Hear the voice, not the room.
Psst, our noise suppression model, removes traffic, music, fans and other people’s voices from speech. It is tuned for the words, so what it cleans is easier to understand, by people and by speech recognition.
Cleaner audio, fewer wrong words.
Noise cancellation, or speech enhancement, separates a voice from everything else in a recording: background noise, echo and other people talking. Most denoisers are tuned to make audio sound pleasant, and that can smear the very sounds that tell words apart.
Psst is tuned for what a voice agent needs: the words. In our benchmark, transcription after Psst makes 36% fewer word errors than after DeepFilterNet3, the strongest published real-time denoiser, and Psst was the only system we tested that improved on not denoising at all. It is strictly causal, working within 32 ms, so it runs live.
- LiveEvery voice agent call goes through Psst before recognition, at no extra charge.
- FilesSend a 16-bit WAV to the API, or drop it in the dashboard, and get it back cleaned.
- Voices tooBuilt to remove other people talking, the most common complaint on real calls.
- Price$0.0014 a minute, $0.084 an hour, billed per second.
One request.
curl -X POST "https://api.lokutor.com/denoise" \
-H "Authorization: Bearer YOUR_API_KEY" \
-F "audio=@noisy.wav" \
--output clean.wavReturns 16-bit PCM WAV. See the API reference.
Common questions.
What is the best noise cancellation for speech recognition?
The one that lowers word errors, which is not always the one that sounds cleanest. In our benchmark, Psst gave 36% fewer word errors after cleaning than DeepFilterNet3, which scores higher on audio-quality metrics but made transcription worse than no denoising at all.
How much does Psst cost?
$0.0014 a minute through the API ($0.084 an hour), billed per second. On voice agent calls it is included in the agent’s price.
Does it remove other people talking?
Yes. Background voices are one of the things it removes, and our benchmark mixes speech with interfering speakers as well as noise.
Can it run on my own hardware?
Yes. Psst is small and strictly causal, so it runs in real time on an ordinary CPU, on your servers or in a device.
Clean your noisiest recording.
Drop it in the dashboard and listen. $5 of credit to start.