elec_rag_ui/src/main/tts/systemTts.ts

import { spawn, execFile, fork, type ChildProcess, type SpawnOptions } from 'child_process'
import { EventEmitter } from 'events'
import { existsSync, readFileSync, writeFileSync, unlinkSync, readdirSync, mkdtempSync, rmSync } from 'fs'
import { join, dirname } from 'path'
import { tmpdir } from 'os'
import { which } from '../utils/which'

export interface SystemTtsOptions {
  rate?: number
  volume?: number
  pitch?: number
  voiceURI?: string
}

interface QueueItem {
  text: string
  resolve: () => void
}

interface TtsEventPayload {
  text: string
  error?: string
}

type Backend = 'bundled-vits' | 'bundled-piper' | 'bundled-espeak-ng' | 'spd-say' | 'espeak-ng' | 'espeak'
type AudioPlayer = 'aplay' | 'paplay' | 'ffplay'

const emitter = new EventEmitter()

let currentProcess: ChildProcess | null = null
let currentPlayer: ChildProcess | null = null
let queue: QueueItem[] = []
let isSpeaking = false
let detectedBackend: Backend | null = null
let detectedBackendPromise: Promise<Backend | null> | null = null
let detectedAudioPlayer: AudioPlayer | null = null
let currentOptions: SystemTtsOptions = {}

// 上次播放结束时间。PulseAudio 的 module-suspend-on-idle 会在空闲约 5 秒后挂起声卡，
// 挂起后恢复播放时 HDMI 通路需要约 1 秒才真正出声，期间写入的音频会被丢弃（表现为吞掉开头）。
// 空闲超过阈值后先播一段静音唤醒音频通路，再播放真实语音。
let lastPlaybackEndAt = 0

// 内嵌 sherpa-onnx VITS（小雅，BZNSYP 标准普通话数据集）路径
let bundledVitsDir: string | null = null
// TTS worker 子进程（独立进程加载 sherpa-onnx-node，避免与 onnxruntime-node DLL 冲突）
let vitsWorker: ChildProcess | null = null
let vitsWorkerReady = false
let vitsSampleRate = 22050
let vitsWorkerInitPromise: Promise<void> | null = null
// 当前播报的分段到达/错误回调（由 speakWithVits 注册）
let vitsSegmentHandler: ((index: number) => void) | null = null
let vitsErrorHandler: ((error: string) => void) | null = null

// 内嵌 Piper 路径
let bundledPiperBin: string | null = null
let bundledPiperDir: string | null = null
let bundledPiperModel: string | null = null
let bundledPiperConfig: string | null = null
// Piper 模型采样率（从模型 config 读取，用于裸流播放）
let bundledPiperSampleRate = 22050
// 流式播放代际标记：stop() 时递增，进行中的分段喂流循环据此中止
let ttsStreamToken = 0

// 内嵌 espeak-ng 路径
let bundledEspeakBin: string | null = null
let bundledEspeakDataDir: string | null = null
let bundledEspeakLibDir: string | null = null

export function onSystemTtsEvent(
  callback: (event: 'started' | 'done' | 'error', payload: TtsEventPayload) => void
): void {
  emitter.on('event', callback)
}

function notify(event: 'started' | 'done' | 'error', payload: TtsEventPayload): void {
  emitter.emit('event', event, payload)
}

function findBundledPiper(): {
  bin: string
  dir: string
  model: string
  config: string
} | null {
  const candidates = [
    join(process.resourcesPath, 'tts', 'piper'),
    join(__dirname, '..', '..', 'resources', 'tts', 'piper'),
    process.env.TTS_BUNDLE_DIR ? join(process.env.TTS_BUNDLE_DIR, 'piper') : ''
  ].filter(Boolean)

  for (const dir of candidates) {
    const bin = join(dir, 'piper')
    const model = join(dir, 'models', 'zh_CN-huayan-medium.onnx')
    const config = join(dir, 'models', 'zh_CN-huayan-medium.onnx.json')
    if (existsSync(bin) && existsSync(model) && existsSync(config)) {
      return { bin, dir, model, config }
    }
  }
  return null
}

/**
 * 查找内嵌的 sherpa-onnx VITS 中文模型（小雅，标准普通话）。
 * 相比 Piper 自带的 huayan（带台湾腔），该模型基于 BZNSYP 标准普通话数据集。
 */
function findBundledVits(): string | null {
  const candidates = [
    join(process.resourcesPath, 'tts', 'vits-xiao_ya'),
    join(__dirname, '..', '..', 'resources', 'tts', 'vits-xiao_ya'),
    process.env.TTS_BUNDLE_DIR ? join(process.env.TTS_BUNDLE_DIR, 'vits-xiao_ya') : ''
  ].filter(Boolean)

  for (const dir of candidates) {
    const model = join(dir, 'zh_CN-xiao_ya-medium.onnx')
    const lexicon = join(dir, 'lexicon.txt')
    const tokens = join(dir, 'tokens.txt')
    if (existsSync(model) && existsSync(lexicon) && existsSync(tokens)) {
      return dir
    }
  }
  return null
}

