Speaker ID

Q: Can I get names in transcripts?

Yes. Add speaker_identification to an A5S Async transcript or an A5S Streaming session with speaker labels.

Know who's speaking, by voice.

Enroll the people you expect to hear, then recognize them from a short clip. A compact model that stays accurate under loud noise and music, and puts real names on transcripts and live streams.

Request access Read the research

aircaps-speaker-id-v1REST API$0.02 per audio hourEnabled on request

Identify · one clip against every enrolled voice4.2 s clip

Clip

AGAna García Best match0.684

BIBen Ito0.208

·Unnamed speaker0.031

Similarity from −1 to 1, best first. You set the threshold.Margin 0.476

3.11%equal error rate across clean speech, noise and music 38%lower error than TitaNet Large 0.5 sof speech is enough to identify $0.02per audio hour

Voices, recognized.

Enroll once, then identify, verify or name speakers across every AirCaps model.

Enroll once

Create a speaker and add clips of them talking alone. Up to 20 enrollments per speaker, 1,000 speakers per account.

Identify or verify

Rank a clip against everyone enrolled, or check it against the one person you expect.

You set the threshold

Scores are similarities from −1 to 1. Measured score ranges in the docs help you choose a cut-off.

Names on transcripts

Add speaker_identification to an A5S Async request and speakers come back with names.

Names in live streams

In A5S Streaming sessions with speaker labels, enrolled people are named as they speak.

Built for noise

3.53% EER under environmental noise and 4.03% under music, with the noise louder than the voice.

Accurate when it is loud.

3.11%

equal error rate on VoxCeleb1-O, averaged over clean speech and four noisy conditions, from a 17M-parameter model. TitaNet Large: 5.06%.

Read the full evaluation

Where it fits.

Meetings

Notes with real names

Transcripts that say Ana and Ben instead of speaker 0 and speaker 1, across every meeting they join.

Assistants

Assistants that know who is talking

Personalize answers for the person speaking on a shared device or call.

Media

Archives by voice

Tag recurring hosts and guests across hours of recordings, without listening to them all.

Enroll, then identify.

Create speakers with an optional name and metadata
Enroll with raw audio, 1 s to 2 minutes
Identify a 0.5 s to 2 minute clip, synchronously
Pass speaker_ids to verify one person

Get an API key Speaker ID guide

# 1. create a speaker
curl https://api.aircaps.com/v1/speakers \
  -H "Authorization: Bearer $AIRCAPS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"name": "Ana García"}'

# 2. enroll a clip of them speaking alone (raw audio, 1 s to 2 min)
curl https://api.aircaps.com/v1/speakers/spk_01k73.../enrollments \
  -H "Authorization: Bearer $AIRCAPS_API_KEY" \
  -H "Content-Type: audio/wav" --data-binary @ana_intro.wav

# 3. identify who is speaking in a clip (0.5 s to 2 min)
curl https://api.aircaps.com/v1/speakers/identify \
  -H "Authorization: Bearer $AIRCAPS_API_KEY" \
  -H "Content-Type: audio/wav" --data-binary @clip.wav

{
  "object": "speaker_identification",
  "model": "aircaps-speaker-id-v1",
  "audio_duration": 4.21,
  "speakers_compared": 3,
  "margin": 0.476,
  "results": [
    {"speaker_id": "spk_01k73a9x...", "name": "Ana García", "score": 0.684},
    {"speaker_id": "spk_01k73a7m...", "name": "Ben Ito", "score": 0.208},
    {"speaker_id": "spk_01k73a5q...", "name": null, "score": 0.031}
  ]
}

Specifications.

Everything else is in the model reference and limits.

Model ID: aircaps-speaker-id-v1
Interface: REST, synchronous: the scores are the response
Clips: 0.5 s to 2 minutes to identify, from 1 s to enroll; up to 100 MB
Speakers: Up to 1,000 per account, 20 enrollments each
Output: A score from −1 to 1 for each enrolled speaker, best first
Rate: 600 requests per minute
Language: None to set: it works from the voice
Price: $0.02 per audio hour; enabled per account on request

Research note · September 2026

Speaker ID-v1: voice recognition that holds up in noise.

Design and evaluation on VoxCeleb1-O with environmental noise and music, against TitaNet, ECAPA-TDNN, ERes2NetV2 and WavLM.

Read the research note

Questions.

How do I get access?

Speaker ID is enabled per account. Talk to sales and we will turn it on.

How long does a clip need to be?

From 0.5 seconds to 2 minutes to identify, and from 1 second to enroll. Crop clips to one person's speech: the whole clip is compared as sent.

Does it decide who is unknown?

No. It returns a score for each enrolled speaker, and your application chooses the threshold. The guide has measured starting points.

Can I get names in transcripts?

Yes. Add speaker_identification to an A5S Async transcript or an A5S Streaming session with speaker labels.

Do you store identification clips?

Identification results are not stored. Enrollments stay until you delete them.

Know who's speaking.

Speaker ID is enabled per account. Tell us what you are building and we will turn it on.

Request access Start free