Text-to-Speech API for 10 Languages: A Developer Guide

Most TTS APIs treat non-English languages as an afterthought. The voice sounds robotic, intonation is wrong, and you can hear the English accent bleeding through. murmr takes a different approach: fewer languages, but each one sounds like a native speaker.

Supported Languages

murmr supports 10 languages, each with full VoiceDesign support:

| Language | Code | VoiceDesign | Streaming | Batch | |----------|------|-------------|-----------|-------| | Chinese (Mandarin) | zh | Yes | Yes | Yes | | English | en | Yes | Yes | Yes | | Japanese | ja | Yes | Yes | Yes | | Korean | ko | Yes | Yes | Yes | | German | de | Yes | Yes | Yes | | French | fr | Yes | Yes | Yes | | Russian | ru | Yes | Yes | Yes | | Portuguese | pt | Yes | Yes | Yes | | Spanish | es | Yes | Yes | Yes | | Italian | it | Yes | Yes | Yes |

Every language supports the full API: VoiceDesign (describe any voice), SSE streaming, batch generation, WebSocket realtime, and saved voices.

Quick Start

Generating speech in any language is a one-parameter change:

# Japanese
curl -X POST https://api.murmr.dev/v1/voices/design \
  -H "Authorization: Bearer $MURMR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "こんにちは、お元気ですか？",
    "voice_description": "A young Japanese woman with a warm, polite tone",
    "language": "ja"
  }' --output speech-ja.wav

# German
curl -X POST https://api.murmr.dev/v1/voices/design \
  -H "Authorization: Bearer $MURMR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "text": "Willkommen bei unserer Anwendung.",
    "voice_description": "A professional German man, clear and authoritative",
    "language": "de"
  }' --output speech-de.wav

Voice Descriptions by Language

VoiceDesign understands voice descriptions in any of the 10 supported languages. You can describe a voice in its native language for the most natural results, or use English descriptions for any language.

Describing Voices in Native Languages

// Chinese — describe the voice in Chinese
{
  "text": "欢迎使用我们的应用程序。",
  "voice_description": "一位温暖的中年女性播音员，语速适中，普通话标准",
  "language": "zh"
}

// Japanese — describe the voice in Japanese
{
  "text": "本日はご利用いただきありがとうございます。",
  "voice_description": "落ち着いた声の若い女性アナウンサー。丁寧で聞きやすい話し方",
  "language": "ja"
}

// Korean — describe the voice in Korean
{
  "text": "안녕하세요, 무엇을 도와드릴까요?",
  "voice_description": "친절하고 전문적인 젊은 여성 상담원",
  "language": "ko"
}

// German — describe the voice in German
{
  "text": "Herzlich willkommen zu unserem Service.",
  "voice_description": "Ein professioneller Nachrichtensprecher mit klarer Aussprache",
  "language": "de"
}

// French — describe the voice in French
{
  "text": "Bienvenue sur notre plateforme.",
  "voice_description": "Une narratrice chaleureuse avec une voix claire et posée",
  "language": "fr"
}

Proven Descriptions for Each Language

Here are voice descriptions that produce consistently good results:

| Language | Example Description | |----------|-------------------| | Chinese | "A professional female news anchor with standard Mandarin pronunciation" | | English | "A warm, clear narrator with a neutral American accent" | | Japanese | "A polite young woman with a gentle, clear speaking style" | | Korean | "A friendly, professional male announcer" | | German | "A confident male speaker with clear Hochdeutsch" | | French | "A warm female narrator with a Parisian accent" | | Russian | "A deep-voiced male narrator with clear diction" | | Portuguese | "A friendly Brazilian woman with a warm tone" | | Spanish | "A professional male narrator with a neutral Latin American accent" | | Italian | "A warm, expressive Italian woman in her 30s" |

Building Multilingual Apps

Dynamic Language Selection

For apps that serve multiple languages, select the TTS language based on user locale:

const LANGUAGE_MAP: Record<string, string> = {
  'zh-CN': 'zh',
  'zh-TW': 'zh',
  'ja-JP': 'ja',
  'ko-KR': 'ko',
  'de-DE': 'de',
  'fr-FR': 'fr',
  'ru-RU': 'ru',
  'pt-BR': 'pt',
  'pt-PT': 'pt',
  'es-ES': 'es',
  'es-MX': 'es',
  'it-IT': 'it',
  'en-US': 'en',
  'en-GB': 'en',
};

function getLanguageCode(locale: string): string {
  return LANGUAGE_MAP[locale] || 'en';
}

