QWEN3-TTS STEPS UP: VOICE CLONING AND VOICE DESIGN

Summary from the TLDR Design newsletter

The Qwen3-TTS family introduces two new models: Qwen3-TTS-VD-Flash for voice design via natural language instructions that control timbre, prosody, and emotion, and Qwen3-TTS-VC-Flash for 3-second voice cloning across 10 languages. Both models deliver highly expressive, humanlike speech with robust text parsing capabilities and automatic tone adjustments based on semantic content. Performance benchmarks show Qwen3-TTS-VD-Flash outperforming GPT-4o-mini-tts and Mimo-audio-7b-instruct on InstructTTS-Eval, while Qwen3-TTS-VC-Flash achieves lower word error rates than MiniMax, ElevenLabs, and GPT-4o-Audio-Preview.