Auto Subtitles — Workers AI Nova-3 + Whisper

// Bind Workers AI as AI in wrangler.toml: [ai] binding = "AI"

// Helper: convert seconds to WebVTT timestamp "HH:MM:SS.mmm"
function toVttTime(s) {
  const h = Math.floor(s / 3600), m = Math.floor((s % 3600) / 60), sec = s % 60
  return String(h).padStart(2,'0') + ':' + String(m).padStart(2,'0') + ':' +
         sec.toFixed(3).padStart(6,'0')
}

// Helper: encode ArrayBuffer to base64 without Node.js Buffer
function bufferToBase64(ab) {
  return btoa(String.fromCharCode(...new Uint8Array(ab)))
}

export default {
  async fetch(request, env) {
    if (request.method !== 'POST') return fetch(request)

    const form     = await request.formData()
    const file     = form.get('file')
    const language = form.get('language') || 'auto'
    const buffer   = await file.arrayBuffer()

    // Model routing based on language:
    //   Thai/ASEAN  ->  Whisper large-v3-turbo  ($0.0005/min)
    //   English/EU  ->  Deepgram Nova-3          ($0.0052/min, lower WER)
    const ASEAN = ['th','id','vi','ms','km','lo','my','tl']
    const isAsean = ASEAN.includes(language)

    let transcript = ''
    let segments   = []

    if (isAsean) {
      // Whisper requires base64-encoded audio — use btoa(), not Node.js Buffer
      const base64 = bufferToBase64(buffer)
      const result = await env.AI.run('@cf/openai/whisper-large-v3-turbo', {
        audio: base64, language
      })
      transcript = result.text ?? ''
      segments   = (result.segments ?? []).map(s => ({
        start: s.start, end: s.end, text: s.text
      }))
    } else {
      // Nova-3 accepts a ReadableStream for audio body
      const stream = new ReadableStream({
        start(c) { c.enqueue(new Uint8Array(buffer)); c.close() }
      })
      const result = await env.AI.run('@cf/deepgram/nova-3', {
        audio: { body: stream, contentType: file.type },
        smart_format: true, utterances: true, detect_language: true
      })
      segments   = (result.utterances ?? []).map(u => ({
        start: u.start, end: u.end, text: u.transcript
      }))
      transcript = result.results?.channels?.[0]?.alternatives?.[0]?.transcript ?? ''
    }

    // Build WebVTT inline — no external helper needed
    const vtt = 'WEBVTT\n\n' + segments.map((s, i) =>
      (i + 1) + '\n' + toVttTime(s.start) + ' --> ' + toVttTime(s.end) + '\n' + s.text
    ).join('\n\n')

    // After STT:
    //   SEA-LION 27B  cleans per-chunk hallucinations (foreign script, garbage words)
    //   GPT-OSS 120B  corrects domain vocabulary across the full transcript

    return Response.json({ transcript, segments, vtt })
  }
}