/**
 * VITS（小雅）的 sherpa-onnx 校验要求指定一个 espeak-ng-data 目录，历史上直接复用
 * Piper bundle 里的那份数据。Piper 自己的引擎（模型/二进制/onnxruntime/phonemize）
 * 在有 VITS 时永远不会被实际调用（detectBackend 第一优先级永远命中 VITS），
 * 所以只保留 piper/espeak-ng-data 这一个子目录，不要求整套 Piper 引擎存在。
 */
function findPiperEspeakDataDir(): string | null {
  const candidates = [
    join(process.resourcesPath, 'tts', 'piper'),
    join(__dirname, '..', '..', 'resources', 'tts', 'piper'),
    process.env.TTS_BUNDLE_DIR ? join(process.env.TTS_BUNDLE_DIR, 'piper') : ''
  ].filter(Boolean)

  for (const dir of candidates) {
    if (existsSync(join(dir, 'espeak-ng-data'))) return dir
  }
  return null
}

function findBundledEspeakNg(): string | null {
  const candidates = [
    join(process.resourcesPath, 'tts', 'espeak-ng'),
    join(__dirname, '..', '..', 'resources', 'tts', 'espeak-ng'),
    process.env.TTS_BUNDLE_DIR ? join(process.env.TTS_BUNDLE_DIR, 'espeak-ng') : ''
  ].filter(Boolean)

  for (const p of candidates) {
    if (existsSync(p)) return p
  }
  return null
}

function checkCommand(cmd: string, env?: NodeJS.ProcessEnv): Promise<boolean> {
  return new Promise((resolve) => {
    execFile(cmd, ['--version'], { env: { ...process.env, ...env } }, (error) => {
      resolve(!error)
    })
  })
}

async function detectAudioPlayer(): Promise<AudioPlayer | null> {
  if (detectedAudioPlayer) return detectedAudioPlayer
  const candidates: AudioPlayer[] = ['aplay', 'paplay', 'ffplay']
  for (const candidate of candidates) {
    if (await which(candidate)) {
      detectedAudioPlayer = candidate
      return candidate
    }
  }
  return null
}

function playerArgs(player: AudioPlayer): string[] {
  switch (player) {
    case 'aplay':
      return ['-t', 'wav', '-']
    case 'paplay':
      return ['-']
    case 'ffplay':
      return ['-nodisp', '-autoexit', '-']
  }
}

function playerArgsForFile(player: AudioPlayer, file: string): string[] {
  switch (player) {
    case 'aplay':
      return ['-t', 'wav', file]
    case 'paplay':
      return [file]
    case 'ffplay':
      return ['-nodisp', '-autoexit', file]
  }
}

/** 空闲超过该时长后，下次播放前先唤醒音频通路（略小于实测的 PulseAudio 约 5s 挂起超时） */
const WAKE_IDLE_THRESHOLD_MS = 4000
/** 唤醒静音时长。实测 HDMI 通路恢复出声约需 1 秒，取 1.5 秒留余量 */
const WAKE_SILENCE_MS = 1500
/** 唤醒播放的最长等待时间，防止播放器异常挂死阻塞播放队列 */
const WAKE_TIMEOUT_MS = WAKE_SILENCE_MS + 3000

/** 播报段间插入的静音停顿时长（叠加在模型自身韵律停顿之上，总停顿时长约 0.2s~0.4s） */
const PAUSE_LONG_MS = 200
const PAUSE_SHORT_MS = 100

/** 生成指定时长的裸静音 PCM（s16le 单声道） */
function buildSilenceRaw(durationMs: number, sampleRate: number): Buffer {
  return Buffer.alloc(Math.floor((sampleRate * durationMs) / 1000) * 2)
}

/**
 * 生成一段 22050Hz / 16bit / 单声道的静音 WAV（与 Piper 输出格式一致）。
 */
function buildSilenceWav(durationMs: number, sampleRate = 22050): Buffer {
  const dataSize = Math.floor((sampleRate * durationMs) / 1000) * 2
  const header = Buffer.alloc(44)
  header.write('RIFF', 0)
  header.writeUInt32LE(36 + dataSize, 4)
  header.write('WAVE', 8)
  header.write('fmt ', 12)
  header.writeUInt32LE(16, 16)
  header.writeUInt16LE(1, 20) // PCM
  header.writeUInt16LE(1, 22) // mono
  header.writeUInt32LE(sampleRate, 24)
  header.writeUInt32LE(sampleRate * 2, 28) // byte rate
  header.writeUInt16LE(2, 32) // block align
  header.writeUInt16LE(16, 34) // bits per sample
  header.write('data', 36)
  header.writeUInt32LE(dataSize, 40)
  return Buffer.concat([header, Buffer.alloc(dataSize)])
}

/**
 * 解析标准 WAV 文件，返回 PCM 数据与采样率。
 */
function extractWavData(buf: Buffer): { data: Buffer; sampleRate: number } {
  let sampleRate = 22050
  let offset = 12
  while (offset + 8 <= buf.length) {
    const id = buf.toString('ascii', offset, offset + 4)
    const size = buf.readUInt32LE(offset + 4)
    if (id === 'fmt ') {
      sampleRate = buf.readUInt32LE(offset + 12)
    } else if (id === 'data') {
      return { data: buf.subarray(offset + 8, offset + 8 + size), sampleRate }
    }
    offset += 8 + size + (size % 2)
  }
  // 解析失败时按 44 字节标准头兜底
  return { data: buf.subarray(44), sampleRate }
}

