pyannote community-1: Local Speaker Diarization

pyannote speaker-diarization-community-1 is the open-source speaker diarization pipeline published at pyannote/speaker-diarization-community-1. The toolkit that loads it is pyannote/pyannote-audio. Product site: pyannote.ai.

This page hangs off Otter.ai alternatives. That is an editorial grouping for local, open diarization. Otter’s commercial product is a meeting agent / transcription suite. community-1 labels speaker turns on an audio file. We do not claim feature parity, live-meeting coverage, or summary quality.

The Hugging Face card is gated (API gated: auto; README HTTP 401 without a token). Public facts below come from the HF API (likes, downloads, license, dates) and the public GitHub README.

What we verified (2026-10-13)

Field	Value	Source
Hugging Face likes	2,840	HF API
HF downloads (monthly counter)	5,444,861	HF API
Card created	2025-04-15	HF createdAt
Card last modified	2025-09-29	HF lastModified
Pipeline tag	automatic-speech-recognition	HF API (diarization tags also present)
Card license	CC-BY-4.0	HF cardData.license
GitHub stars / forks	10,638 / 1,117	GitHub API pyannote/pyannote-audio
GitHub license	MIT	GitHub API
GitHub created	2016-03-07	GitHub API
Toolkit description	“Neural building blocks for speaker diarization…”	GitHub API

The GitHub description field (API) is: “Neural building blocks for speaker diarization: speech activity detection, speaker change detection, overlapped speech detection, speaker embedding.” Homepage on that repo: https://www.pyannote.ai. Last push in the 2026-10-13 check: 2026-10-07.

How to run (from the official README)

The README section “community-1 open-source speaker diarization” lists four steps:

1. Install ffmpeg (needed by torchcodec).
2. Install with uv add pyannote.audio (recommended) or pip install pyannote.audio.
3. Accept user conditions on the gated Hugging Face card.
4. Create a Hugging Face access token.

Python sketch copied from that README (token placeholder unchanged):

import torch
from pyannote.audio import Pipeline
from pyannote.audio.pipelines.utils.hook import ProgressHook

pipeline = Pipeline.from_pretrained(
    "pyannote/speaker-diarization-community-1",
    token="HUGGINGFACE_ACCESS_TOKEN")
pipeline.to(torch.device("cuda"))
with ProgressHook() as hook:
    output = pipeline("audio.wav", hook=hook)
for turn, speaker in output.speaker_diarization:
    print(f"start={turn.start:.1f}s stop={turn.end:.1f}s speaker_{speaker}")

That snippet is documentation, not a forkalt test.

precision-2 in the same README is a premium service (Pipeline.from_pretrained("pyannote/speaker-diarization-precision-2", token="PYANNOTEAI_API_KEY")) that “runs on pyannoteAI servers”. Do not confuse it with community-1.

Benchmarks (upstream, 2025-09)

The README publishes diarization error rate (%, lower is better) last updated 2025-09. Examples copied for identification, not as a ranking against Otter (Otter is not in the table):

Set	legacy 3.1	community-1	precision-2
AISHELL-4	12.2	11.7	11.4
AMI (IHM)	18.8	17.0	12.9
VoxConverse v0.3	11.2	11.2	8.5

The README says community-1 “brings significant improvement in terms of speaker counting and assignment” versus the 3.1 legacy pipeline. Self-hosted speed on an NVIDIA H100 80GB HBM3: community-1 31s per hour of audio on AMI IHM (~1h files); precision-2 14s per hour on the same row. forkalt has not reproduced DER or speed.

Optional telemetry (PYANNOTE_METRICS_ENABLED) is documented; the README states they do not track information that could identify the user.

License

• Pipeline card: CC-BY-4.0 (HF API, 2026-10-13).
• Toolkit repo: MIT (GitHub API, 2026-10-13).

Those identifiers are different. Use the license attached to the artifact you ship. Not legal advice.

What we did not invent

• Word-error-rate or meeting-summary quality versus Otter, Descript, or Whisper-based note apps is not sourced here.
• We have not opened the gated card README.
• Hardware other than the README’s H100 speed row is not yet verified.

The rest of that 2025-09 DER table, still upstream:

Set	legacy 3.1	community-1	precision-2
AliMeeting (ch. 1)	24.5	20.3	15.2
AMI (SDM)	22.7	19.9	15.6
AVA-AVD	49.7	44.6	37.1
CALLHOME part 2	28.5	26.7	16.6
DIHARD 3 full	21.4	20.2	14.7
Ego4D (dev.)	51.2	46.8	39.0
MSDWild	25.4	22.8	17.3
RAMC	22.2	20.8	10.5
REPERE phase2	7.9	8.9	7.4

On DIHARD 3 (~5 min files) the README also lists community-1 at 37s per hour of audio versus precision-2 at 14s, on the same H100. Lower DER is better; speed is self-hosted on that GPU. Otter.ai does not appear in the table.

HF tags on the card include speaker-diarization, speaker-change-detection, voice-activity-detection, overlapped-speech-detection, and several arXiv ids (2104.03603, 2111.14448, 2012.01477, 2110.07058). Those ids match papers named beside the benchmark rows. We do not summarize the papers here.

See also Video & Audio, Otter.ai alternatives, and models.