corpora of multiple languages in the globalized world today, the amount of data would be too mass and complex. Performance and processing speed would then inevitably become an issue.
However, we can look at the speech of all languages from an articulatory point of view. By using articulatory features, it is possible to represent all phones of all languages with limited numbers of features. With these features, speech can be easily handled regardless of language.
In this research, a set of articulatory features is proposed to represent phones of Mandarin and English, based on three corpora: TIMIT, COSPRO and DSP. Artificial neural networks are used to train the relationship between acoustics and articulatory features. With accuracy tested, the models may be ready for future applications.
|
|---|
