Correct forced decoder ids
#38
by
patrickvonplaten
- opened
Any reason why the generation config doesn't match the config here: https://huggingface.co/openai/whisper-tiny/blob/be0ba7c2f24f0127b27863a23a08002af4c2c279/config.json#L26
Config is wrong in this case - the language at index 1 shouldn't be specified in the forced decoder ids. It should be inferred automatically from the audio.
This inconsistency between config and generation config is observed across all Whisper models - I'll open Hub PRs to correct them!