/**
 * 把一段静音拼到 PCM 语音前面，打包成完整 WAV。
 * 静音与语音在同一条音频流中，硬件在静音期间完成唤醒，语音部分不会被吞。
 */
function combineWavWithLeadingSilence(speech: Buffer, sampleRate: number, silenceMs: number): Buffer {
  const silenceSize = Math.floor((sampleRate * silenceMs) / 1000) * 2
  const dataSize = silenceSize + speech.length
  const header = Buffer.alloc(44)
  header.write('RIFF', 0)
  header.writeUInt32LE(36 + dataSize, 4)
  header.write('WAVE', 8)
  header.write('fmt ', 12)
  header.writeUInt32LE(16, 16)
  header.writeUInt16LE(1, 20) // PCM
  header.writeUInt16LE(1, 22) // mono
  header.writeUInt32LE(sampleRate, 24)
  header.writeUInt32LE(sampleRate * 2, 28)
  header.writeUInt16LE(2, 32)
  header.writeUInt16LE(16, 34)
  header.write('data', 36)
  header.writeUInt32LE(dataSize, 40)
  return Buffer.concat([header, Buffer.alloc(silenceSize), speech])
}

/**
 * 通过当前播放器播放一段静音，唤醒被 PulseAudio 挂起的声卡/HDMI 通路。
 * 播放器异常时按超时返回，保证播放队列不被卡死。
 */
async function wakeAudioPath(): Promise<void> {
  const player = await detectAudioPlayer()
  if (!player) return
  console.log(`[systemTts] wake audio path with ${WAKE_SILENCE_MS}ms silence (${player})`)
  await new Promise<void>((resolve) => {
    const proc = spawn(player, playerArgs(player), { stdio: ['pipe', 'ignore', 'ignore'] })
    guardStdin(proc)
    let settled = false
    const done = () => {
      if (settled) return
      settled = true
      clearTimeout(timer)
      resolve()
    }
    const timer = setTimeout(() => {
      try {
        proc.kill('SIGTERM')
      } catch {
        // ignore
      }
      done()
    }, WAKE_TIMEOUT_MS)
    proc.on('exit', done)
    proc.on('error', done)
    if (writeToStdin(proc, buildSilenceWav(WAKE_SILENCE_MS))) {
      try {
        proc.stdin?.end()
      } catch {
        // ignore
      }
    } else {
      done()
    }
  })
}

function detectBackend(): Promise<Backend | null> {
  if (detectedBackendPromise) return detectedBackendPromise
  detectedBackendPromise = (async () => {
    // 1. 优先使用内嵌 sherpa-onnx VITS 小雅模型（BZNSYP 标准普通话）。
    // 依赖 Piper bundle 里的 espeak-ng-data 作为 dataDir（sherpa-onnx 校验需要），
    // 不需要 Piper 引擎本身（bin/model/config），见 findPiperEspeakDataDir 注释。
    const bundledVits = findBundledVits()
    const piperDataDir = findPiperEspeakDataDir()
    if (bundledVits && piperDataDir) {
      bundledVitsDir = bundledVits
      bundledPiperDir = piperDataDir
      detectedBackend = 'bundled-vits'
      console.log('[systemTts] detected backend:', detectedBackend, 'at', bundledVitsDir)
      return detectedBackend
    }

    // 2. 内嵌 Piper（神经网络中文女声，带台湾腔，仅作兜底）
    const bundledPiper = findBundledPiper()
    if (bundledPiper) {
      const player = await detectAudioPlayer()
      if (player) {
        const env: NodeJS.ProcessEnv = {
          LD_LIBRARY_PATH:
            bundledPiper.dir + (process.env.LD_LIBRARY_PATH ? ':' + process.env.LD_LIBRARY_PATH : ''),
          ESPEAK_DATA_PATH: join(bundledPiper.dir, 'espeak-ng-data')
        }
        if (await checkCommand(bundledPiper.bin, env)) {
          bundledPiperBin = bundledPiper.bin
          bundledPiperDir = bundledPiper.dir
          bundledPiperModel = bundledPiper.model
          bundledPiperConfig = bundledPiper.config
          try {
            const cfg = JSON.parse(readFileSync(bundledPiper.config, 'utf-8'))
            bundledPiperSampleRate = cfg?.audio?.sample_rate ?? 22050
          } catch {
            bundledPiperSampleRate = 22050
          }
          detectedBackend = 'bundled-piper'
          console.log('[systemTts] detected backend:', detectedBackend, 'at', bundledPiperBin)
          return detectedBackend
        }
      }
    }

    // 2. 内嵌 espeak-ng 兜底
    const bundled = findBundledEspeakNg()
    if (bundled) {
      const dataDir = dirname(bundled)
      const libDir = join(dataDir, 'lib')
      const env: NodeJS.ProcessEnv = {
        ESPEAK_DATA_PATH: dataDir,
        LD_LIBRARY_PATH: libDir + (process.env.LD_LIBRARY_PATH ? ':' + process.env.LD_LIBRARY_PATH : '')
      }
      if (await checkCommand(bundled, env)) {
        bundledEspeakBin = bundled
        bundledEspeakDataDir = dataDir
        bundledEspeakLibDir = libDir
        detectedBackend = 'bundled-espeak-ng'
        console.log('[systemTts] detected backend:', detectedBackend, 'at', bundledEspeakBin)
        return detectedBackend
      }
    }

