Pyannote.audio + Replicate = 💛

This is an implementation of pyannote.audio in a cog wrapper to easily run speaker diarization via replicate to save you the trouble of dependency hell 😇.

How this model works

The model takes an input audio file in the audio parameter. Then, it runs speaker diarization and returns a list of audio files containing the individual speaker turns within the audio file split by speaker and index. The output URLs contain encoded information in the file name to make working with the outputs easier. The format for the file name is {index}_{speaker}_{duration} which resolves to 0_SPEAKER_01_16. Duration is in seconds. Index refers to the order of speaker turns.

Building this model with Cog

1. SSH into a Linux environment with a GPU
2. Install Cog (using replicate/codespaces if you're using GitHub Codespaces)
3. Create a HuggingFace token and add it to predict.py as HUGGINGFACE_TOKEN (TOOD: Move it out of predict.py somehow.. maybe into a script that caches the weights)
4. Accept license aggrements for these two models on HuggingFace:

• https://huggingface.co/pyannote/speaker-diarization
• https://huggingface.co/pyannote/segmentation

1. Run cog predict -i [email protected]

Then:

1. Create a new model at replicate.com/create
2. Run cog push r8.im/your-username/your-model-name