One Voice Per Language

For consistent branding, save a VoiceDesign voice for each language your app supports. Saved voices produce identical output every time:

// Map locales to saved voice IDs
const VOICE_IDS: Record<string, string> = {
  en: 'voice_abc123',
  ja: 'voice_def456',
  de: 'voice_ghi789',
  fr: 'voice_jkl012',
  es: 'voice_mno345',
};

async function speak(text: string, locale: string) {
  const language = getLanguageCode(locale);
  const voiceId = VOICE_IDS[language];

  const response = await fetch('https://api.murmr.dev/v1/audio/speech', {
    method: 'POST',
    headers: {
      'Authorization': `Bearer ${process.env.MURMR_API_KEY}`,
      'Content-Type': 'application/json',
    },
    body: JSON.stringify({
      input: text,
      voice: voiceId,
    }),
  });

  return response.arrayBuffer();
}

Language-Specific Tips

Chinese (Mandarin)

Mandarin is a tonal language, and murmr handles tones correctly from text. No pinyin annotation needed — just send standard simplified or traditional Chinese characters.

{
  "text": "今天天气真好，我们去公园散步吧。",
  "voice_description": "A cheerful young Chinese woman",
  "language": "zh"
}

Japanese

Japanese text can mix kanji, hiragana, and katakana. The model handles all three scripts natively. For best results with names or unusual kanji readings, add furigana in parentheses.

{
  "text": "東京タワーは日本で最も有名な観光地の一つです。",
  "voice_description": "A calm male narrator for a travel guide",
  "language": "ja"
}

Korean

Korean's agglutinative grammar produces long compound words. The model handles these naturally, including proper spacing and intonation for formal (합니다체) and informal (해요체) speech levels.

{
  "text": "오늘 회의에서 논의된 내용을 정리해 드리겠습니다.",
  "voice_description": "A professional female business presenter",
  "language": "ko"
}

European Languages (German, French, Spanish, Portuguese, Italian, Russian)

European languages work reliably with standard voice descriptions. A few tips:

• German: Use "Hochdeutsch" for standard German pronunciation. Regional dialects (Bavarian, Swiss) are not reliably reproduced.
• French: Specify "Parisian" for metropolitan French. Canadian French is less consistent.
• Spanish: Both Castilian and Latin American accents work. Specify in the description.
• Portuguese: Specify "Brazilian" or "European" — they sound quite different.
• Russian: The model handles Cyrillic text natively. No transliteration needed.

Audio Formats

All languages support the same output formats:

| Format | Endpoint | Use Case | |--------|----------|----------| | PCM (24kHz, 16-bit, mono) | Streaming | Real-time playback | | WAV | Batch (default) | Download, processing | | MP3 | Batch | Web distribution | | Opus | Batch | Low bandwidth | | AAC | Batch | Mobile apps | | FLAC | Batch | Archival quality |

Set the format via the response_format parameter on batch endpoints:

curl -X POST https://api.murmr.dev/v1/audio/speech \
  -H "Authorization: Bearer $MURMR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "input": "Bonjour le monde",
    "voice": "voice_abc123",
    "response_format": "mp3"
  }' --output speech.mp3

Latency by Language

All 10 languages run on the same Qwen3-TTS model with identical inference performance. There's no latency penalty for non-English languages — a Japanese request generates as fast as an English one.

Typical latency (streaming, first audio chunk):

• SSE streaming: ~300ms time-to-first-chunk
• WebSocket: ~200ms time-to-first-chunk
• Batch: ~1-3s for complete audio (varies by text length)

What's Not Supported (Yet)

• Mixed-language text: Switching between languages mid-sentence produces inconsistent results. Split into separate requests per language instead.
• Regional dialects: Standard accents work well, but regional variants (Bavarian German, Kansai Japanese, Cantonese) are unreliable.
• Code-switching: Common in multilingual communities (e.g., Spanglish), but not well supported. Use the dominant language.

Get Started

1. Sign up at murmr.dev — the free plan includes 10,000 characters/month
2. Try voices in the Voice Playground — test all 10 languages interactively
3. Check the API reference for the full parameter spec
4. Read the VoiceDesign guide for voice description best practices