Skip to content

transcription gets stuck repeating the same words #616

@dotmobo

Description

@dotmobo

Sometimes, the generated transcription gets stuck repeating the same words, and there is a warning :

Compression ratio threshold is not met with temperature 0.0

The problem is that the default value for temperature is 0 in speaches. The problem can be solved if we pass a higher temperature but it will apply to all the transcript.

The default behavior in faster-whisper is that if temperature is not specified, it is considered 0 and only when compression ratio threshold or log probability threshold are not met for a certain segment then it will try with temperature 0.2 and so on from the list [0, 0.2, 0.4, 0.6, 0.8, 1.0] for that segment.

See fix in #615 (original fix in #553) : Change the default value for temperature to the same default list used in faster-whisper to let it handle retries for segments where compression ratio threshold or log probability threshold criterias are not met.

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions