How is the speaker-independency ? about lyra HOT 3 OPEN

The codec has been trained on thousands of hours of speech in more than 70 languages and tested on many different speakers and accents. It relies only on log-mel spectrograms per frame and no long-term speaker id. For some audio samples you can check out our blog or generate them yourself using encoder_main and decoder_main.

from lyra.

The codec has been trained on thousands of hours of speech in more than 70 languages and tested on many different speakers and accents. It relies only on log-mel spectrograms per frame and no long-term speaker id. For some audio samples you can check out our blog or generate them yourself using encoder_main and decoder_main.

Did the training data including 70 languages as you said contain Chinese speech data? If it did, what is the proportion of Chinese speech data?

from lyra.

Our training data definitely contained Chinese speech data. But since we combined several different datasets it is difficult to estimate the precise proportion. Please see reply in #32.

from lyra.