Panel scribe

A hiring debrief is three or four people talking over one speakerphone. The useful record is not the words, it is who said them: which interviewer loved the system design, which one flagged the thin testing story. One microphone, several voices, and plain transcription flattens them into one wall of text.

Diarization is the hero, so it drives the stack. pyannoteAI Live is a cloud WebSocket: stream 16 kHz PCM, get speaker-turn labels back in real time, no local model and no GPU. Deepgram Nova-3 transcribes the words, Cerebras GPT-OSS 120B writes the scorecard, and Rime reads it back. The trio is the voice; pyannote is the ears.

The sidecar POSTs for a pre-authorized socket, then streams the room's audio to it:

resp = await self._http.post(
    PYANNOTE_REST_URL, headers={"Authorization": f"Bearer {self._api_key}"}
)
resp.raise_for_status()
body = await resp.json()
url = body.get("stream", {}).get("url") or body.get("url") or body.get("wsUrl")
self._ws = await self._http.ws_connect(url)

That audio comes from forking the STT frames, so one source feeds both Deepgram and pyannote with no second subscription:

async def stt_node(self, audio, model_settings):
    async def tee():
        async for frame in audio:
            self.sidecar.feed_frame(frame)
            yield frame

    async for event in Agent.default.stt_node(self, tee(), model_settings):
        yield event

The agent has to stay quiet. A panel talks among themselves, so every finished turn records its line, prefixed with the active speaker, then raises StopResponse, unless the turn contains the word scribe:

speaker = self.sidecar.display_speaker()
if self.sidecar.current_speaker is not None:
    new_message.content = [f"[{speaker}] {text}"]
self.transcript.append({"speaker": speaker, "text": text})
...
if TRIGGER not in text.lower():
    raise StopResponse()

Because each line enters the model's context already tagged with its speaker, the recap needs no separate state: on "scribe, recap" the LLM reads its own attributed history and calls publish_scorecard.

Diarization needs more than one voice. A solo tester hears nothing interesting, so play a panel recording into the mic or gather a few people. The eval feeds pre-labeled turns to test the recap without audio.

Build it from an empty folder in the full walkthrough, or talk to the finished agent at https://playground.mahimai.ca/demos/panel-scribe.