Skip to content

Question about prompt settings for evaluations #1884

@jzhe727

Description

@jzhe727

For CosyVoice 3 posted metrics, what is the prompting setup for them? Is it following the exact setup as in example.py ('You are a helpful assistant.<|endofprompt|>" + target text)? Also, does the text normalization method (WeText vs TTSFRD) matter?

I ask because the results I am getting are slightly different from published for CosyVoice3 0.5B no RL, (seedtts-zh bolded are mine: 0.780 -> 0.7786 SIM and 1.16 -> 1.341 WER).

If I am prompting Cosyvoice correctly I might need to check my evaluation pipeline. (For example, if Paraformer transcribes a different character with exact same pronunciation, should that count towards WER?)

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type
    No fields configured for issues without a type.

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions