Alibaba Qwen launches Qwen-Audio-3.0-TTS for real-time and premium speech
千问大模型 · wechat · 2026-07-20
Alibaba Qwen has officially launched Qwen-Audio-3.0-TTS, a new text-to-speech model family with two variants: Flash for real-time interaction and Plus for higher-quality generation.
Key upgrades
- Fine-grained tag control: supports structured tags such as [gasp], [giggles], and [angry] to control breathing, emotion, and delivery details.
- Free-style natural language control: users can describe role, mood, scene, and speaking rate in plain language without speech-specific annotation knowledge.
- Multilingual and dialect coverage: supports 16 languages including Chinese, English, Japanese, Korean, and German, plus new coverage for Arabic, Vietnamese, Malay, and Filipino.
- Dialect synthesis: now covers 20 Chinese dialects, including Cantonese, Chongqing, Northeastern, Shaanxi, Shanghai, Sichuan, and Yunnan.
- Robustness: improved handling of noisy or reverberant reference audio for voice cloning.
Reported results
- On ArtificialAnalysis, Qwen-Audio-3.0-TTS-Plus took the top spot.
- On 16-language evaluations, the family reportedly reached SOTA in 10 languages for word/character error rate, with especially strong gains in Korean and Malay.
- In speaker similarity tests, Plus was best across all 16 languages, while Flash placed second in 15 cases.
The post also says the models are now available on Alibaba Cloud Bailian and via the project blog.
Related event: Alibaba Releases Qwen-Audio-3.0-TTS Voice Synthesis Model(5 posts)→
More from Models
- Anthropic's Terra returns with better writing and 2X usage vs 6.1 Sol — weswinder · 2026-10-27
- User catches model failing to parse 400-paper file, citing stale news sources instead — jungofthewon · 2026-10-27
- Odd eval result: Gemini 4.1 Flash gets more accurate when asked to control its CoT — dfrsrchtwts · 2026-10-27
- Poster surprised that v4.1 flash gets more accurate when asked to control its CoT — dfrsrchtwts · 2026-10-27
- Dev claims Anthropic is baking 'sabotage' into model behavior for 'wrong' engineering — cephaloform · 2026-10-27
- Gemini fails at listing unread emails, tells user to search Gmail manually — gethackteam · 2026-10-27