Qwen3-TTS

Qwen3-TTS is Alibaba Qwen's open-weight text-to-speech family. It ships in 0.6B and 1.7B sizes with three task variants: Base for voice cloning and general TTS, CustomVoice with 9 preset speakers controlled by style prompts, and VoiceDesign, which creates entirely new voices from plain-language descriptions. It covers 10 languages (Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian) and is built for real-time streaming, with a first audio packet in around 97 ms. Honest limitations: the streaming latency figures and benchmark claims are provider-reported from Qwen's own test setups, so treat them as lab results rather than guarantees on your hardware. Fine-grained voice design still needs a good English or Chinese prompt to produce a stable voice, and like all cloning models it carries misuse risk for impersonation, which users should weigh before deploying.

Provider-reported (technical report): Seed-TTS zero-shot cloning, 1.7B-Base scores WER 1.24 on test-en (claimed state of the art among compared open systems) and 0.77 on test-zh. On a 10-language multilingual TTS set it reports the lowest WER in 6 of 10 languages and the highest speaker similarity in all 10 versus MiniMax-Speech and ElevenLabs Multilingual v2. On InstructTTSEval, the 1.7B VoiceDesign model claims state of the art among open-source models on Description-Speech Consistency and Response Precision in Chinese and English. The 12Hz tokenizer codec reports PESQ wideband 3.21, STOI 0.96, UTMOS 4.16 on LibriSpeech test-clean.