    // 3. 系统 TTS 兜底
    const candidates: Array<Exclude<Backend, 'bundled-piper' | 'bundled-espeak-ng'>> = [
      'spd-say',
      'espeak-ng',
      'espeak'
    ]
    for (const candidate of candidates) {
      if (await which(candidate)) {
        if (await checkCommand(candidate)) {
          detectedBackend = candidate
          console.log('[systemTts] detected backend:', detectedBackend)
          return detectedBackend
        }
      }
    }

    detectedBackend = null
    console.log('[systemTts] no system TTS backend available')
    return detectedBackend
  })()
  return detectedBackendPromise
}

function buildPiperCommand(text: string, options: SystemTtsOptions, sentenceSilence = '0.2'): {
  cmd: string
  args: string[]
  options?: SpawnOptions
} | null {
  if (!bundledPiperBin || !bundledPiperModel || !bundledPiperConfig) return null

  const rate = options.rate ?? 1
  // Piper length_scale: 1.0 正常，>1 更慢，<1 更快
  const lengthScale = Math.max(0.5, Math.min(2, 1 / rate))

  const env: NodeJS.ProcessEnv = {
    ...process.env,
    LD_LIBRARY_PATH:
      bundledPiperDir! + (process.env.LD_LIBRARY_PATH ? ':' + process.env.LD_LIBRARY_PATH : ''),
    ESPEAK_DATA_PATH: join(bundledPiperDir!, 'espeak-ng-data')
  }

  return {
    cmd: bundledPiperBin,
    args: [
      '--model',
      bundledPiperModel,
      '--config',
      bundledPiperConfig,
      '--length_scale',
      String(lengthScale),
      '--sentence_silence',
      sentenceSilence,
      '--output_file',
      '-'
    ],
    options: { env, stdio: ['pipe', 'pipe', 'pipe'] }
  }
}

function buildEspeakArgs(
  backend: 'bundled-espeak-ng' | 'espeak-ng' | 'espeak',
  text: string,
  options: SystemTtsOptions
): { cmd: string; args: string[]; options?: SpawnOptions } {
  const rate = options.rate ?? 1
  const volume = options.volume ?? 1
  const pitch = options.pitch ?? 1
  const wpm = Math.round(140 + (rate - 1) * 100)
  const amp = Math.round(50 + volume * 100)
  const espeakPitch = Math.round(((pitch - 0.1) / 1.9) * 99)
  const cmd = backend === 'bundled-espeak-ng' ? bundledEspeakBin! : backend

  const spawnOptions: SpawnOptions = { stdio: ['ignore', 'ignore', 'pipe'] }
  if (backend === 'bundled-espeak-ng' && bundledEspeakDataDir && bundledEspeakLibDir) {
    spawnOptions.env = {
      ...process.env,
      ESPEAK_DATA_PATH: bundledEspeakDataDir,
      LD_LIBRARY_PATH:
        bundledEspeakLibDir + (process.env.LD_LIBRARY_PATH ? ':' + process.env.LD_LIBRARY_PATH : '')
    }
  }

  const voice = backend === 'espeak' ? 'zh' : 'cmn'
  return {
    cmd,
    args: ['-v', voice, '-s', String(wpm), '-a', String(amp), '-p', String(espeakPitch), text],
    options: spawnOptions
  }
}

function buildCommand(
  backend: Backend,
  text: string,
  options: SystemTtsOptions
): { cmd: string; args: string[]; options?: SpawnOptions } | null {
  const rate = options.rate ?? 1
  const volume = options.volume ?? 1

  switch (backend) {
    case 'bundled-vits':
      return null // sherpa-onnx 进程内合成，不走外部命令
    case 'bundled-piper':
      return buildPiperCommand(text, options)
    case 'spd-say': {
      const spdRate = Math.round((rate - 1) * 50)
      const spdVolume = Math.round(volume * 100)
      const args = ['-l', 'zh', '-r', String(spdRate), '-i', String(spdVolume), text]
      if (options.voiceURI) args.push('-t', options.voiceURI)
      return { cmd: 'spd-say', args }
    }
    case 'bundled-espeak-ng':
    case 'espeak-ng':
    case 'espeak': {
      return buildEspeakArgs(backend, text, options)
    }
    default:
      return null
  }
}

function playNext(): void {
  if (isSpeaking || queue.length === 0) return
  const item = queue.shift()
  if (!item) return

  void speakInternal(item.text).then(() => {
    item.resolve()
    playNext()
  })
}

async function speakInternal(text: string): Promise<void> {
  const backend = await detectBackend()
  if (!backend) {
    console.error('[systemTts] 无可用 TTS 后端')
    notify('error', { text, error: 'no backend' })
    return
  }

  // 声卡/HDMI 通路可能已休眠。Piper 合成耗时长（约 1.5 秒），若先播唤醒静音再等合成完
  // 以新流起播，间隙期间通路会重新休眠导致二次吞音，因此 Piper 路径把静音拼进语音同一条流；
  // 其他后端起播快，沿用播前静音唤醒。
  const needWake = Date.now() - lastPlaybackEndAt > WAKE_IDLE_THRESHOLD_MS

  console.log(`[systemTts] speak (${backend}):`, text.slice(0, 80))
  notify('started', { text })

