thomasgauthier

/

csm-1b-hf

like 9

Hi there,

Really great work on this, thank you.

QQ. regarding the training, I assume you can just focus on training speaker_1 in order to force the model to respond in a particular voice. I have some voice audio which I'm testing with this and other models, so the .jsonl could be formatted:

{
  "messages": [
    {
      "role": "speaker_1",
      "content": [
        {"type": "text", "text": "A voice I want to force the model to use to speak."},
        {"type": "audio", "url": "clips/example_audio_01.wav"}
      ]
    },
    {
      "role": "speaker_1",
      "content": [
        {"type": "text", "text": "Another example of the voice I want the model to use."},
        {"type": "audio", "url": "clips/example_audio_02.wav"}
      ]
    }
  ],
  "training_mask": [false, true]
}

I'm doing this as I've not had much success with voice cloning.

Thanks.