Bark builtin speakers and voice cloning questions

[ROBERT-MCDOWELL]

I couldn't find any informations about bark builtin speakers, even speakers = tts.speakers didn't return anything,
but on the coqui-tts doc it's mentioned about speaker='ljspeech' option... any clue? thanks

[eginhard]

Bark doesn't have any default speakers. The speaker_id described in the Bark docs is just an arbitrary label that you can assign yourself. It's basically just a mechanism to precompute and store the embeddings for later reuse.

[ROBERT-MCDOWELL]

ah ok, it's more clear indeed. and about the speaker directory, is it where the .wav or embeddings should be? can we choose whatever directory?

[ROBERT-MCDOWELL]

ok. understood, but I still do not catch how coqui select the right speaker if in the directory has several wav files and speaker_id is an arbitrary label ? unless the whole directory is taken as reference mixing all voice files??

[eginhard]

For the code below you should have a folder bark_voices/myspeaker123/ that can contain an arbitrarily named file, say xy.wav. You will get an error if there is more than one file (it's probably possible to adapt that code to handle multiple reference files). After running it bark_voices/myspeaker123/xy.npz is stored for later calls.

bark = TTS("tts_models/multilingual/multi-dataset/bark").to("cuda")
bark.tts_to_file(text="Hello world", voice_dir="bark_voices/", speaker="myspeaker123")

I'll probably make this kind of mechanism more intuitive to use and document it better soon. It will be useful for other models, especially kNN-VC where you ideally have longer reference audio that you don't want to reprocess every time.

[eginhard]

Bark might also have an issue with the cloning not working at all, but I haven't looked into it: coqui-ai#3570

[ROBERT-MCDOWELL]

ok now it's clear. but now doing that won't work according to the issue coqui-ai#3570? I can take a look to the coqui-tts source too if it can help.

[ROBERT-MCDOWELL]

according to the coqui-ai#3570 I suspect the window path the issue... it shouldn't be harcoded so that's why bark ignores it.
what is questioning me is where all these random voices come from if bark has no builtin voices?