Can you give us a code for us to reproduce the problem?

Just running with any Fairseq model normally, the same way as with XTTS (which clones just fine, version 2 included): tts.exe --use_cuda true --model_name tts_models/[lang]/fairseq/vits --text "Testing voice cloning with Fairseq on Windows." --speaker_wav Test.wav --out_path Fairseq.wav

Can you try this? https://tts.readthedocs.io/en/latest/inference.html#example-voice-cloning-by-a-single-speaker-tts-model-combining-with-the-voice-conversion-model

TTS with VC is not supported on terminal AFAIR

The thing is that running tts.exe --use_cuda true --model_name tts_models/multilingual/multi-dataset/xtts_v2 --language_idx [lang] --text "Testing voice cloning with XTTS on Windows." --speaker_wav Test.wav --out_path XTTS.wav clones the voice perfectly fine. The problem is with Fairseq models, where the argument --speaker_wav seems to be ignored, using the generic male voice.

@Poccapx
for non-voice cloning models, you need to run the resulting TTS audio through a voice conversion model. See
https://github.com/coqui-ai/TTS#voice-conversion-models

XTTS is a voice-cloning model which does this on its own. (And actually can't do without any cloning audio file).

That’s very important, thank you. Is there a list for models regarding --model_name "<language>/<dataset>/<model_name>"?

pretty sure there is currently only one official one, and that is voice_conversion_models/multilingual/vctk/freevc24

(not sure if you have to keep the first part "voice_conversion_models" out from the --model_name argument, as i am not using the CLI)

you can find a list of all models here: https://github.com/coqui-ai/TTS/blob/dev/TTS/.models.json#L924

Right! In the string tts --out_path output/path/speech.wav --model_name "<language>/<dataset>/<model_name>" --source_wav <path/to/speaker/wav> --target_wav <path/to/reference/wav> what is the difference between arguments --out_path and --target_wav?

--source_wav is the speech audio you want to convert.
--target_wav is the speech you want the source_wav to be converted into
--out_path is the finished converted audio.

Using voice_conversion_models/multilingual/vctk/freevc24 on top of a Fairseq output has worked. The voice cloning quality is nowhere near that of XTTS, but at least that way it’s possible to switch from the default male voice to a female one. For Fairseq as a non-voice cloning model, is the --speaker_wav argument always pointless or are there instances where it is used with Fairseq? Since it is present in these two examples, and that got me thinking that something was wrong with my initial setup.

sorry for the late reply.

for the first example link, its because the tts_with_vc_to_file() function does the voice conversion internally already. (thats what the "with_vc" part in the function name means)

About your second example, i have actually no idea. I would guess it has to do with the encoder model and not with the TTS model. But thats just a guess. So maybe i was wrong and you can somehow convert speakers using some vocoder models. Haven't found anything in the documentation about it, so maybe ask in the discussions https://github.com/coqui-ai/TTS/discussions about it.

I hate to make advertising, but in case you want, you can give my Application Whispering Tiger a try. It has multiple TTS plugins (including coqui TTS) and together with the RVC Plugin and a RVCv2 model, you can have probably the currently best voice conversion available. (its currently windows only though)

should be fixed by now.

你能试试这个吗？ https://tts.readthedocs.io/en/latest/inference.html#example-voice-cloning-by-a-single-speaker-tts-model-combining-with-the -语音转换模型

AFAIR 终端不支持带有 VC 的 TTS

UnboundLocalError: cannot access local variable 'dataset' where it is not associated with a value