GitHub radar
A free model that works out who's speaking, and when
Pyannote released a free model that figures out who spoke when in an audio recording, without manual labeling. On public benchmarks, it is noticeably more accurate than the previous 3.1 version.
Pyannote builds open models for speech analysis and is already known to teams who produce podcasts and video. The new Community-1 model listens to a recording and marks who is speaking and when, sorting the conversation by voice before any transcription happens. Compared to the previous 3.1 version, errors dropped noticeably: on the AISHELL-4 benchmark the error rate fell from 12.2% to 11.7%, and on the AliMeeting recording it fell from 24.5% to 20.3%. The model can take the number of speakers as a hint if you already know it, and it produces results without an internet connection, running entirely on your own machine. Access to the weights is free under the CC-BY-4.0 license, but the download is gated: before downloading, you need to fill out a short form on the page and get a Hugging Face token.
Why a vibe-coder should care
For a podcast or a video with several people, this saves the most tedious part of editing - manually marking who speaks where. The gap with the previous version is not dramatic but consistent: on recordings with several participants, the model confuses voices noticeably less often. There is a small catch: the download is gated - before running it, you need to accept the terms on the model page and set up a Hugging Face token, even though the weights themselves are free.
How to install
Copy this and send it to your agent — Claude Code, Codex, any of them:
Set up Pyannote's speaker-labeling model for me: https://huggingface.co/pyannote/speaker-diarization-community-1 - get a Hugging Face token, install pyannote.audio, and run it on this recording (I'll give you the file), showing who spoke when.
Runs on a regular computer without a GPU - it processes on the CPU by default, a GPU just speeds it up. You need Python and the pyannote.audio library.
Open on Hugging Face▌ More finds