Great Work Labs
Models / Audio

Clear

Speech enhancement
clear-1

Background noise out, voice in front, at podcast loudness.

Price$0.001 per minute
Speed (our servers)612 to 766 ms
LanguagesAny (works on the sound of speech, not words)
InputAudio, up to 5 min

Clear takes a recording of someone talking and gives it back with the background noise removed and room echo reduced: fans, traffic, keyboards, hum, a busy kitchen. The result is 48 kHz audio, mastered to podcast loudness by default (or Spotify, YouTube, broadcast, your own target, or none), as MP3, WAV, FLAC or M4A.

strength blends in some of the original when full cleanup sounds too processed. In our tests on the VoiceBank+DEMAND test set it scored a PESQ of 3.16 (the noisy input scores 1.97).

Use it

curl https://labs.greatwork.company/v1/run \
  -H "Authorization: Bearer $GWL_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
  "model": "clear-1",
  "input": "https://labs.greatwork.company/examples/kitchen-voice-memo.mp3"
}'

Set GWL_API_KEY to a key from your dashboard. The bare name clear 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. MP3, WAV, M4A, OGG, FLAC, WebM, MP4 and more, up to 25 MB and 5 minutes.

NameTypeDefaultWhat it does
loudnessstring"podcast"podcast (-16 LUFS, the default), apple_podcasts (-16), spotify or youtube (-14), broadcast (-23), none (no mastering), or a LUFS target like -18. Quiet recordings get at most 12 dB of gain.
strengthnumber1How much cleanup, 0 to 1. Lower values blend in some of the original sound.
formatstring
mp3 | wav | flac | m4a
"mp3"Output format. MP3 is 192 kbps.
sample_rateinteger48000Output sample rate: 16000, 22050, 24000, 44100 or 48000.

Output

FieldTypeMeaning
audiostringThe cleaned audio as a data URL (data:audio/mpeg;base64,...).
formatstringmp3, wav, flac or m4a.
sample_rateintegerOutput sample rate.
durationnumberLength in seconds.
input_lufsnumberLoudness of the cleaned speech before mastering (not with loudness: none).
output_lufsnumberLoudness after mastering.
true_peak_dbnumberPeak level after mastering (at most -1.5 dB).
input_snr_dbnumberHow noisy the input was: estimated speech-to-noise ratio in dB (lower is noisier).

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

Examples

Kitchen voice memo

{
  "model": "clear-1",
  "input": "/examples/kitchen-voice-memo.mp3"
}
{
  "audio": "/examples/out/clear-1-0.mp3",
  "format": "mp3",
  "sample_rate": 48000,
  "duration": 7.03,
  "input_lufs": -26,
  "output_lufs": -16.2,
  "true_peak_db": -1.7,
  "input_snr_db": -6.4
}

Captured on our production servers: 763 ms, 8 seconds, $0.000136.

Noisy call

{
  "model": "clear-1",
  "input": "/examples/noisy-call.mp3"
}
{
  "audio": "/examples/out/clear-1-1.mp3",
  "format": "mp3",
  "sample_rate": 48000,
  "duration": 5.93,
  "input_lufs": -29.1,
  "output_lufs": -17.1,
  "true_peak_db": -1.7,
  "input_snr_db": 1.1
}

Captured on our production servers: 715 ms, 6 seconds, $0.000102.

Echoey room

{
  "model": "clear-1",
  "input": "/examples/echoey-room.mp3"
}
{
  "audio": "/examples/out/clear-1-2.mp3",
  "format": "mp3",
  "sample_rate": 48000,
  "duration": 5.71,
  "input_lufs": -21,
  "output_lufs": -16,
  "true_peak_db": -1.7,
  "input_snr_db": 0.8
}

Captured on our production servers: 612 ms, 6 seconds, $0.000102.

YouTube loudness

{
  "model": "clear-1",
  "input": "/examples/kitchen-voice-memo.mp3",
  "loudness": "youtube"
}
{
  "audio": "/examples/out/clear-1-3.mp3",
  "format": "mp3",
  "sample_rate": 48000,
  "duration": 7.03,
  "input_lufs": -26,
  "output_lufs": -14.7,
  "true_peak_db": -1.7,
  "input_snr_db": -6.4
}

Captured on our production servers: 766 ms, 8 seconds, $0.000136.

Good for

  • Podcast and interview tracks recorded at home
  • Voice notes and voicemails before transcription
  • Video voiceovers recorded on a phone
  • Call recordings for review or training

Limits and caveats

  • Room echo is reduced, not removed: in a big or bare room some of the tail stays.
  • Made for speech: music and singing in the background get treated as noise.
  • When full cleanup sounds too processed, lower strength to keep some of the original.
  • Up to 5 minutes per call: split longer recordings.

Pricing

$0.001 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).