Speaker Diarization

OpenASR can label transcript segments with per-speaker identities. Diarization is a model-agnostic stage that wraps around any ASR executor — it composes with every native model family additively, using neural VAD plus speaker segmentation, embedding, and clustering.

All processing is local. No network calls are made beyond a one-time, consented capability-pack install. Only L2-normalized embeddings are stored — raw enrollment audio is never written to the voiceprint store. Voice matching is a diarization convenience, not authentication.

Enabling diarization

Pass --diarize to transcribe or live:

openasr transcribe --diarize recording.wav
openasr live --diarize

Segments are labeled with anonymous speakers in zero-padded format: SPEAKER_00, SPEAKER_01, and so on.

If the required WeSpeaker speaker-embedder capability pack is not already installed, passing --diarize triggers an automatic install — the flag itself is the consent. The install reuses the same progress UI as openasr pull. If --diarize is passed on a backend or pack that does not support diarization, the CLI fails closed with an error rather than silently skipping it.

Forcing a speaker count

For offline transcription you can force an exact cluster count:

openasr transcribe --diarize --speakers 3 recording.wav

--speakers accepts a positive integer and requires --diarize. It is available on transcribe only — live does not support a forced speaker count.

Voice-match enrollment

Enroll a voice-match profile so diarized output uses a display name instead of an anonymous label:

openasr speaker enroll enroll.wav --name "Alice" --match-similarity 0.6

Option	Description	Default
`<input>`	16 kHz mono PCM16 WAV with at least five seconds of speech.	(required)
`--name`	Display name used when this voice matches.	`SPEAKER_ME`
`--match-similarity`	Cosine similarity threshold (0–1) required for a match.	`0.5`

If your audio is not in the right format, convert it with ffmpeg:

ffmpeg -i input.mp3 -ac 1 -ar 16000 -c:a pcm_s16le enroll.wav

On success the CLI prints the profile name, ID, input path, and storage location. The profile takes effect on the next diarized session.

Clearing profiles

Remove all local voice-match profiles:

openasr speaker clear

This deletes the on-disk voiceprint store entirely.

Storage location

Profiles are stored at OPENASR_HOME/diarize/voiceprints.json by default. Override the path with the OPENASR_SPEAKER_PROFILES environment variable.

Server API

The server exposes a REST surface for managing voice-match profiles. All /v1/speakers routes are operator-only — they require the pairing admin token when pairing auth is enabled.

List profiles

curl -s http://127.0.0.1:8080/v1/speakers

Each profile in the response includes:

Field	Type	Description
`id`	string	Profile identifier.
`name`	string	Display name.
`created_at`	string	Creation timestamp.
`sample_seconds`	number	Duration of the enrollment audio.
`compatible`	bool	Whether the stored embedding matches the currently active speaker-embedder.

Enroll a profile

curl -s http://127.0.0.1:8080/v1/speakers \
  -F name=Alice \
  -F [email protected]

Rename a profile

curl -s -X PATCH http://127.0.0.1:8080/v1/speakers/{id} \
  -H "Content-Type: application/json" \
  -d '{"name": "Bob"}'

Re-enroll a profile

Replace a profile’s stored embedding from a new recording without changing its ID:

curl -s -X POST http://127.0.0.1:8080/v1/speakers/{id}/reenroll \
  -F wav=@new_sample.wav

Delete a profile

curl -s -X DELETE http://127.0.0.1:8080/v1/speakers/{id}

Returns { "id": "...", "deleted": true }.

Privacy model

Embeddings only — raw enrollment audio is never persisted. The voiceprint store contains L2-normalized embeddings and the active speaker-embedder identity.
Local processing — diarization runs entirely on-device. The only network activity is the one-time capability-pack download, which only happens when you pass --diarize and the pack is missing.
Not authentication — voice matching assigns display names during diarization. It is not an identity-verification mechanism.
No telemetry.

See the CLI reference for the full flag list and Server API for the complete endpoint table.