Phone-level speaker embedding based speaker adaptation method audio demo

Here we provide some audio samples generated aftering different speaker embedding methods, and explain the abbreviations below

• OracleVocode: orignal acoustic feature re-resynthesized by vocoder
• Xvec: use xvector as speaker embedding
• UttEmb: use reference audios to get utterance level speaker embedding using reference encoder
• Attentron: use reference audios to get frame level speaker embedding using attention based reference encoder
• PhnEmb(proposed): use predictor to get phon level speaker embedding
• +Adapt: also updating the LSTM in decode

test speaker1 `KHW`

test speaker1 `KHW` with adaptation

test speaker2 `HJX`

test speaker2 `HJX` with adaptation