  // 渲染层已在停顿标点后断行，按行分段
  const segments = text
    .split('\n')
    .map((s) => s.trim())
    .filter((s) => s.length > 0)

  if (backend === 'bundled-vits') {
    // sherpa-onnx 进程内合成（小雅模型，标准普通话）
    return speakWithVits(segments.length > 0 ? segments : [text], needWake)
  }

  const command = buildCommand(backend, text, currentOptions)
  if (!command) {
    notify('error', { text, error: 'build command failed' })
    return
  }

  if (backend === 'bundled-piper') {
    // aplay 可用时走流式路径：按行逐段合成，裸流喂给同一个 aplay，
    // 段间插入精确时长的静音停顿，且全程同一条音频流不会被吞音。
    if (detectedAudioPlayer === 'aplay' && segments.length > 0) {
      const streamCmd = buildPiperCommand(text, currentOptions, '0')
      if (streamCmd) {
        return speakWithPiperStreaming(streamCmd, segments, needWake)
      }
    }
    return speakWithPiper(command, text, needWake)
  }

  if (needWake) {
    await wakeAudioPath()
  }

  return new Promise<void>((resolve) => {
    currentProcess = spawn(command.cmd, command.args, command.options)

    let stderr = ''
    currentProcess.stderr?.on('data', (chunk: Buffer) => {
      stderr += chunk.toString()
    })

    currentProcess.on('error', (err) => {
      console.error('[systemTts] spawn error:', err)
      cleanup()
      notify('error', { text, error: err.message })
      resolve()
    })

    currentProcess.on('exit', (code) => {
      cleanup()
      if (code !== 0 && stderr) {
        console.warn('[systemTts] speak stderr:', stderr)
      }
      notify('done', { text })
      resolve()
    })

    isSpeaking = true
  })
}

let piperTmpCounter = 0

/**
 * 分段流式播报（仅 aplay + Piper）：
 * - 一次 piper 进程按行逐段合成到临时目录（每行一个 WAV，仅一次模型加载）；
 * - 单个 aplay 以裸 PCM 流模式运行，合成出一段就立刻喂一段；
 * - 段间按标点类型插入静音停顿（长停顿 PAUSE_LONG_MS / 短停顿 PAUSE_SHORT_MS）；
 * - 全程同一条音频流，需要唤醒时只在流首写一段裸静音，不会出现间隙二次吞音。
 */
function speakWithPiperStreaming(
  command: { cmd: string; args: string[]; options?: SpawnOptions },
  segments: string[],
  needWake: boolean
): Promise<void> {
  return new Promise<void>((resolve) => {
    const sampleRate = bundledPiperSampleRate
    const segDir = mkdtempSync(join(tmpdir(), 'tts-seg-'))
    const myToken = ++ttsStreamToken
    const isAlive = () => myToken === ttsStreamToken && !resolved
    const fullText = segments.join('')

    let resolved = false
    const done = (error?: string) => {
      if (resolved) return
      resolved = true
      cleanup()
      try {
        rmSync(segDir, { recursive: true, force: true })
      } catch {
        // ignore
      }
      if (error) {
        notify('error', { text: fullText, error })
      } else {
        notify('done', { text: fullText })
      }
      resolve()
    }

    // piper：按行逐段合成，每行输出一个 WAV 到临时目录
    const piperArgs: string[] = []
    for (let i = 0; i < command.args.length; i++) {
      if (command.args[i] === '--output_file') {
        piperArgs.push('--output_dir', segDir)
        i++ // 跳过原本的 '-' 占位
      } else {
        piperArgs.push(command.args[i])
      }
    }
    currentProcess = spawn(command.cmd, piperArgs, command.options)
    guardStdin(currentProcess)

    // aplay：裸 PCM 流，采样率与模型一致
    currentPlayer = spawn(
      'aplay',
      ['-t', 'raw', '-f', 'S16_LE', '-r', String(sampleRate), '-c', '1', '-'],
      { stdio: ['pipe', 'ignore', 'pipe'] }
    )
    guardStdin(currentPlayer)

    // 声卡/HDMI 通路休眠时，先在流首写一段裸静音：静音播放期间 piper 正在合成首段，
    // 通路完成唤醒后语音紧接着到达，既不吞音也几乎不增加等待
    if (needWake) {
      writeToStdin(currentPlayer, buildSilenceRaw(WAKE_SILENCE_MS, sampleRate))
      console.log(`[systemTts] stream ${WAKE_SILENCE_MS}ms leading silence to wake audio path`)
    }

    let piperStderr = ''
    let playerStderr = ''
    let piperExited = false
    currentProcess.stderr?.on('data', (chunk: Buffer) => {
      piperStderr += chunk.toString()
    })
    currentPlayer.stderr?.on('data', (chunk: Buffer) => {
      playerStderr += chunk.toString()
    })

    currentProcess.on('error', (err) => {
      console.error('[systemTts] piper spawn error:', err)
      done(err.message)
    })
    currentPlayer.on('error', (err) => {
      console.error('[systemTts] player spawn error:', err)
      done(err.message)
    })

    currentProcess.on('exit', (code) => {
      piperExited = true
      if (code !== 0 && piperStderr) {
        console.warn('[systemTts] piper stderr:', piperStderr)
      }
      if (code !== 0 && !resolved) {
        try {
          currentPlayer?.kill('SIGTERM')
        } catch {
          // ignore
        }
        done('piper failed')
      }
    })

    currentPlayer.on('exit', (code) => {
      if (code !== 0 && playerStderr) {
        console.warn('[systemTts] player stderr:', playerStderr)
      }
      done(code !== 0 ? 'player failed' : undefined)
    })

    currentProcess.stdin?.write(segments.join('\n') + '\n', 'utf-8', (err) => {
      if (err) {
        console.error('[systemTts] write to piper stdin failed:', err)
      }
      currentProcess?.stdin?.end()
    })

