Language Support

Multilingual capabilities and language-specific features of Qwen3-TTS

Overview

Qwen3-TTS provides comprehensive multilingual support, covering 10 major languages with native-quality synthesis. The models are trained on diverse multilingual data and support cross-lingual voice cloning and generation.

Supported Languages

Chinese

普通话, Beijing Dialect, Sichuan Dialect

English

American, British, and neutral accents

Japanese

Standard Japanese (標準語)

Korean

Standard Korean (표준어)

German

Standard German (Hochdeutsch)

French

European French (français européen)

Russian

Standard Russian (русский)

Portuguese

Brazilian and European Portuguese

Spanish

European and Latin American Spanish

Italian

Standard Italian (italiano)

Language Quality Comparison

Content Consistency (WER/CER ↓)

Word Error Rate (WER) or Character Error Rate (CER) on multilingual test set - lower is better:

Language	1.7B-Base	0.6B-Base	Quality Tier
Chinese	0.928	1.145	⭐⭐⭐ Excellent
English	0.934	0.836	⭐⭐⭐ Excellent
Korean	1.755	1.741	⭐⭐⭐ Excellent
German	1.235	1.089	⭐⭐ Very Good
Italian	0.948	1.534	⭐⭐ Very Good
Portuguese	1.526	2.254	⭐⭐ Very Good
Spanish	1.126	1.491	⭐⭐ Very Good
French	2.858	2.931	⭐ Good
Russian	3.212	4.458	⭐ Good
Japanese	3.823	6.404	⭐ Good

Chinese, English, and Korean achieve the best content accuracy, making them ideal for production applications requiring high precision.

Speaker Similarity (Cosine Similarity ↑)

Speaker embedding similarity on voice cloning tasks - higher is better:

Language	1.7B-Base	0.6B-Base	Quality Tier
English	0.775	0.829	⭐⭐⭐ Excellent
Portuguese	0.817	0.794	⭐⭐⭐ Excellent
Spanish	0.814	0.812	⭐⭐⭐ Excellent
Italian	0.817	0.792	⭐⭐⭐ Excellent
Chinese	0.799	0.811	⭐⭐⭐ Excellent
Korean	0.799	0.812	⭐⭐⭐ Excellent
Russian	0.792	0.781	⭐⭐⭐ Excellent
Japanese	0.788	0.798	⭐⭐ Very Good
German	0.775	0.769	⭐⭐ Very Good
French	0.714	0.700	⭐⭐ Very Good

All languages achieve strong speaker similarity (>0.70), indicating excellent voice cloning capabilities across the board.

Speaker Native Languages

For CustomVoice models, the following 9 premium speakers are available:

Speaker	Voice Description	Native Language	Recommended Languages
Vivian	Bright, slightly edgy young female	Chinese	Chinese, English
Serena	Warm, gentle young female	Chinese	Chinese, English
Uncle_Fu	Seasoned male, low mellow timbre	Chinese	Chinese
Dylan	Youthful Beijing male, clear natural	Chinese (Beijing)	Chinese, English
Eric	Lively Chengdu male, slightly husky	Chinese (Sichuan)	Chinese
Ryan	Dynamic male, strong rhythmic drive	English	English, Chinese
Aiden	Sunny American male, clear midrange	English	English
Ono_Anna	Playful Japanese female, light nimble	Japanese	Japanese, English
Sohee	Warm Korean female, rich emotion	Korean	Korean, English

For best quality, use each speaker’s native language. However, all speakers can speak any of the 10 supported languages with reasonable quality.

Cross-Lingual Capabilities

Qwen3-TTS supports cross-lingual voice cloning, allowing you to clone a voice in one language and generate speech in another.

