Great Work Labs
Models / Audio

Who

Speaker diarization
who-1

Who spoke when, from a recording. No transcript needed.

Price$0.0003 per minute
Speed (our servers)956 to 966 ms
LanguagesAny (works on voices, not words)
InputAudio, up to 10 min
Built onpyannote/segmentation-3.0 + WeSpeaker ResNet34 (MIT)

Who splits a recording into turns and labels each one with a speaker (SPEAKER_1, SPEAKER_2...), with start and end times and each person's total talk time. It recognizes the same voice across the whole file, so a speaker who comes back later keeps their label.

Set num_speakers when you know how many people are talking. For a transcript with speakers, send speakers: true to Scribe instead: it runs Who for you.

Use it

curl https://labs.greatwork.company/v1/run \
  -H "Authorization: Bearer $GWL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "who-1",
  "input": "https://labs.greatwork.company/examples/two-voices.m4a"
}'

Set GWL_API_KEY to a key from your dashboard. The bare name who works too and follows new versions.

Options

input: An audio or video file: an https URL, a data URL, or a file in a multipart upload, up to 25 MB and 10 minutes.

NameTypeDefaultWhat it does
num_speakersintegerExact number of speakers, if you know it.
max_speakersinteger8Upper limit when the number isn't known.
thresholdnumber0.5How similar two voices must be to count as one speaker. Raise it if two people get merged, lower it if one person gets split.

Output

FieldTypeMeaning
speakersintegerHow many speakers were found.
segmentsobject[]Turns in order: speaker, start, end (seconds).
speaking_timeobjectSeconds of speech per speaker.
durationnumberAudio length in seconds.

Every response wraps this in {"id", "model", "output", "usage": {"units", "unit", "cost_usd"}, "latency_ms"}.

Examples

Two people

{
  "model": "who-1",
  "input": "/examples/two-voices.m4a"
}
{
  "speakers": 2,
  "segments": [
    {
      "speaker": "SPEAKER_1",
      "start": 0,
      "end": 3.89
    },
    {
      "speaker": "SPEAKER_2",
      "start": 4.04,
      "end": 8.91
    },
    {
      "speaker": "SPEAKER_1",
      "start": 9,
      "end": 12.55
    },
    {
      "speaker": "SPEAKER_2",
      "start": 12.69,
      "end": 17.23
    }
  ],
  "speaking_time": {
    "SPEAKER_1": 7.44,
    "SPEAKER_2": 9.41
  },
  "duration": 17.41
}

Captured on our production servers: 956 ms, 18 seconds, $0.00009.

Known count

{
  "model": "who-1",
  "input": "/examples/two-voices.m4a",
  "num_speakers": 2
}
{
  "speakers": 2,
  "segments": [
    {
      "speaker": "SPEAKER_1",
      "start": 0,
      "end": 3.89
    },
    {
      "speaker": "SPEAKER_2",
      "start": 4.04,
      "end": 8.91
    },
    {
      "speaker": "SPEAKER_1",
      "start": 9,
      "end": 12.55
    },
    {
      "speaker": "SPEAKER_2",
      "start": 12.69,
      "end": 17.23
    }
  ],
  "speaking_time": {
    "SPEAKER_1": 7.44,
    "SPEAKER_2": 9.41
  },
  "duration": 17.41
}

Captured on our production servers: 966 ms, 18 seconds, $0.00009.

Good for

  • Split meeting recordings by speaker
  • Measure talk time on sales calls
  • Label podcast hosts and guests
  • Find the customer's turns in a support call

Limits and caveats

  • Speakers are numbered, not named.
  • Overlapping speech is given to one speaker.
  • Very short turns (under a quarter second) are dropped.
  • Similar voices (same person on two mics, close relatives) can merge: set num_speakers when you know it.

Pricing

$0.0003 per minute. Billed per second of audio, rounded up, at least one second. Failed calls aren't charged, and a call never costs more than its estimate (POST /v1/estimate tells you before you run it).

Built on

pyannote/segmentation-3.0 + WeSpeaker ResNet34 (with our clustering), 1.5M + 6.6M parameters. License: MIT and CC-BY-4.0. We run it on our servers.