    // 喂流循环：按合成产出顺序把各段裸 PCM 写入 aplay，段间插入对应时长的静音
    void (async () => {
      let fed = 0
      while (fed < segments.length && isAlive()) {
        let files: string[] = []
        try {
          files = readdirSync(segDir)
            .filter((f) => f.endsWith('.wav'))
            .sort()
        } catch {
          break
        }
        // 文件写入完成判定：已有更新的分段文件，或 piper 已整体退出
        if (files.length > fed && (piperExited || files.length > fed + 1)) {
          try {
            const wav = readFileSync(join(segDir, files[fed]))
            const { data } = extractWavData(wav)
            // 播放器可能刚被 stop() kill，写入失败说明本轮播报已终止
            if (!writeToStdin(currentPlayer, data)) break
            if (fed < segments.length - 1) {
              const longPause = /[。；！？!?]\s*$/.test(segments[fed])
              writeToStdin(
                currentPlayer,
                buildSilenceRaw(longPause ? PAUSE_LONG_MS : PAUSE_SHORT_MS, sampleRate)
              )
            }
            fed++
          } catch {
            // 文件可能尚未写完，稍后重试
          }
        } else {
          await new Promise((r) => setTimeout(r, 60))
        }
      }
      if (isAlive()) {
        try {
          currentPlayer?.stdin?.end()
        } catch {
          // ignore
        }
      }
    })()

    isSpeaking = true
  })
}

/**
 * 解析 TTS worker 脚本路径（对齐 ASR worker 的查找顺序）。
 */
function resolveVitsWorkerScriptPath(): string {
  // electron-vite 生产构建后 worker 脚本会输出到 out/main/
  const builtPathJs = join(__dirname, 'tts-worker.js')
  if (existsSync(builtPathJs)) {
    return builtPathJs
  }
  const builtPathCjs = join(__dirname, 'tts-worker.cjs')
  if (existsSync(builtPathCjs)) {
    return builtPathCjs
  }
  // 开发模式：项目根目录 resources/tts-worker.cjs
  const devPath = join(process.cwd(), 'resources', 'tts-worker.cjs')
  if (existsSync(devPath)) {
    return devPath
  }
  // electron-builder 打包目录
  const prodPath = join(process.resourcesPath, 'tts-worker.cjs')
  if (existsSync(prodPath)) {
    return prodPath
  }
  throw new Error('找不到 TTS worker 脚本')
}

/**
 * 懒启动 TTS worker 并加载模型（首次约 4 秒，之后 worker 常驻复用）。
 */
async function ensureVitsWorker(): Promise<void> {
  if (vitsWorker && !vitsWorker.killed && vitsWorker.connected && vitsWorkerReady) return
  if (vitsWorkerInitPromise) return vitsWorkerInitPromise

  vitsWorkerInitPromise = new Promise<void>((resolve, reject) => {
    const scriptPath = resolveVitsWorkerScriptPath()
    console.log('[systemTts] 启动 TTS worker：', scriptPath)

    const w = fork(scriptPath, [], {
      stdio: ['ignore', 'pipe', 'pipe', 'ipc'],
      execPath: process.execPath,
      env: {
        ...process.env,
        // 让 Electron 以类 Node.js 模式运行子进程，确保 native addon 可加载
        ELECTRON_RUN_AS_NODE: '1'
      }
    })
    vitsWorker = w
    vitsWorkerReady = false

    const timeout = setTimeout(() => {
      reject(new Error('TTS worker 初始化超时'))
    }, 30000)

    w.on('message', (msg: unknown) => {
      if (!msg || typeof msg !== 'object') return
      const m = msg as { type?: string; sampleRate?: number; error?: string; index?: number }
      if (!m.type) return
      if (m.type === 'ready') {
        vitsWorkerReady = true
        vitsSampleRate = m.sampleRate ?? 22050
        clearTimeout(timeout)
        resolve()
      } else if (m.type === 'segment') {
        vitsSegmentHandler?.(m.index ?? -1)
      } else if (m.type === 'error') {
        vitsErrorHandler?.(m.error ?? 'unknown')
      }
    })
    w.stdout?.on('data', (data: Buffer) => {
      const line = data.toString().trim()
      if (line) console.log('[tts-worker]', line)
    })
    w.stderr?.on('data', (data: Buffer) => {
      const line = data.toString().trim()
      if (line) console.error('[tts-worker err]', line)
    })
    w.on('exit', (code) => {
      console.warn('[systemTts] TTS worker 退出:', code)
      vitsWorker = null
      vitsWorkerReady = false
      vitsWorkerInitPromise = null
      vitsErrorHandler?.('worker exited')
    })

    w.send({
      cmd: 'init',
      modelDir: bundledVitsDir,
      dataDir: join(bundledPiperDir!, 'espeak-ng-data'),
      numThreads: 2
    })
  })

  try {
    await vitsWorkerInitPromise
  } catch (err) {
    vitsWorkerInitPromise = null
    throw err
  }
}

