What p99lab makes
p99lab makes speech models for real conversations. Two models cover the two hard parts of a conversation:
- Audrey turns speech into text. It is a hosted API: one HTTP request per recording, or a WebSocket for live audio. It is built for business calls, with accented speakers, people talking over each other and noisy phone lines, and it writes names, numbers and amounts the way people write them.
- turn-1 decides when the speaker has finished. It is open weights under Apache-2.0, so you run it yourself, next to any speech recognizer, language model and voice.
The models
Speech to text · Hosted API
Audrey 0.7
Accurate English text from calls and meetings. Files over HTTP, live audio over a WebSocket.
Turn detection · Open weights
turn-1
Knows when the speaker has finished. The larger model, for servers.
Turn detection · Open weights
turn-1-mini
End of turn and interruptions, in 5.3M parameters.
Helpful links
- Quickstart Send your first recording to Audrey.
- Live streaming Get text while people are still talking.
- API reference Every endpoint, field and message.
- Errors and limits Status codes, messages and limits.
- Integrations turn-1 with Pipecat and LiveKit, or any stack.
Getting access
Sign up in the console: every new account starts with $100 of free
credit, and no card is needed. Create API keys there, and follow usage and balance. The base URL is
https://api.p99lab.com.
From $0.15 per audio hour (batch) · from $0.25 (streaming). Volume pricing available.
turn-1 and turn-1-mini are free to download from Hugging Face.