Audrey 0.7
Speech recognition by p99lab, built for voice agents and calls: phone audio, accented speech, meetings, noisy rooms. Hosted; file and live-streaming APIs. Self-serve at console.p99lab.com.
Results
Hugging Face Open ASR Leaderboard, English — #3 of 60 systems (table read on 26 Oct 2026), average WER 4.98. Measured by the leaderboard maintainers through this API, not by us. Eleven test sets, two of them private to the leaderboard.
| Test set | WER |
|---|---|
| AMI (meetings) | 7.28 |
| Earnings-22 (earnings calls, chunked) | 5.23 |
| GigaSpeech (podcasts, audiobooks) | 7.09 |
| LibriSpeech clean / other | 1.14 / 2.64 |
| SPGISpeech (financial calls) | 2.22 |
| URGENT 2024 (degraded, noisy speech) | 8.99 🥉 |
| Voice Arena Monsoon | 3.22 |
| VoxPopuli (European Parliament, accented English) | 2.33 |
| Private scripted | 2.78 |
| Private conversational | 11.85 🥇 |
| Average (11 sets) | 4.98 🥉 |
The leaderboard row was scored in October 2026 on the build served at that time. The build served today scores VoxPopuli 1.79 and Earnings-22 5.23 on the same clips with the same scoring (our measurement, Artificial Analysis' normaliser).
Other languages. FLEURS (read speech, 250 clips per language, our measurement; word error rate, character error rate for zh / ja):
Spanish 3.1 · Portuguese 3.6 · French 4.2 · German 4.9 · Japanese 5.0 · Mandarin 5.4 · Hindi 12.8. Thirty languages are accepted
(language=<code>); English is the strongest and the one we optimise for.
Scope of these numbers. English, single channel, 16 kHz or any common audio/video container. The public sets above are read speech, meetings, earnings calls and parliamentary speech; our own tests add phone-band agent calls and far-field rooms. Not measured here: overlapping speakers (no diarization yet) and audio under 0.05 s.
Try it
Sign up at console.p99lab.com, create a key, send a file:
curl -X POST "https://api.p99lab.com/v1/transcribe" \
-H "Authorization: Bearer $P99LAB_API_KEY" \
-F "file=@call.wav"
{"text": "…", "audio_seconds": 12.4, "model": "audrey-0.7", "formatted": true}
Written forms (numbers, dates, currency, e-mails, URLs) are formatted by default; format=false returns the raw words.
A prompt of up to 500 characters (names, products, terms) biases recognition for that request. Files up to 2 hours.
Live streaming
wss://api.p99lab.com/v1/stream: send 16 kHz PCM, receive partials while the caller speaks and a final at the end of each turn.
Turn detection is built in for English (our turn-1-mini), so finals arrive at the end of
the caller's turn rather than after a fixed silence. Finals are typically ready about 150 ms after the caller stops, on warm capacity.
Full reference, limits and error codes: p99lab.com/docs. Integrations:
LiveKit · Pipecat.
Pricing
| Use | Price |
|---|---|
| Files | $0.15 per audio hour |
| Live streaming | $0.25 per audio hour |
| Trial | free hours with each new account |
Usage is counted in seconds of audio received. Pilots and on-premises deployments: research@p99lab.com.
Privacy
Audio is held only while the transcript is produced and is then discarded. We do not store, log or train on the audio you send or the transcripts we return. Policy: p99lab.com/privacy.
The name
In 1952, Bell Labs built the first machine that could recognise speech. It understood spoken digits. They called it Audrey.
p99lab
p99lab builds models for real-time conversation: speech recognition and turn-taking. See also turn-1-mini.