/**
 * 分段流式播报（sherpa-onnx VITS 小雅模型，标准普通话）：
 * - 合成在独立 worker 进程中进行（避免与主进程 onnxruntime-node 冲突）；
 * - 主进程把各段裸 PCM 顺序喂给同一个 aplay 流，段间插入对应时长的静音停顿；
 * - 需要唤醒时流首写裸静音，与首段合成时间重叠，几乎零额外等待。
 * worker 初始化失败时回退到 Piper（huayan）。
 */
function speakWithVits(segments: string[], needWake: boolean): Promise<void> {
  return new Promise<void>((resolve) => {
    const fullText = segments.join('')
    const finish = (error?: string) => {
      if (error) {
        notify('error', { text: fullText, error })
      } else {
        notify('done', { text: fullText })
      }
      resolve()
    }

    void (async () => {
      try {
        await ensureVitsWorker()
      } catch (err) {
        console.error('[systemTts] vits worker 初始化失败，回退到 Piper:', err)
        if (bundledPiperBin) {
          const cmd = buildPiperCommand(fullText, currentOptions, '0.2')
          if (cmd) {
            await speakWithPiper(cmd, fullText, needWake)
            finish()
            return
          }
        }
        finish('vits init failed')
        return
      }

      const player = await detectAudioPlayer()
      if (!player) {
        finish('no audio player')
        return
      }
      // 非 aplay 播放器极少出现，回退到 Piper 文件播放路径
      if (player !== 'aplay') {
        const cmd = buildPiperCommand(fullText, currentOptions, '0.2')
        if (cmd) {
          await speakWithPiper(cmd, fullText, needWake)
          finish()
          return
        }
        finish('aplay required for vits')
        return
      }

      const sampleRate = vitsSampleRate
      const speed = Math.max(0.5, Math.min(2, currentOptions.rate ?? 1))
      const gain = Math.max(0, Math.min(1, currentOptions.volume ?? 1))
      const myToken = ++ttsStreamToken
      const segDir = mkdtempSync(join(tmpdir(), 'tts-vits-'))

      let resolved = false
      const isAlive = () => myToken === ttsStreamToken && !resolved
      const done = (error?: string) => {
        if (resolved) return
        resolved = true
        // 通知 worker 停止后续合成（模型保持加载，worker 常驻不退出）
        try {
          vitsWorker?.send({ cmd: 'cancel' })
        } catch {
          // ignore
        }
        vitsSegmentHandler = null
        vitsErrorHandler = null
        cleanup()
        try {
          rmSync(segDir, { recursive: true, force: true })
        } catch {
          // ignore
        }
        finish(error)
      }

      currentPlayer = spawn(
        'aplay',
        ['-t', 'raw', '-f', 'S16_LE', '-r', String(sampleRate), '-c', '1', '-'],
        { stdio: ['pipe', 'ignore', 'pipe'] }
      )
      guardStdin(currentPlayer)

      if (needWake) {
        writeToStdin(currentPlayer, buildSilenceRaw(WAKE_SILENCE_MS, sampleRate))
        console.log(`[systemTts] stream ${WAKE_SILENCE_MS}ms leading silence to wake audio path`)
      }

      let playerStderr = ''
      currentPlayer.stderr?.on('data', (chunk: Buffer) => {
        playerStderr += chunk.toString()
      })
      currentPlayer.on('error', (err) => {
        console.error('[systemTts] player spawn error:', err)
        done(err.message)
      })
      currentPlayer.on('exit', (code) => {
        if (code !== 0 && playerStderr) {
          console.warn('[systemTts] player stderr:', playerStderr)
        }
        done(code !== 0 ? 'player failed' : undefined)
      })

      // 分段到达跟踪
      const arrived = new Set<number>()
      let lastProgressAt = Date.now()
      vitsSegmentHandler = (idx) => {
        arrived.add(idx)
        lastProgressAt = Date.now()
      }
      vitsErrorHandler = (err) => {
        console.warn('[systemTts] TTS worker error:', err)
        lastProgressAt = Date.now()
      }

      vitsWorker!.send({ cmd: 'synth', dir: segDir, segments, speed, gain })

      // 喂流循环：分段就绪一段喂一段，段间插入对应时长的静音
      void (async () => {
        let fed = 0
        while (fed < segments.length && isAlive()) {
          if (arrived.has(fed)) {
            try {
              const pcm = readFileSync(join(segDir, `seg-${fed}.pcm`))
              // 播放器可能刚被 stop() kill，写入失败说明本轮播报已终止
              if (!writeToStdin(currentPlayer, pcm)) break
              if (fed < segments.length - 1) {
                const longPause = /[。；！？!?]\s*$/.test(segments[fed])
                writeToStdin(
                  currentPlayer,
                  buildSilenceRaw(longPause ? PAUSE_LONG_MS : PAUSE_SHORT_MS, sampleRate)
                )
              }
              fed++
              lastProgressAt = Date.now()
            } catch {
              // 文件可能尚未写完，稍后重试
              await new Promise((r) => setTimeout(r, 30))
            }
          } else {
            if (Date.now() - lastProgressAt > 60000) {
              // 合成卡死：杀掉 worker（下次播报会重新拉起）并结束本轮
              try {
                vitsWorker?.kill('SIGTERM')
              } catch {
                // ignore
              }
              done('synth timeout')
              return
            }
            await new Promise((r) => setTimeout(r, 50))
          }
        }
        if (isAlive()) {
          try {
            currentPlayer?.stdin?.end()
          } catch {
            // ignore
          }
        }
      })()

      isSpeaking = true
    })().catch((err) => {
      console.error('[systemTts] speakWithVits error:', err)
      finish(String(err))
    })
  })
}

function speakWithPiper(
  command: { cmd: string; args: string[]; options?: SpawnOptions },
  text: string,
  withLeadingSilence: boolean
): Promise<void> {
  return new Promise<void>((resolve) => {
    const player = detectedAudioPlayer
    if (!player) {
      notify('error', { text, error: 'no audio player' })
      resolve()
      return
    }

    const seq = ++piperTmpCounter
    const synthFile = join(tmpdir(), `tts-synth-${process.pid}-${seq}.wav`)
    const wakeFile = join(tmpdir(), `tts-wake-${process.pid}-${seq}.wav`)

    let resolved = false
    const done = (error?: string) => {
      if (resolved) return
      resolved = true
      cleanup()
      for (const f of [synthFile, wakeFile]) {
        try {
          unlinkSync(f)
        } catch {
          // ignore
        }
      }
      if (error) {
        notify('error', { text, error })
      } else {
        notify('done', { text })
      }
      resolve()
    }

