oaistt · /v1/audio/transcriptions · /v1/realtime · part of stt

oaistt

oaistt transcribes speech with OpenAI-compatible endpoints: OpenAI itself, a local proxy, or a gateway in front of other providers' models. It is stt fixed to that API. Batch mode uploads a WAV; live sessions stream PCM over a realtime WebSocket at the rate the model asks for.

oaistt meeting.m4a                                   # multipart batch
oaistt --model whisper-1 --lang en --prompt "Collective" talk.wav
oaistt --stream talk.wav                             # gpt-realtime-whisper, 24 kHz
oaistt --stream --model mai-transcribe-2-streaming   # through a proxy, 16 kHz
oaistt --endpoint http://localhost:4000 --stream     # a local proxy

Sample rates

One client, any compatible model. A model named mai-transcribe* streams at 16 kHz even when it is reached through an OpenAI-style proxy.

oaistt --print-request --stream --model mai-transcribe-2-streaming
{"url":"wss://api.openai.com/v1/realtime?model=mai-transcribe-2-streaming&intent=transcription",
 "rate":16000,"auth_header":"authorization","model":"mai-transcribe-2-streaming",
 "session_update":{"type":"session.update","session":{"type":"transcription","audio":{"input":{
   "format":{"type":"audio/pcm","rate":16000},"noise_reduction":null,
   "transcription":{"model":"mai-transcribe-2-streaming"},"turn_detection":{"type":"server_vad"}}}}}}

Requests and URLs

Standard requests. Batch is a multipart file, model, response_format=json and optional language and prompt; realtime is the GA session.update with type: transcription.
Any base URL, once. Origins, /v1 URLs, proxy prefixes and full paths normalise without doubling /v1; HTTP(S) and WS(S) are allowed, redirects are not followed.
The provider's VAD, or none. gpt-realtime-whisper rejects turn detection, so it commits once at the end of input and still streams text as it arrives.
Keys from the environment. OPENAI_API_KEY as a Bearer header; models from OPENAI_STT_MODEL and OPENAI_REALTIME_TRANSCRIPTION_MODEL.
Prompts where they exist. --prompt biases OpenAI's models toward names and terms; Azure MAI rejects it.
Everything else is shared. Input selection, events, the daemon relay and the Nix module are the stt core's.