세미나

In this talk, we will raise a challenging question: Can we find the “elementary particles” of a person’s speech in one language and use them for rendering his/her voice in a different language? A positive “yes” answer and the found “elementary particles” can then have many useful applications, e.g. mixed code text-to-speech (TTS), language learning, speech-to-speech translation, etc. We try to answer the question by limiting ourselves first to how to train a TTS of a target language with speech corpus collected from a speaker in his own mother tongue, which is different from the target language, along with a speech corpus recorded by a reference speaker (in the same target language). We then use “trajectory tiling algorithm,” invented by us for synthesizing high quality, unit selection TTS, to “tile” the trajectories of all sentences in the reference speaker’s corpus with the most appropriate speech segments in the monolingual speaker’s data. To make the tiling proper across two different (reference and source) speakers, the difference between their speech signals needs to be equalized with appropriate vocal tract length normalization, e.g., a bilinear warping function or formant mapping. All tiled sentences are then used to train a new HMM-based TTS of the monolingual speaker but in the reference speaker’s language. Different length units of the ‘elementary particles” have been tried and a label-less frame length (10 ms) segments have been found to yield the best TTS quality. Some preliminary results also show that training a speech recognizer with speech data of different languages tends to improve the ASR performance in each individual language. Also, in addition to the fact that audio “elementary particles” of human speech in different languages can be discovered as frame-level speech segments, the mouth shapes of a mono-lingual speaker have also been found adequate for rendering the lips movement of talking heads in different languages. Various demos will be shown to illustrate our findings.