Inconsistent in selecting the speaker voice

[JernikGaldine]

Hey, guys I tried generating output for speaker ID 0 at a time, but I'm not getting a consistent voice—the voice seems to change across generations. Did you experience the same issue? Were you able to get the same voice every time, or did it vary for you too?

[Ssviki]

I’ve encountered the same issue—when generating output for speaker ID 0, the voice isn't consistent across generations. It tends to vary, even though I’m using the same ID. Would be great to know if there’s a workaround or if this is a known limitation.

[ZackHodari]

The model uses speaker ID to reference the context, if the speaker does not exist in the context the model will generate a new voice. If you want the model to follow a specific speaker identity you can add "audio prompts" to the context, this is like voice cloning

[ZackHodari]

The model should be pretty reliable at following the context, including speaker identity

[JernikGaldine]

But I can't get the same voice, even after fine-tuning and keeping the speaker ID as 0 in the index. I still get different voices on each attempt.

[ZackHodari]

Is the voice you want provided in the context? If you provide the target voice in the context, use a specific speaker ID (e.g. [0]) and then refer to the same speaker ID during generation it should follow the speaker identity. It might not be perfect but it should work the majority of the time

Fine-tuning should not be necessary, you can try this with the base model

[JernikGaldine]

can you please share me code for using the presistent voice using the context you are mentioning '

[ZackHodari]

You can check the readme https://github.com/SesameAILabs/csm?tab=readme-ov-file#generate-with-context