Cross-Lingual Performance

Mixed Error Rate (WER for English, CER for others) on cross-lingual benchmark - lower is better:

Task	1.7B-Base	0.6B-Base	Quality
Korean → English	3.09	3.48	⭐⭐⭐ Excellent
Japanese → English	3.04	3.95	⭐⭐⭐ Excellent
English → Chinese	4.77	5.66	⭐⭐ Very Good
Korean → Japanese	3.67	4.17	⭐⭐ Very Good
Korean → Chinese	4.82	8.12	⭐⭐ Very Good
English → Korean	5.14	6.83	⭐⭐ Very Good
Japanese → Korean	5.59	6.86	⭐⭐ Very Good
English → Japanese	7.21	7.74	⭐ Good
Chinese → Japanese	8.40	9.29	⭐ Good

The 1.7B model generally performs better on cross-lingual tasks, especially for Korean and Japanese target languages.

Cross-Lingual Example

from qwen_tts import Qwen3TTSModel

model = Qwen3TTSModel.from_pretrained(
    "Qwen/Qwen3-TTS-12Hz-1.7B-Base"
)

# Clone English voice and speak Chinese
wavs, sr = model.generate_voice_clone(
    text="你好，很高兴见到你。",  # Chinese text
    language="Chinese",
    ref_audio="english_speaker.wav",  # English reference
    ref_text="Hello, nice to meet you."
)

Language-Specific Considerations

Chinese (中文)

Strengths:

• Excellent accuracy (WER ~0.93 for 1.7B)
• Strong dialect support (Beijing, Sichuan)
• Native speakers available (Vivian, Serena, Uncle_Fu, Dylan, Eric)

Considerations:

• Tone accuracy is critical; may occasionally flatten in complex prosody
• Text input should use simplified or traditional Chinese consistently
• Pinyin input not officially supported

Best Practices:

wavs, sr = model.generate_custom_voice(
    text="这是一段测试文本，包含了复杂的语调和情感。",
    language="Chinese",  # Always specify
    speaker="Vivian",
    instruct="温柔地说，带有一点兴奋的语气"  # Chinese instructions work best
)

English

Strengths:

• Excellent accuracy (WER ~0.93 for 1.7B)
• Multiple native speakers (Ryan, Aiden)
• Strong cross-lingual source language

Considerations:

• Accents: models default to neutral/American accent
• British English: supported but may sound slightly American-influenced
• Contractions and informal speech handled well

Best Practices:

wavs, sr = model.generate_custom_voice(
    text="I'm really excited to show you what we've built!",
    language="English",
    speaker="Ryan",
    instruct="enthusiastic and energetic, slightly faster pace"
)

Japanese (日本語)

Strengths:

• Native speaker available (Ono_Anna)
• Good speaker similarity (0.788)

Considerations:

• Higher character error rate (~3.8-6.4%)
• Pitch accent may not always be perfect
• Kanji, hiragana, and katakana all supported

Best Practices:

wavs, sr = model.generate_custom_voice(
    text="こんにちは、お元気ですか？今日はいい天気ですね。",
    language="Japanese",
    speaker="Ono_Anna",  # Use native speaker for best quality
    instruct="明るく元気な声で"  # Japanese instructions recommended
)

Korean (한국어)

Strengths:

• Excellent accuracy (WER ~1.75)
• Native speaker available (Sohee)
• Strong cross-lingual capabilities

Considerations:

• Hangul input only (no romanization)
• Handles formal and informal speech

Best Practices:

wavs, sr = model.generate_custom_voice(
    text="안녕하세요, 만나서 반갑습니다.",
    language="Korean",
    speaker="Sohee",
    instruct="따뜻하고 친근한 목소리로"  # Korean instructions
)

German (Deutsch)

Strengths:

• Good accuracy (WER ~1.09-1.24)
• Good speaker similarity (0.77)

Considerations:

• Compound words handled well
• Umlauts (ä, ö, ü) supported
• May occasionally anglicize pronunciation

Best Practices:

wavs, sr = model.generate_voice_clone(
    text="Guten Tag! Wie geht es Ihnen heute?",
    language="German",
    ref_audio="german_speaker.wav",
    ref_text="Hallo, schön Sie kennenzulernen."
)

French (Français)

Strengths:

• Handles liaison and elision
• Reasonable accuracy for European languages

Considerations:

• Moderate error rate (~2.86-2.93)
• Nasal vowels may be approximated
• Accents (é, è, ê, etc.) should be included

