add-dictation

// server (any runtime with fetch + FormData)
export async function transcribe(blob: Blob, filename: string) {
  const form = new FormData();
  form.append("model", "grok-voice-transcribe-2.0"); // latest from docs; re-read model-selection
  form.append("format", "true");     // written-form numbers/currency; requires language
  form.append("language", "en");
  // form.append("keyterm", "Acme"); // repeat per term, ≤100 terms × 50 chars
  form.append("file", blob, filename); // last
  const res = await fetch("https://api.x.ai/v1/stt", {
    method: "POST",
    headers: { Authorization: `Bearer ${process.env.XAI_API_KEY}` },
    body: form,
  });
  if (!res.ok) throw new Error(`STT ${res.status}`); // 400 bad input, 413 >500 MB, 429 back off, 502 url fetch failed, 503 retry
  return (await res.json()) as {
    text: string; language: string; duration: number;
    words?: { text: string; start: number; end: number; speaker?: number }[];
    channels?: { index: number; text: string; words: unknown[] }[];
  };
}

// client
const mime = MediaRecorder.isTypeSupported("audio/webm;codecs=opus") ? "audio/webm;codecs=opus" : "audio/mp4";
const rec = new MediaRecorder(stream, { mimeType: mime });
const parts: BlobPart[] = [];
rec.ondataavailable = (e) => parts.push(e.data);
rec.onstop = async () => {
  const fd = new FormData();
  fd.append("file", new Blob(parts, { type: mime }), "dictation");
  const { text } = await (await fetch("/api/dictation", { method: "POST", body: fd })).json();
  insertAtCursor(text);
};
rec.start(); // second tap: rec.stop()

import { WebSocketServer, WebSocket } from "ws";

new WebSocketServer({ port: 8788 }).on("connection", (client) => {
  const q = new URLSearchParams({ model: "grok-voice-transcribe-2.0", sample_rate: "16000", encoding: "pcm", interim_results: "true", language: "en" });
  const up = new WebSocket(`wss://api.x.ai/v1/stt?${q}`, { headers: { Authorization: `Bearer ${process.env.XAI_API_KEY}` } });
  up.on("message", (d) => client.send(d.toString()));                       // transcript.* and error events
  client.on("message", (d, isBinary) => up.readyState === WebSocket.OPEN && up.send(d, { binary: isBinary })); // audio + finalize/audio.done
  const end = () => { client.close(); up.close(); };
  up.on("close", end); up.on("error", end); client.on("close", end);
});

const ws = new WebSocket(relayUrl); ws.binaryType = "arraybuffer";
const ctx = new AudioContext({ sampleRate: 16000 }); // if ctx.sampleRate !== 16000, downsample in the worklet
await ctx.audioWorklet.addModule("/pcm16-worklet.js"); // Float32 → Int16LE, posts one 3,200-byte frame per 100 ms
const node = new AudioWorkletNode(ctx, "pcm16");
ctx.createMediaStreamSource(stream).connect(node);
let ready = false;
node.port.onmessage = (e) => ready && ws.readyState === WebSocket.OPEN && ws.send(e.data);

let committed = "", locked = "", live = "";
ws.addEventListener("message", (e) => {
  const ev = JSON.parse(e.data);
  if (ev.type === "transcript.created") ready = true;
  else if (ev.type === "transcript.partial") {
    if (ev.speech_final) { committed += ev.text + " "; locked = ""; live = ""; } // complete stitched utterance
    else if (ev.is_final) { locked += ev.text + " "; live = ""; }                // chunk final: text will not change
    else live = ev.text;                                                          // interim: may change
    render(committed + locked + live);
  } else if (ev.type === "transcript.done") ws.close();                          // after audio.done
  else if (ev.type === "error") showError(ev.message);                           // most errors close the socket
});
// stop: ws.send(JSON.stringify({ type: "audio.done" }))
// push-to-talk release: ws.send(JSON.stringify({ type: "Finalize" })) then keep streaming (docs show both `finalize` and `Finalize`; the examples use `Finalize`)

import os, requests
r = requests.post(
    "https://api.x.ai/v1/stt",
    headers={"Authorization": f"Bearer {os.environ['XAI_API_KEY']}"},
    data=[("model", "grok-voice-transcribe-2.0"), ("format", "true"), ("language", "en")],
    files={"file": ("dictation.webm", blob, "audio/webm")},  # requests sends data fields before files
)
r.raise_for_status(); text = r.json()["text"]
# streaming: websockets.connect(url, additional_headers={"Authorization": f"Bearer {key}"}); await ws.send(pcm_bytes)