Text-to-Speech Cheat Sheet

Practical Snippet

Key points: (1) normalise the text first — numbers into reading form, symbols into words, abbreviations expanded — which buys far more than any parameter tweak; (2) comprehensibility collapses above a rate of 1.2, so speed up by cutting content rather than accelerating speech; (3) raising pitch slightly increases duration, the opposite of intuition, so change rate alone when shortening; (4) long text must be synthesised in chunks or it tends to be truncated; (5) Web Speech accepts rate up to 10 but naturalness degrades sharply past 1.2.

// ① 文本正规化：数字与符号先转成可读形式，胜过调参数
function normalize(text) {
  return text
    .replace(/[¥$]s?([d,]+)/g, (_, n) => Number(n.replace(/,/g, '')) + ' 元')  // ¥1,200 → 一千二百元
    .replace(/&/g, ' 和 ')
    .replace(/(d{4})-(d{2})-(d{2})/g, '$1 年 $2 月 $3 日')
    .replace(/Dr./g, 'Doctor')
    .replace(/e.g./g, 'for example')
    .replace(/etc./g, 'et cetera')
}
console.log(normalize('2024-06-01 & Dr. Smith ¥1,200 etc.'))

// ② Web Speech API 调参：小步试探，不要一次跳到 1.5
function speak(text, { rate = 1.0, pitch = 1.0, volume = 1.0 } = {}) {
  if (!('speechSynthesis' in window)) throw new Error('不支持 Web Speech')
  const u = new SpeechSynthesisUtterance(normalize(text))
  u.lang = 'zh-CN'
  u.rate = Math.min(Math.max(rate, 0.1), 10)     // 数值合法但 >1.3 自然度差
  u.pitch = Math.min(Math.max(pitch, 0), 2)
  u.volume = Math.min(Math.max(volume, 0), 1)
  speechSynthesis.speak(u)
  return u
}
speak('测试语速', { rate: 1.1 })   // ✅ 先小幅试
speak('测试语速', { rate: 1.5 })   // ⚠️ 往往变成尖锐急促

// ③ 长文要分句合成，否则长文本会被截断或卡住
function speakLong(text, chunk = 200) {
  const s = normalize(text)
  for (let i = 0; i < s.length; i += chunk) {
    speak(s.slice(i, i + chunk), { rate: 1.0 })
  }
}

// ④ 先 stop 再 speak，避免队列堆积
speechSynthesis.cancel()
speak('重来一次')

// ⑤ 服务端 TTS（更可控，但要自己处理 SSML 与采样率）
// POST /tts  { text, voice, rate, pitch, format: 'mp3', sampleRate: 24000 }
// 24kHz 是清晰度与体积的平衡点；16kHz 会明显发闷