The hard part is no longer “what was said.” It is who said it, and when

Convert messy, multi-speaker audio and see speaker timelines, segments, and speaker-attributed transcripts.

Contact Sales

No credit card required

precision-3

Choose a sample audio

Or upload your audio

Choose file

MP3, WAV, M4A, MP4, or OGG · Up to 250 MB and 1 hour.

See when each speaker talks

Choose a sample or file, then press Run to see speaker turns.

Read docs Open Playground

Speech AI
AI Editing
We switched from AWS Transcribe for the accuracy, and tested it against three other platforms. Pyannote won. Now it's pure set-it-and-forget-it.
Voice AI
AI Meeting
AI Video
AI Transcription
Dubbing only works if every line lands on the right voice. Pyannote keeps speaker attribution reliable across languages, at production volume.
Call Analytics
Medical scribe
Voice QA
AI Transcription
I've seen the system work flawlessly at high data volume. Very high reliability, high confidence in the output.
Dubbing
AI Meeting

Teams choose us to power Voice AI tools at scale

Explore Customer Story

Risk of incorrect evaluations would make the value of our product go to zero if misassigned. If speaker attribution is wrong, the value of our evaluations goes to zero.

Risk of incorrect evaluations would make the value of our product go to zero if misassigned. If speaker attribution is wrong, the value of our evaluations goes to zero.

Risk of incorrect evaluations would make the value of our product go to zero if misassigned. If speaker attribution is wrong, the value of our evaluations goes to zero.

Speaker diarization is critical to the Notta experience because so many downstream features depend on knowing who said what.

Speaker diarization is critical to the Notta experience because so many downstream features depend on knowing who said what.

Speaker diarization is critical to the Notta experience because so many downstream features depend on knowing who said what.

There might be a hundred-millisecond overlap where I'm ending a sentence and you start speaking. In a fine-tuning dataset, even that small a chunk would create a problem.

There might be a hundred-millisecond overlap where I'm ending a sentence and you start speaking. In a fine-tuning dataset, even that small a chunk would create a problem.

There might be a hundred-millisecond overlap where I'm ending a sentence and you start speaking. In a fine-tuning dataset, even that small a chunk would create a problem.

We’re focused on really formal meetings: town halls, government sessions. It’s important in this kind of meeting minutes that you can understand who said what.

We’re focused on really formal meetings: town halls, government sessions. It’s important in this kind of meeting minutes that you can understand who said what.

Dubbing only works if every line lands on the right voice. Pyannote keeps speaker attribution reliable across languages, at production volume.

Our models

Research-backed, production-ready AI models

pyannoteAI is built on 12 years of speech processing and speaker diarization research. Our models have been open source since 2016, and we are committed to the open-source community.

1  
2  
3  
4  
5  
6
7
8
9
10
11
12

# download pretrained pipeline from Huggingface
from pyannote.audio import Pipeline
pipeline = Pipeline.from_pretrained('pyannote/speaker-diarization-community-1', token="HUGGINGFACE_TOKEN")

# perform speaker diarization locally
output = pipeline('/path/to/audio.wav')

# enjoy state-of-the-art speaker diarization
for turn, speaker in output.speaker_diarization:
print(f"{speaker} speaks between t={turn.start}s and t={turn.end}s")

Open source roots