Great Work Labs
Models / Audio

Ear

Spoken language detection
ear-1

Which language is being spoken, from the first 30 seconds.

Price$0.10 per 1,000 files
Speed (our servers)1777 to 2216 ms
Languages~100 spoken languages
InputAudio, up to 1 min
Built onopenai/whisper-base (Apache-2.0)

Ear listens to the start of a recording and names the spoken language out of about 100, with a confidence and the runner-up guesses. Use it to route calls and voice notes to the right team, or to pick the language for a transcription model before you pay for the whole file.

Use it

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

Set GWL_API_KEY to a key from your dashboard. The bare name ear 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. Only the first 30 seconds are used.

NameTypeDefaultWhat it does
top_kinteger3How many candidate languages to return.

Output

FieldTypeMeaning
languagestringISO 639-1 code (en, es, ja).
namestringEnglish name of the language.
scorenumberConfidence, 0 to 1.
candidatesobject[]Top guesses with language, name, score.
listenednumberSeconds of audio used (at most 30).

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

Examples

Spanish

{
  "model": "ear-1",
  "input": "/examples/spanish.m4a"
}
{
  "language": "es",
  "name": "Spanish",
  "score": 0.994,
  "candidates": [
    {
      "language": "es",
      "name": "Spanish",
      "score": 0.994
    },
    {
      "language": "en",
      "name": "English",
      "score": 0.0016
    },
    {
      "language": "fr",
      "name": "French",
      "score": 0.0005
    }
  ],
  "listened": 8.68
}

Captured on our production servers: 2216 ms, 1 file, $0.0001.

English

{
  "model": "ear-1",
  "input": "/examples/voice-note.m4a",
  "top_k": 5
}
{
  "language": "en",
  "name": "English",
  "score": 0.9914,
  "candidates": [
    {
      "language": "en",
      "name": "English",
      "score": 0.9914
    },
    {
      "language": "hi",
      "name": "Hindi",
      "score": 0.001
    },
    {
      "language": "tl",
      "name": "Tagalog",
      "score": 0.0009
    },
    {
      "language": "tr",
      "name": "Turkish",
      "score": 0.0007
    },
    {
      "language": "el",
      "name": "Greek",
      "score": 0.0006
    }
  ],
  "listened": 10.73
}

Captured on our production servers: 1777 ms, 1 file, $0.0001.

Good for

  • Route voicemails and calls by language
  • Pick the right transcription or translation model
  • Tag podcast and video libraries by language
  • Spot callers who need an interpreter

Limits and caveats

  • Needs speech in the first 30 seconds: trim long intros or music first.
  • Closely related languages (Spanish and Galician, Norwegian and Danish) can be confused.

Pricing

$0.10 per 1,000 files. One unit per file, any length (only the first 30 seconds are analyzed). 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

openai/whisper-base (language head), 74M parameters. License: Apache-2.0. We run it on our servers.