大家好，我正在开发一个对接FreeSWITCH的AI语音机器人，核心流程是：

1. FreeSWITCH通过WebSocket把用户的通话音频传给我的Node.js服务
2. 服务用Azure STT完成语音转写，识别用户意图
3. 生成回复文本后，用Azure TTS转为RAW PCM音频
4. 把PCM按20ms的帧分片，通过WebSocket回传给FreeSWITCH，让用户听到机器人的回复

目前步骤1和2都正常工作，能正确转写用户的语音，但步骤3和4完成后，用户那边只能听到无声，完全没有机器人的语音输出。我已经检查过Azure TTS的配置，也确认生成的PCM数据有内容，但就是无法在通话中播放出来。

以下是我的完整代码，麻烦大家帮忙看看哪里可能出问题了：

import fs from "fs";
import path from "path";
import { WebSocketServer } from "ws";
import sdk from "microsoft-cognitiveservices-speech-sdk";
import pino from "pino";
import "dotenv/config";

/* ============================
   CONFIG
============================ */
const WS_PORT = 3000;
const RECORD_DIR = "./calls";
const logger = pino({ level: "info" });

if (!fs.existsSync(RECORD_DIR)) {
  fs.mkdirSync(RECORD_DIR, { recursive: true });
}

/* ============================
   AZURE BASE CONFIG
============================ */
const baseSpeechConfig = sdk.SpeechConfig.fromSubscription(
  process.env.AZURE_SPEECH_KEY,
  process.env.AZURE_SPEECH_REGION
);

baseSpeechConfig.speechRecognitionLanguage = "en-IN";

/* ============================
   WEBSOCKET SERVER
============================ */
const wss = new WebSocketServer({
  port: WS_PORT,
  host: "127.0.0.1",
  perMessageDeflate: false
});

logger.info(` audio_fork WS listening on 127.0.0.1:${WS_PORT}`);

/* ============================
   TTS → RAW PCM
============================ */
function synthesizeToRawPCM(text, sampleRate) {
  return new Promise((resolve, reject) => {
    const chunks = [];

    const speechConfig = sdk.SpeechConfig.fromSubscription(
      process.env.AZURE_SPEECH_KEY,
      process.env.AZURE_SPEECH_REGION
    );

    speechConfig.speechSynthesisOutputFormat =
      sampleRate === 16000
        ? sdk.SpeechSynthesisOutputFormat.Raw16Khz16BitMonoPcm
        : sdk.SpeechSynthesisOutputFormat.Raw8Khz16BitMonoPcm;

    const pullStream = sdk.AudioOutputStream.createPullStream();
    const audioConfig = sdk.AudioConfig.fromStreamOutput(pullStream);
    const synth = new sdk.SpeechSynthesizer(speechConfig, audioConfig);

    synth.speakTextAsync(
      () => {
        let chunk;
        while ((chunk = pullStream.read()) !== undefined) {
          if (chunk?.byteLength) {
            chunks.push(Buffer.from(chunk));
          }
        }
        synth.close();
        resolve(Buffer.concat(chunks));
      },
      err => {
        synth.close();
        reject(err);
      }
    );
  });
}

/* ============================
   SEND AUDIO → FREESWITCH
   (20ms frames, realtime)
============================ */
async function sendAudio(ws, pcm, sampleRate) {
  if (ws.readyState !== ws.OPEN) return;

  const FRAME_BYTES = sampleRate === 16000 ? 640 : 320;

  for (let i = 0; i < pcm.length; i += FRAME_BYTES) {
    const frame = pcm.slice(i, i + FRAME_BYTES);

    ws.send(JSON.stringify({
      type: "playAudio",
      data: {
        audioContentType: "raw",
        sampleRate,
        audioContent: frame.toString("base64")
      }
    }));

    await new Promise(r => setTimeout(r, 20));
  }
}

/* ============================
   CONNECTION HANDLER
============================ */
wss.on("connection", ws => {
  logger.info(" audio_fork connected");