Best Practices:

wavs, sr = model.generate_voice_clone(
    text="Bonjour ! Comment allez-vous aujourd'hui ?",
    language="French",
    ref_audio="french_speaker.wav",
    ref_text="Enchanté de faire votre connaissance."
)

Russian (Русский)

Strengths:

• Strong speaker similarity (0.79)
• Handles Cyrillic script

Considerations:

• Moderate error rate (~3.2-4.5)
• Stress patterns may not always be perfect
• Cyrillic input required

Best Practices:

wavs, sr = model.generate_voice_clone(
    text="Привет! Как дела?",
    language="Russian",
    ref_audio="russian_speaker.wav",
    ref_text="Здравствуйте, рад познакомиться."
)

Portuguese (Português)

Strengths:

• Excellent speaker similarity (0.817)
• Supports both Brazilian and European variants

Considerations:

• Error rate ~1.5-2.3
• Diacritics (ã, õ, ç) should be included
• May default to Brazilian pronunciation

Best Practices:

wavs, sr = model.generate_voice_clone(
    text="Olá! Como você está hoje?",
    language="Portuguese",
    ref_audio="portuguese_speaker.wav",
    ref_text="Prazer em conhecê-lo."
)

Spanish (Español)

Strengths:

• Excellent speaker similarity (0.814)
• Good accuracy (WER ~1.13-1.49)
• Supports European and Latin American variants

Considerations:

• Accent marks (á, é, í, ó, ú) should be included
• ñ character supported
• May default to Castilian pronunciation

Best Practices:

wavs, sr = model.generate_voice_clone(
    text="¡Hola! ¿Cómo estás hoy?",
    language="Spanish",
    ref_audio="spanish_speaker.wav",
    ref_text="Mucho gusto en conocerte."
)

Italian (Italiano)

Strengths:

• Excellent speaker similarity (0.817)
• Good accuracy (WER ~0.95-1.53)
• Handles double consonants well

Considerations:

• Accent marks (à, è, é, ì, ò, ù) should be included
• Regional accents not explicitly supported

Best Practices:

wavs, sr = model.generate_voice_clone(
    text="Ciao! Come stai oggi?",
    language="Italian",
    ref_audio="italian_speaker.wav",
    ref_text="Piacere di conoscerti."
)

Automatic Language Detection

Qwen3-TTS supports automatic language detection when language="Auto" is specified:

# Automatic language detection
wavs, sr = model.generate_custom_voice(
    text="Hello, 你好, こんにちは",
    language="Auto",  # Detects mixed languages
    speaker="Vivian"
)

Limitations of Auto mode:

• May misdetect short phrases or ambiguous text
• Mixed-language text (code-switching) may be inconsistent
• For best quality, explicitly specify the language when known

Multilingual Generation Tips

1. Choose the Right Speaker

   Use native speakers for best quality:

   • Chinese text → Vivian, Serena, Uncle_Fu, Dylan, Eric
   • English text → Ryan, Aiden
   • Japanese text → Ono_Anna
   • Korean text → Sohee
2. Specify Language Explicitly

   Always specify language when known to avoid detection errors:

   language="Chinese"  # Better than language="Auto"
   
3. Use Native-Language Instructions

   Instructions work best in the same language as the content:

   # Chinese text with Chinese instruction
   text="你好", instruct="温柔的语气"
   
   # English text with English instruction  
   text="Hello", instruct="gentle tone"
   
4. Test Cross-Lingual Carefully

   Cross-lingual quality varies by language pair. Test thoroughly:

   • English → Chinese: Very good
   • Japanese → Chinese: Good
   • Chinese → Japanese: Moderate

Language Roadmap

Future language support (tentative):

• Arabic (العربية)
• Hindi (हिन्दी)
• Turkish (Türkçe)
• Vietnamese (Tiếng Việt)
• Thai (ไทย)

Check the GitHub repository for updates.

Next Steps

Voice Cloning

Learn how to clone voices across languages

Custom Voice

Use premium speakers in different languages

Voice Design

Create voices with language-specific characteristics

Examples

See multilingual code examples