16.09-音频转文本

录音/语音消息 → ASR 转文本 → 文本 → LLM 处理（摘要、提取、问答）

async function transcribeWithWorkersAI(
  audioBuffer: ArrayBuffer,
  ai: Ai
): Promise<string> {
  const response = await ai.run(
    '@cf/openai/whisper',
    { audio: [...new Uint8Array(audioBuffer)] }
  )
  return response.text
}

async function transcribeWithOpenAI(
  audioBuffer: ArrayBuffer,
  apiKey: string
): Promise<{ text: string; segments?: any[] }> {
  const formData = new FormData()
  formData.append(
    'file',
    new Blob([audioBuffer], { type: 'audio/mpeg' }),
    'audio.mp3'
  )
  formData.append('model', 'whisper-1')
  formData.append('language', 'zh')
  // 加这两行可以拿到 segment 级时间戳
  formData.append('response_format', 'verbose_json')
  formData.append('timestamp_granularities[]', 'segment')
 
  const response = await fetch(
    'https://api.openai.com/v1/audio/transcriptions',
    {
      method: 'POST',
      headers: { Authorization: `Bearer ${apiKey}` },
      body: formData,
    }
  )
 
  const data = await response.json()
  return {
    text: data.text,
    segments: data.segments,
  }
}

async function transcribeWithAssemblyAI(
  audioUrl: string,
  apiKey: string
) {
  // 发起异步转写请求
  const response = await fetch('https://api.assemblyai.com/v2/transcript', {
    method: 'POST',
    headers: { Authorization: apiKey, 'Content-Type': 'application/json' },
    body: JSON.stringify({
      audio_url: audioUrl,
      language_code: 'zh',
      speaker_labels: true,
      auto_chapters: true,
    }),
  })
  const { id } = await response.json()
 
  // 轮询直到完成（Workers 里可以用 Cron Trigger 或 Durable Object 替代）
  let result: any
  while (true) {
    const status = await fetch(
      `https://api.assemblyai.com/v2/transcript/${id}`,
      { headers: { Authorization: apiKey } }
    )
    result = await status.json()
    if (result.status === 'completed') break
    if (result.status === 'error') throw new Error(result.error)
    await new Promise((r) => setTimeout(r, 3000))
  }
 
  return {
    text: result.text,
    utterances: result.utterances ?? [],
    chapters: result.chapters ?? [],
  }
}

async function transcribeWithDeepgram(
  audioBuffer: ArrayBuffer,
  apiKey: string
): Promise<string> {
  const response = await fetch(
    'https://api.deepgram.com/v1/listen?model=nova-2&language=zh',
    {
      method: 'POST',
      headers: {
        Authorization: `Token ${apiKey}`,
        'Content-Type': 'audio/mpeg',
      },
      body: audioBuffer,
    }
  )
 
  const data = await response.json()
  return data.results.channels[0].alternatives[0].transcript
}

// 浏览器端代码：用 Web Audio API 重采样为 16kHz 单声道
async function resampleTo16k(audioFile: File): Promise<AudioBuffer> {
  const arrayBuffer = await audioFile.arrayBuffer()
  const audioCtx = new AudioContext({ sampleRate: 16000 })
  return audioCtx.decodeAudioData(arrayBuffer)
}
 
// 拿到 AudioBuffer 后，用 OfflineAudioContext 或第三方库编码为 WAV Blob
// 也可以用 lamejs 编码为 MP3，recorder.js 导出为 WebM

async function transcribeLongAudio(
  chunks: ArrayBuffer[],
  apiKey: string
): Promise<string> {
  const results: string[] = []
 
  for (const chunk of chunks) {
    const text = await transcribeWithOpenAI(chunk, apiKey)
    results.push(text.text)
  }
 
  return results.join('\n\n')
}

interface Segment {
  start: number
  end: number
  text: string
}
 
function formatSRT(segments: Segment[]): string {
  return segments
    .map((seg, i) => {
      const start = formatTime(seg.start)
      const end = formatTime(seg.end)
      return `${i + 1}\n${start} --> ${end}\n${seg.text.trim()}\n`
    })
    .join('\n')
}
 
function formatTime(seconds: number): string {
  const h = Math.floor(seconds / 3600)
  const m = Math.floor((seconds % 3600) / 60)
  const s = Math.floor(seconds % 60)
  const ms = Math.round((seconds % 1) * 1000)
  return `${pad(h)}:${pad(m)}:${pad(s)},${pad(ms, 3)}`
}
 
function pad(n: number, digits = 2): string {
  return String(n).padStart(digits, '0')
}

function cleanTranscription(text: string): string {
  return text
    // 去除重复词（口吃）
    .replace(/(\S+)\s+\1+/g, '$1')
    // 去除常见填充词
    .replace(/\s*(嗯|啊|那个|就是|然后|这个|呃)\s*/g, '')
    // 合并连续空格
    .replace(/\s{2,}/g, ' ')
    .trim()
}

async function addPunctuation(
  text: string,
  apiKey: string
): Promise<string> {
  const response = await fetch(
    'https://api.openai.com/v1/chat/completions',
    {
      method: 'POST',
      headers: {
        Authorization: `Bearer ${apiKey}`,
        'Content-Type': 'application/json',
      },
      body: JSON.stringify({
        model: 'gpt-4o-mini',
        messages: [
          { role: 'system', content: '为以下 ASR 转写文本添加标点符号。保持原文内容不变，不增删文字。' },
          { role: 'user', content: text },
        ],
      }),
    }
  )
  const data = await response.json()
  return data.choices[0]?.message?.content ?? text
}

import { Hono } from 'hono'
 
type Bindings = {
  AI: Ai
  BUCKET: R2Bucket
  ASR_PROVIDER: 'workers-ai' | 'openai'
  OPENAI_API_KEY: string
}
 
const app = new Hono<{ Bindings: Bindings }>()
 
app.post('/api/transcribe', async (c) => {
  const body = await c.req.parseBody()
  const file = body['file'] as File
 
  if (!file) {
    return c.json({ error: '缺少音频文件' }, 400)
  }
 
  // 校验文件格式
  const allowedTypes = [
    'audio/mpeg',
    'audio/wav',
    'audio/mp4',
    'audio/x-m4a',
    'audio/webm',
  ]
  if (!allowedTypes.includes(file.type)) {
    return c.json({ error: `不支持的格式：${file.type}` }, 400)
  }
 
  // 校验文件大小（25MB）
  const maxSize = 25 * 1024 * 1024
  if (file.size > maxSize) {
    return c.json({ error: '文件超过 25MB 限制' }, 400)
  }
 
  const audioBuffer = await file.arrayBuffer()
 
  // 根据配置选择转写方案
  let text: string
  let segments: Segment[] | undefined
 
  if (c.env.ASR_PROVIDER === 'workers-ai') {
    text = await transcribeWithWorkersAI(audioBuffer, c.env.AI)
  } else {
    const result = await transcribeWithOpenAI(
      audioBuffer,
      c.env.OPENAI_API_KEY
    )
    text = result.text
    segments = result.segments
  }
 
  // 清理口语化表达
  const cleanedText = cleanTranscription(text)
 
  return c.json({
    text: cleanedText,
    segments,
    metadata: {
      originalSize: file.size,
      mimeType: file.type,
      provider: c.env.ASR_PROVIDER,
    },
  })
})
 
export default app