pyannote support

kentchang · August 15, 2023, 9:00pm

Hello,

I'm on Prodigy 1.11.14 and I've been struggling to get pyannote to work. My primary source of reference is Audio and Video · Prodigy · An annotation tool for AI, Machine Learning & NLP.

I installed pyannote (2.1.1) the recommended way:

pip install -qq https://github.com/pyannote/pyannote-audio/archive/refs/heads/develop.zip

Question 1: Is pyannote.sad.manual still supported?

The example on the Audio & Video page:

prodigy pyannote.sad.manual speech_activity ./data/wav  -chunk 5

yields ✘ Can't find recipe or command 'pyannote.sad.manual'.

Following the link on that page, I was brought to this in the pyannote repo. It actually doesn't look like that manual is there in the most recent version (2.1.1). A quick inspection of version 1.1.1 of pyannote (via github) seems to suggest it does exist in the earlier version.

Question 2: Is it possible to have pyannote in the loop at all?

Even using this from their tutorial:

prodigy pyannote.audio dataset /path/to/audio/directory pyannote/speaker-segmentation

I'm seeing Could not download 'pyannote/speaker-segmentation' pipeline. I've logged in with huggingface-cli login and loading the pipeline directly in Python works, so I'm not sure why I'm seeing that message.

I tried putting the hard-coded token in the file just to test it out, which fixed that error, but now I'm seeing:

FileNotFoundError: [Errno 2] No such file or directory: '.../site-packages/pyannote/audio/interactive/common/controller.js'

Any pointers would be greatly appreciated! Thanks.

koaning · August 16, 2023, 11:29am

Hi there!

I followed this link in our docs which indeed seems to be point to an empty folder. So it seems that the model the tutorial uses isn't available anymore ... which means that I'll need to dive into the history a bit to see if we can recover it via some other method.

Do you have a link where you saw this? I just want to make sure that I have an overview of all the links from our side that point to the pyannotate side. I did find this doc on their end, but I haven't found the associated code for it just yet.

github.com

pyannote/pyannote-audio/blob/develop/tutorials/prodigy.md#recipes

# Annotating your own data with 💥 Prodigy

Manually segmenting and labeling audio data is time consuming.  For speaker diarization, depending on the required level of precision, it may take more than 10 times the duration of a recording to annotate it.

## Table of content

* [Recipes](#recipes)
* [Keyboard shortcuts](#keyboard-shortcuts)

## Recipes

`pyannote.audio` comes with a bunch of [💥 Prodigy](https://prodi.gy) recipes designed to speed things up a bit.

|   Recipe              | Usage                                                 |
|-----------------------|-------------------------------------------------------|
 🦻 `pyannote.audio`    | Annotate with a [pretrained pipeline](https://huggingface.co/models?other=pyannote-audio-pipeline) in the loop
 🧐 `pyannote.review`   | Merge multiple annotations
 🤲 `pyannote.diff`     | Show differences between two annotations
 🗄 `pyannote.database` | Dump annotations as [`pyannote.database`](https://github.com/pyannote/pyannote-database/) protocols

This file has been truncated. show original

I'll respond more in a bit after I've done some digging!

koaning · August 16, 2023, 11:39am

I'll discuss this internally, it does look like the Prodigy features may have been deprecated on their end. Per the information in this PR:

In particular, this commit seems to remove the Prodigy recipes