    // 先合成到临时文件，再整体播放：
    // 需要唤醒时把静音拼进同一条音频流，避免“唤醒静音→合成间隙→新流”造成的二次吞音
    const piperArgs = command.args.slice(0, -1).concat(synthFile)
    currentProcess = spawn(command.cmd, piperArgs, command.options)
    guardStdin(currentProcess)

    let piperStderr = ''
    currentProcess.stderr?.on('data', (chunk: Buffer) => {
      piperStderr += chunk.toString()
    })

    currentProcess.on('error', (err) => {
      console.error('[systemTts] piper spawn error:', err)
      done(err.message)
    })

    currentProcess.on('exit', (code) => {
      if (resolved) return
      if (code !== 0) {
        if (piperStderr) {
          console.warn('[systemTts] piper stderr:', piperStderr)
        }
        done('piper failed')
        return
      }

      let fileToPlay = synthFile
      if (withLeadingSilence) {
        try {
          const wav = readFileSync(synthFile)
          const { data, sampleRate } = extractWavData(wav)
          writeFileSync(wakeFile, combineWavWithLeadingSilence(data, sampleRate, WAKE_SILENCE_MS))
          fileToPlay = wakeFile
          console.log(`[systemTts] prepend ${WAKE_SILENCE_MS}ms silence to wake audio path`)
        } catch (err) {
          console.warn('[systemTts] 拼接唤醒静音失败，直接播放原始语音:', err)
        }
      }

      currentPlayer = spawn(player, playerArgsForFile(player, fileToPlay), {
        stdio: ['ignore', 'ignore', 'pipe']
      })

      let playerStderr = ''
      currentPlayer.stderr?.on('data', (chunk: Buffer) => {
        playerStderr += chunk.toString()
      })

      currentPlayer.on('error', (err) => {
        console.error('[systemTts] player spawn error:', err)
        done(err.message)
      })

      currentPlayer.on('exit', (playerCode) => {
        if (playerCode !== 0 && playerStderr) {
          console.warn('[systemTts] player stderr:', playerStderr)
        }
        done(playerCode !== 0 ? 'player failed' : undefined)
      })
    })

    currentProcess.stdin?.write(text, 'utf-8', (err) => {
      if (err) {
        console.error('[systemTts] write to piper stdin failed:', err)
      }
      currentProcess?.stdin?.end()
    })

    isSpeaking = true
  })
}

function cleanup(): void {
  currentProcess = null
  currentPlayer = null
  isSpeaking = false
  lastPlaybackEndAt = Date.now()
}

/**
 * 给子进程 stdin 挂 error 监听。进程被 kill 后管道断开（EPIPE），
 * 未监听的 'error' 事件会抛为未捕获异常，导致 Electron 主进程弹出
 * "A JavaScript error occurred in the main process" 对话框。
 */
function guardStdin(proc: ChildProcess | null): void {
  proc?.stdin?.on('error', (err: NodeJS.ErrnoException) => {
    if (err?.code !== 'EPIPE') {
      console.warn('[systemTts] child stdin error:', err)
    }
  })
}

/** 安全写入子进程 stdin：进程可能刚被 kill、管道已断开，写入前检查可写状态 */
function writeToStdin(proc: ChildProcess | null, data: Buffer | string): boolean {
  const stdin = proc?.stdin
  if (!stdin || stdin.destroyed || !stdin.writable) return false
  try {
    stdin.write(data)
    return true
  } catch {
    return false
  }
}

export function configureSystemTts(options: SystemTtsOptions): void {
  currentOptions = { ...currentOptions, ...options }
}

export async function systemTtsSpeak(text: string): Promise<void> {
  if (!text.trim()) return
  return new Promise<void>((resolve) => {
    queue.push({ text: text.trim(), resolve })
    playNext()
  })
}

export function systemTtsStop(): void {
  // 递增代际标记，进行中的分段喂流循环会据此中止
  ttsStreamToken++
  // 先销毁 stdin 再 kill：否则喂流循环在 kill 后的间隙仍可能向已断开的管道写入，
  // 触发 EPIPE 未捕获异常（Electron 弹出 "A JavaScript error occurred" 对话框）
  if (currentProcess) {
    try {
      currentProcess.stdin?.destroy()
    } catch {
      // ignore
    }
    try {
      currentProcess.kill('SIGTERM')
    } catch {
      // ignore
    }
    currentProcess = null
  }
  if (currentPlayer) {
    try {
      currentPlayer.stdin?.destroy()
    } catch {
      // ignore
    }
    try {
      currentPlayer.kill('SIGTERM')
    } catch {
      // ignore
    }
    currentPlayer = null
  }
  queue = []
  isSpeaking = false
  // 注意：这里不能刷新 lastPlaybackEndAt。播报前渲染进程总会先调一次 stop()，
  // 若此时没有在播内容，刷新时间戳会让"空闲时长"判断永远失效，唤醒逻辑永远不会触发。
  // 被 kill 的播放进程退出时会经 cleanup() 记录结束时间，无需在此处理。
}

export async function systemTtsAvailable(): Promise<boolean> {
  return (await detectBackend()) !== null
}