Getting Started with @habemus-papadum/aiui-stt

Realtime STT with the pencil's contract: segments in, cumulative deltas + finals out.

Pick an engine

	scribeTransport (DEFAULT)	openaiTranscriptionTransport
Vendor	ElevenLabs Scribe v2 realtime	OpenAI realtime, transcription sessions
Credential	single-use token per connect (connectUrl())	ek_ via the oracle's KeySource
Word timestamps	yes (segment-rebased)	no (wire has none)
Word confidence	yes (logprobs)	yes (token→word fold)
Language hint / keyterms	yes / yes	no / no

Both engines: 24 kHz PCM16 in, client-committed segments (push-to-talk), vendorVad: false in V1.

Wire it

import { createStt, scribeTransport, sttSignals } from "@habemus-papadum/aiui-stt";
import { micAudioSource, MicPicker } from "@habemus-papadum/aiui-stt/mic";

const stt = createStt({
  transport: scribeTransport({
    connectUrl: () => fetch("/my/mint/route").then((r) => r.text()), // your credential seam
    keyterms: ["morphogen", "aiui"],
  }),
  audio: micAudioSource(),
});

stt.connect();                       // eager: overlap the mint with the arm→talk gap
const seg = stt.beginSegment();      // key down
stt.endSegment();                    // key up → the segment's final arrives via events
await stt.drain();                   // before composing: wait for in-flight finals

beginSegment auto-connects and the handle queues ops through the mint window, so "hold the key and talk immediately" loses nothing even without the eager connect().

Observe it

sttSignals(stt) gives Solid signals: status, partial (the open segment's cumulative text), finals (append-only, with per-word data when the engine has it), lastError. Feed partial into aiui-viz/modal's LiveDiffText for the streaming-text look. For anything deeper, subscribe to the raw event stream (stt.subscribe), which includes diagnostic events — config echoes, vendor self-commits, unhandled message types.