  const callId = Date.now();
  let sampleRate = null;
  let speaking = false;

  let pushStream;
  let recognizer;

  const rawFile = path.join(RECORD_DIR, `call-${callId}.raw`);
  const transcriptFile = path.join(RECORD_DIR, `call-${callId}.txt`);

  const audioRecorder = fs.createWriteStream(rawFile, { flags: "a" });
  const transcriptWriter = fs.createWriteStream(transcriptFile, { flags: "a" });

  function initSTT(rate) {
    logger.info(` Initializing Azure STT @ ${rate} Hz`);

    const format = sdk.AudioStreamFormat.getWaveFormatPCM(rate, 16, 1);
    pushStream = sdk.AudioInputStream.createPushStream(format);

    const audioCfg = sdk.AudioConfig.fromStreamInput(pushStream);
    recognizer = new sdk.SpeechRecognizer(baseSpeechConfig, audioCfg);

    recognizer.recognized = async (_, e) => {
      if (e.result.reason !== sdk.ResultReason.RecognizedSpeech) return;
      if (speaking) return;

      const text = e.result.text?.trim();
      if (!text) return;

      transcriptWriter.write(`USER: ${text}\n`);
      logger.info(` ${text}`);

      speaking = true;

      const reply = text.toLowerCase().includes("support")
        ? "Sure. Please describe your issue."
        : "Sorry, can you repeat that?";

      transcriptWriter.write(`BOT: ${reply}\n`);

      try {
        const pcm = await synthesizeToRawPCM(reply, sampleRate);
        await sendAudio(ws, pcm, sampleRate);
      } catch (err) {
        logger.error(err, "TTS failed");
      }

      speaking = false;
    };

    recognizer.startContinuousRecognitionAsync();
  }

  ws.on("message", (data, isBinary) => {
    if (!isBinary) {
      try {
        const meta = JSON.parse(data.toString());

        if (meta?.sampling_rate) {
          if (meta.sampling_rate === "16k" || meta.sampling_rate === "16000" || meta.sampling_rate === "16") {
            sampleRate = 16000;
          } else if (meta.sampling_rate === "8k" || meta.sampling_rate === "8000" || meta.sampling_rate === "8") {
            sampleRate = 8000;
          }

          logger.info(` FreeSWITCH sampleRate=${sampleRate}`);

          if (!recognizer && sampleRate) {
            initSTT(sampleRate);
          }
        }
      } catch { }
      return;
    }

    audioRecorder.write(data);

    if (!speaking && pushStream) {
      pushStream.write(data);
    }
  });

  ws.on("close", () => {
    logger.info(`❌ Call ${callId} ended`);
    audioRecorder.end();
    transcriptWriter.end();
    pushStream?.close();
    recognizer?.stopContinuousRecognitionAsync();
  });
});

/* ============================
   SHUTDOWN
============================ */
process.on("SIGINT", () => {
  logger.info("Shutting down WS server");
  wss.close();
  process.exit(0);
});

我自己怀疑过几个点，但还没找到问题：

• 音频格式匹配问题：我用的是和FreeSWITCH传来的采样率一致的RAW PCM（16位单声道），Azure TTS的输出格式也对应设置了，但会不会还有其他参数不匹配？
• 帧大小计算问题：16000Hz下20ms的帧应该是16000*0.02*2=640字节，8000Hz下是8000*0.02*2=320字节，这个计算是对的吗？
• WebSocket消息格式问题：我用的是type: "playAudio"的JSON格式，FreeSWITCH的WebSocket音频播放是不是需要特定的消息格式？
• PCM数据的字节序问题：Azure TTS输出的是小端还是大端？会不会和FreeSWITCH要求的不一致？

麻烦有经验的朋友帮忙看看，或者给我一些排查的方向，谢谢大家！