Qwen3-TTS Open Source Release

Qwen3-TTS Family is now open sourced: Voice Design, CustomVoice, and Base, 5 models (0.6B & 1.8B), Support for 10 languages.

Key Features

• Dual-Track modeling for extreme bidirectional streaming
• First audio packet after processing just a single character (97ms)
• Multi-codebook model series: 1.7B and 0.6B sizes
• Voice Design, Custom Voice (clone), Base

Benchmark

• 10 languages average WER: 1.835%
• Speaker similarity: 0.789
• Outperforms MiniMax and ElevenLabs
• Cross-lingual voice clone SOTA, surpassing CosyVoice3