Realtime speech-to-text

Three steps to your first transcript:

1. Get a license key — denthub-issued, scoped to your account.
2. Open the WebSocket and send a start frame.
3. Stream PCM16 audio and read result messages.

const key = "YOUR_LICENSE_KEY";
const ws = new WebSocket(
  "wss://api.denthub.ai/stt/realtime/stream?key=" + key + "&lang=en&backend=densper3"
);
ws.binaryType = "arraybuffer";

ws.onopen = () => {
  // 1) send the start frame (text) before any audio
  ws.send(JSON.stringify({
    type: "start",
    sample_rate: 16000,
    codec: "pcm_s16le",
    channels: 1,
    lang: "en",
    backend: "densper3",
  }));
  startMic(ws); // 2) then stream raw PCM16
};

ws.onmessage = (e) => {
  const msg = JSON.parse(e.data);
  if (msg.type === "result") console.log(`[${msg.utt_index}] ${msg.text}`);
};

async function startMic(ws) {
  const stream = await navigator.mediaDevices.getUserMedia({ audio: true });
  const ctx = new AudioContext({ sampleRate: 16000 });
  await ctx.audioWorklet.addModule("/pcm-worklet.js");
  const node = new AudioWorkletNode(ctx, "pcm-processor");
  ctx.createMediaStreamSource(stream).connect(node);
  node.port.onmessage = (e) => ws.send(e.data); // binary PCM16LE frames
}

curl can't stream WS binary frames — use wscat/Postman for manual testing (REST only for /health).

Your key rides as the key query parameter on the WebSocket URL, because browsers can't set custom headers on a wss:// connection. Keep the key server-side wherever possible; where it must live in browser code, treat it as a short-lived, scoped credential.

Send raw PCM16LE, 16 kHz, mono binary frames — no WAV header, no base64, no JSON wrapping.

Any frame size works; ~1024 bytes (32 ms) is a good default. Send continuously — the gateway runs VAD server-side, so you don't need to gate on silence.

You get one result message per utterance — final only, no partials. Each carries the transcribed text plus recording-relative timestamps, so you can line up an utterance (or a word) against the audio you sent.

{
  "type": "result",
  "utt_index": 0,
  "session_id": "b1f7…",
  "abs_start_utc": "2026-07-10T05:12:01.80Z",
  "abs_end_utc":   "2026-07-10T05:12:05.20Z",
  "duration_ms": 3400,
  "text": "Patient presents with sensitivity on the upper right quadrant.",
  "segments": [
    { "start_ms": 800, "end_ms": 4200,
      "text": "Patient presents with sensitivity…", "language": "en" }
  ],
  "words": [ { "word": "Patient", "start_ms": 800, "end_ms": 1180 }, … ],
  "status": "ok",
  "forced_cut": false,
  "continues": false
}

Timestamps (start_ms/end_ms, abs_*_utc) come from a sample-count clock, immune to jitter or processing time. Use them — never message arrival time — to place an utterance on your timeline.

Send a known sample file and assert you get it back: the same audio in should produce the same transcript, utterance count, and timestamps out.

• Prepare a 16 kHz mono PCM16 clip of known speech.
• Stream it, collect every result, and join the text fields.
• Assert the transcript, the utt_index count, and the start_ms/end_ms alignment.

Prefer to poke it by hand first? Open the Playground →