实时语音转文本，延迟低于 200 毫秒：架构指南

type TranscriptState = {
  committed: string[]; // finalized segments, never rewritten
  current: string;     // latest partial, overwritten on each update
};

const onPartial = (s: TranscriptState, text: string): TranscriptState =>
  ({ ...s, current: text });

const onFinal = (s: TranscriptState, text: string): TranscriptState =>
  ({ committed: [...s.committed, text], current: "" });

// pcm-worklet.ts - registered via audioContext.audioWorklet.addModule()
class PCMWorklet extends AudioWorkletProcessor {
  process(inputs: Float32Array[][]) {
    const channel = inputs[0]?.[0]; // mono; Float32, range [-1, 1]
    if (!channel) return true;
    const pcm = new Int16Array(channel.length);
    for (let i = 0; i < channel.length; i++) {
      const s = Math.max(-1, Math.min(1, channel[i]));
      pcm[i] = s < 0 ? s * 0x8000 : s * 0x7fff;
    }
    // Transfer the buffer to the main thread without copying.
    this.port.postMessage(pcm.buffer, [pcm.buffer]);
    return true;
  }
}
registerProcessor("pcm-worklet", PCMWorklet);

// client.ts - runs in the browser. ws is an open WebSocket to your server.
const audioContext = new AudioContext({ sampleRate: 16000 });
await audioContext.audioWorklet.addModule("pcm-worklet.js");

const mediaStream = await navigator.mediaDevices.getUserMedia({
  audio: { channelCount: 1, echoCancellation: true, noiseSuppression: true },
});

const source = audioContext.createMediaStreamSource(mediaStream);
const worklet = new AudioWorkletNode(audioContext, "pcm-worklet");

// Forward each PCM frame to the server the moment it is produced.
worklet.port.onmessage = (e: MessageEvent<ArrayBuffer>) => {
  if (ws.readyState === WebSocket.OPEN) ws.send(e.data);
};
source.connect(worklet);

// Manual commit: tell the server to finalize the current segment.
const commit = () => ws.send(JSON.stringify({ type: "commit" }));

// server.ts - Node, using the `ws` library. ELEVENLABS_API_KEY and the
// recognizer URL come from the environment; see the Speech to Text reference
// for the exact path and query parameters.
import { WebSocketServer, WebSocket } from "ws";

new WebSocketServer({ port: 8080 }).on("connection", (client) => {
  // The API key stays on the server, never on the wire to the browser.
  const recognizer = new WebSocket(process.env.RECOGNIZER_WSS_URL!, {
    headers: { "xi-api-key": process.env.ELEVENLABS_API_KEY! },
  });

  // Browser -> recognizer: forward binary PCM, translate control messages.
  client.on("message", (data, isBinary) => {
    if (recognizer.readyState !== WebSocket.OPEN) return;
    if (isBinary) recognizer.send(data); // raw PCM bytes
    else if (JSON.parse(data.toString()).type === "commit")
      recognizer.send(sendCommit());
  });

  // Recognizer -> browser: normalize events into a stable shape.
  recognizer.on("message", (raw) => {
    const event = parseRecognizerEvent(raw.toString());
    if (event && client.readyState === WebSocket.OPEN)
      client.send(JSON.stringify(event));
  });

  // ... open handshake, queueing pre-open audio, and teardown on close/error
});

// The single place that knows the recognizer's wire format.
const sendCommit = (): string => JSON.stringify({ type: "commit" });

type NormalizedEvent =
  | { type: "partial"; text: string }
  | { type: "final"; text: string }
  | { type: "vad"; speaking: boolean };

function parseRecognizerEvent(raw: string): NormalizedEvent | null {
  const msg = JSON.parse(raw);
  if (msg.is_final === true || msg.type === "final")
    return { type: "final", text: msg.text ?? "" };
  if (msg.type === "vad") return { type: "vad", speaking: !!msg.speaking };
  if (typeof msg.text === "string")
    return { type: "partial", text: msg.text };
  return null;
}

// stream-stt.ts - pace ~100ms chunks at real time, then commit the tail.
const SAMPLE_RATE = 16000, CHUNK_MS = 100;
const CHUNK_BYTES = (SAMPLE_RATE * 2 * CHUNK_MS) / 1000; // 3200 bytes
const ws = new WebSocket(process.env.RECOGNIZER_WSS_URL!, {
  headers: { "xi-api-key": process.env.ELEVENLABS_API_KEY! },
});

// Send: walk the PCM buffer in 100ms chunks, sleeping between to mimic a
// live source. For audio that already arrives in real time, drop the sleep.
async function sendAudio(pcm: Buffer) {
  for (let off = 0; off < pcm.length; off += CHUNK_BYTES) {
    ws.send(pcm.subarray(off, off + CHUNK_BYTES));
    await new Promise((r) => setTimeout(r, CHUNK_MS));
  }
  ws.send(JSON.stringify({ type: "commit" })); // finalize the trailing segment
}

// Receive: print partials in place, append finals.
ws.on("message", (raw) => {
  const e = parseRecognizerEvent(raw.toString());
  if (e?.type === "final") console.log(`[final]   ${e.text}`);
  else if (e?.type === "partial") process.stdout.write(`[partial] ${e.text}\r`);
});

// benchmark.ts - measurement core; one StreamFn adapter per provider.
type StreamFn = (
  audioPath: string,
  onEvent: (kind: "partial" | "final", text: string) => void,
  result: RunResult
) => Promise<void>; // adapter sets result.lastChunkSentAt on the final chunk

interface RunResult {
  firstPartialMs?: number;
  finalLagMs?: number;
  hypothesis: string;
  partialEdits: number;
  lastChunkSentAt: number;
  startedAt: number;
}

async function measure(streamFn: StreamFn, audioPath: string): Promise<RunResult> {
  const result: RunResult = {
    hypothesis: "", partialEdits: 0, lastChunkSentAt: 0,
    startedAt: performance.now(),
  };
  let prevPartial = "";

  await streamFn(audioPath, (kind, text) => {
    const now = performance.now();
    if (kind === "partial") {
      if (text && result.firstPartialMs === undefined)
        result.firstPartialMs = now - result.startedAt;
      if (text !== prevPartial) { result.partialEdits++; prevPartial = text; }
    } else { // final
      result.hypothesis = result.hypothesis ? `${result.hypothesis} ${text}` : text;
      if (result.lastChunkSentAt)
        result.finalLagMs = now - result.lastChunkSentAt;
    }
  }, result);

  return result;
}