No credit card required
Built by the team behind the most-used diarization model: pyannote.audio - 1B+ download.
In production in 500+ enterprises - Synthesia, Descript, SmallestAI, Notta, Gladia, UpMeet...
Diarization output
Live-1 · 300 ms
SPEAKER_00
SPEAKER_01
SPEAKER_02
{ "speaker": "SPEAKER_01", "start": 7.412, "end": 11.089, "confidence": 0.97}
Overlap marked at 00:14 — two speakers active in the same instant.
Twelve years of diarization research
Precision-3: highest accuracy, batch or streaming
The commercial model, for teams shipping diarization to end users.
Batch mode
Post-call analytics, compliance archives, and any workload where you hold the full recording
Streaming mode
Live-1, so speaker labels arrive as the audio does at 300 ms latency
Voiceprints and speaker identification
Confidence scores at turn and sample level, input parameters, and output scores to tune detection and interpret results
We respect your privacy by design
No customer data has ever been used to train our models.
Available as a Cloud API, and On-Premise
On-devices, or air-gapped when the audio cannot leave your infrastructure
Community-1: Open-Source, cost-efficient at volume
The open-source model, self-hosted and free.
Batch diarization
Running on your own hardware
Fine-tune on your own data
The only version you can fine-tune on your own annotated data
pyannote.audio
Backed by the pyannote.audio ecosystem and its community
In production today
Meeting intelligence
Readable, speaker-attributed transcripts that hold up in overlapping or noisy audio, so summaries and action items land on the right person.
Financial services compliance
Timestamped, speaker-attributed records that support audits, risk review, and dispute resolution.
Healthcare AI scribing
Clean separation of clinician, patient, and third parties, so clinical notes are attributed correctly and need less manual repair.
Voice agent evaluation
Turn-taking, latency and interruption patterns measured from the audio itself, to tell where an agent talks over its user.
Media and podcast production
Voice-aligned segments for editing, subtitling, dubbing and multilingual delivery.
Training data for speech models
Segmented and speaker-labelled audio for training, fine-tuning and evaluation.
1
Create an API key from the dashboard.
2
Send your first request with the audio you want to process.
3
Read back a list of segments: speaker label, start, end, confidence.
Language-agnostic
Precise timestamps
Batch, streaming, on-device
Sovereign
curl
Every plan includes audio hours. Developer covers up to 170 hours per month, Starter up to 1,030, Enterprise is priced on volume.
It partitions audio by speaker. You send a file or a stream, and it returns who spoke, when they started, when they stopped, and how confident the model is. It estimates the speaker count on its own, which is what separates diarization from channel splitting.
Diarization does not transcribe: it answers "who spoke when", and speech-to-text answers "what was said". Run both, and you get a speaker-attributed transcript. Diarization also does not name people. It returns anonymous labels such as SPEAKER_00. Attaching a real identity is speaker identification, and it needs an enrolled voiceprint.
01
Voice activity detection separates speech from silence, music, and noise.
02
Segmentation cuts speech where the active speaker changes. Overlap handling happens here: end-to-end models can mark two speakers as active at the same instant.
03
Speaker embeddings project each segment into a space where distance reflects voice similarity, not words or language.
04
Clustering groups those embeddings. Each cluster is one speaker, and the cluster count is the estimated speaker count.
The standard metric is diarization error rate (DER): the share of speech time attributed to the wrong speaker. It sums false alarms (non-speech labelled as speech), missed speech (usually the quieter voice in an overlap), and speaker confusion (right speech, wrong label). Lower is better, and it is time-weighted, so one mislabelled monologue costs more than several mislabelled interjections.
Published numbers are a screening tool, not a verdict. Benchmark datasets are not your audio. Run your own recordings through the model before you commit.
How many speakers can diarization detect?
Yes, absolutely! You can sign up for a free trial and enjoy 150 hours of identification and 10 voiceprints for one month. You will have access to our latest premium models to ensure accuracy and precision for your speech processing needs.
Does diarization include transcription?
Yes, absolutely! You can sign up for a free trial and enjoy 150 hours of identification and 10 voiceprints for one month. You will have access to our latest premium models to ensure accuracy and precision for your speech processing needs.
Can diarization run in real time?
Yes, absolutely! You can sign up for a free trial and enjoy 150 hours of identification and 10 voiceprints for one month. You will have access to our latest premium models to ensure accuracy and precision for your speech processing needs.
My STT provider already includes diarization. Why add pyannote?
Yes, absolutely! You can sign up for a free trial and enjoy 150 hours of identification and 10 voiceprints for one month. You will have access to our latest premium models to ensure accuracy and precision for your speech processing needs.
Does diarization work in my language?
Yes, absolutely! You can sign up for a free trial and enjoy 150 hours of identification and 10 voiceprints for one month. You will have access to our latest premium models to ensure accuracy and precision for your speech processing needs.