openai: realtime STT (transcription session) rejects every model with invalid_model — caused by the `model` URL query param

[rinarakaki]

Versions

@livekit/agents-plugin-openai 1.4.4 (code unchanged on current main: plugins/openai/src/stt.ts)

Behavior

With useRealtime: true (the default for realtime-capable models), the OpenAI realtime transcription session dies immediately after session.update with:

{"type":"error","error":{"type":"invalid_request_error","code":"invalid_model",
 "message":"Model \"gpt-4o-mini-transcribe\" is not supported in transcription mode."}}

followed by ws close 4000 reason=invalid_request_error.invalid_model. Every transcription model fails the same way (gpt-4o-transcribe, gpt-4o-mini-transcribe, whisper-1), so realtime STT via this plugin currently produces zero transcripts.

Root cause

buildRealtimeSttUrl appends the model to the websocket URL:

url.searchParams.set('intent', 'transcription');
url.searchParams.set('model', model);

As of 2026-06, OpenAI's realtime endpoint treats a ?model= query param as selecting a conversation session; the subsequent transcription-mode session.update is then rejected for any transcription model. Verified directly against wss://api.openai.com/v1/realtime?intent=transcription:

• Without &model=...: identical session.update → session.updated, and audio streams transcribe correctly (deltas + completed events)
• With &model=...: the invalid_model error above, for every model tried

The model is already conveyed in the documented place for transcription sessions — session.update → audio.input.transcription.model — so the URL param is redundant.

Repro (raw ws, isolates the param)

const ws = new WebSocket('wss://api.openai.com/v1/realtime?intent=transcription&model=gpt-4o-mini-transcribe',
  { headers: { Authorization: `Bearer ${process.env.OPENAI_API_KEY}` } });
ws.on('open', () => ws.send(JSON.stringify({ type: 'session.update', session: { type: 'transcription', audio: { input: {
  format: { type: 'audio/pcm', rate: 24000 },
  transcription: { model: 'gpt-4o-mini-transcribe', language: 'en' },
  turn_detection: { type: 'server_vad', threshold: 0.5, prefix_padding_ms: 600, silence_duration_ms: 350 },
} } } })));
ws.on('message', (d) => console.log(String(d)));
// → error invalid_model + close 4000. Remove `&model=...` from the URL → session.updated + working transcription.

Suggested fix

Drop url.searchParams.set('model', model) from buildRealtimeSttUrl (keep intent=transcription). We've been running this as a one-line dist patch in production and realtime transcription works correctly.

Aggravating factor

The failure is completely silent at default log levels: the plugin throws a plain Error (not APIError, so the base class doesn't retry/log it), the base stt class emits error on the STT instance — and an application without an error listener gets an ERR_UNHANDLED_ERROR inside a floating promise, which the job child process logs at DEBUG. The stream then drains cleanly, so the app just sees zero transcription events. A logger.warn on the realtime ws error path (like the Deepgram plugin's retry logging) would make this class of failure visible.