Qwen3-TTS-Flash Update: 49 Timbres, 10 Languages, and 9 Dialects

Qwen has released an update to Qwen3-TTS-Flash, a flagship text-to-speech (TTS) model designed for natural and expressive speech synthesis across multiple timbres, languages, and dialects. The update significantly expands the model's vocal diversity and linguistic reach, making it available via the Qwen API.

Expanded Timbre and Persona Support

Qwen3-TTS-Flash now offers over 49 high-quality timbres. These voices cover a broad spectrum of genders, ages, regional traits, and character profiles to suit various application scenarios. Specific character personas include:

  • Momo: Playful and quirky
  • Ono Anna: Warm and supportive childhood friend
  • Vivian: Proud and forthright "tough girl"
  • Elias: Strict instructor
  • Eldric Sage: Wise elder
  • Bunny: Cute loli

Multilingual and Multi-dialect Capabilities

Global Language Support

The model supports 10 major languages: Chinese, English, German, Italian, Portuguese, Spanish, Japanese, Korean, French, and Russian. In performance benchmarking using the MiniMax TTS multilingual test set, Qwen3-TTS-Flash achieved a lower average word error rate (WER) than GPT-4o-Audio-Preview, ElevenLabs, and MiniMax.

Chinese Dialect Synthesis

Qwen3-TTS-Flash provides authentic reproduction of local accents and linguistic nuances for nine Chinese dialects:

  • Mandarin
  • Hokkien
  • Wu
  • Cantonese
  • Sichuanese
  • Beijing
  • Nanjing
  • Tianjin
  • Shaanxi

Improvements in Prosody and Human-Likeness

Qwen3-TTS-Flash features enhanced adaptive adjustments for speech rate and prosody based on the provided textual input. Compared to previous versions, these improvements result in speech that more closely approximates real human conversation in terms of rhythm and intonation.

Technical Implementation and API Usage

Developers can access the model via the DashScope SDK using the qwen3-tts-flash-2025-11-27 model identifier. The API allows users to specify the voice (timbre) and language_type to ensure correct pronunciation and natural intonation.

import dashscope

response = dashscope.MultiModalConversation.call(
    model="qwen3-tts-flash-2025-11-27",
    api_key=os.getenv("DASHSCOPE_API_KEY"),
    text=text,
    voice="Ryan",
    language_type="English",
    stream=False
)

Sources