Chinese support

[MonolithFoundation]

would consider support Chinese?

[cantabile-kwok]

That is a good question. We want to see how this model works on Chinese, but the core problem is not about model or dataset; it is about the speech tokens. Since in the paper we use vq-wav2vec, which is only trained on English Librispeech corpus, we don't expect it to generalize very well to Chinese. We need to find another token which contains limited timbre information and enough prosody information for Chinese, which seems a bit hard. Training a vq-wav2vec on Chinese dataset is also a larger project. Hence, we would not train this on Chinese unless there is a satisfactory speech token ready to use.

Nevertheless, the language restriction is only on the source speech. For the target reference, any language is feasible (i.e. no problem from English content to Chinese speaker).