Skip to content

Speaker diarization support in Granite 4.1 Plus? #90

Description

@victorhooi

This is pretty awesome!

I saw there's support for the IBM Granite 4.1 models, which I've been keen to try for ages - in particular the Granite 4.1 Plus mode with timestamps and speaker diarization.

However, I saw that transcribe.cpp doesn't currently support this. It says here:

Speaker diarization | No (upstream supports via prompt; not exposed in v1 of transcribe.cpp)

When you say "upstream supports via prompt" - do you mean that the only way to enable it is to change the prompt, is that right? (Meta question - I'm curious why they do it like this - as opposed to something more deterministic, like a flag - but perhaps I'm showing my modern ML models newbieness here):

E.g. from the HF page:

SAA_PROMPT = "<|audio|> Speaker attribution: Transcribe and denote who is speaking by adding [Speaker 1]: and [Speaker 2]: tags before speaker turns."

or

TS_PROMPT = "<|audio|> Timestamps: Transcribe the speech. After each word, add a timestamp tag showing the end time in centiseconds, e.g. hello [T:45] world [T:82]"

Is that quite hard to add to transcribe.cpp though? As in, it's not a simple change, is it?

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