Qwen-Audio-3.0-TTS: More Multilingual, Easier to Direct

Our text-to-speech model, now across 16 languages.

Qwen-Audio-3.0-TTS is our latest text-to-speech model release. It ships as two variants from the same lineage:

• Flash: tuned for real-time interaction, with a first-packet latency at 300ms-level.

• Plus: tuned for high-quality generation, where naturalness and timbre fidelity matter more than speed.

API: Model Studio

Release Blog: Blog

This release focuses on four things developers actually run into in production: broader language coverage, natural-language style control, fine-grained tag control, and robustness when the reference audio isn’t clean.

Qwen-Audio-3.0-TTS-Plus currently ranks #1 on Artificial Analysis, the independent third-party TTS leaderboard.

Here’s what changed and what the numbers look like.

---

Multilingual Coverage Across 16 Languages

Qwen-Audio-3.0-TTS was optimized across English, Chinese, Japanese, Korean, German, and 16 languages total, plus improved fidelity on several Chinese dialects.

Supported languages (16): Arabic, Chinese, English, French, German, Indonesian, Italian, Japanese, Korean, Malay, Portuguese, Russian, Spanish, Tagalog, Thai, Vietnamese.

*Full language support rolling out soon.

• WER/CER (lower is better): Qwen-Audio-3.0-TTS family demonstrates the strongest overall multilingual intelligibility, achieving the best WER/CER in 10 of 16 languages. Flash delivers the lowest average WER/CER at 3.87, while Plus remains highly competitive at 3.96, both outperforming the other systems on average.

• Speaker Similarity (higher is better): Qwen-Audio-3.0-TTS shows a clear and consistent advantage in speaker similarity. Plus ranks first across all 16 languages with an average SS of 82.75, while Flash follows at 80.44, demonstrating strong and robust voice-preservation quality across diverse languages.

Demos

Reference audio:

Outputs:

• Arabic

أعلنت الحكومة اليوم عن خطة جديدة لتطوير البنية التحتية في المناطق الريفية بتكلفة تقدر بمليارات الدولارات.

• Indonesian

Sekolah dasar membentuk karakter anak melalui pendidikan akademik dan moral sejak dini.

• Portuguese

No programa de hoje, vamos descobrir os sabores autênticos da culinária alentejana, conhecida por seus pratos à base de pão, ervas aromáticas e azeite de oliva extra virgem.

• Thai

เราขออภัยสำหรับความล่าช้าในการจัดส่งคำสั่งซื้อของคุณ เนื่องจากสภาพอากาศที่ไม่ดีซึ่งส่งผลกระทบต่อระบบโลจิสติกส์ของเรา

• Vietnamese

Sức khỏe thể chất và tinh thần luôn song hành với nhau nên ta phải chăm sóc cả hai một cách cân bằng và điều độ.

• Malay

Perbicaraan juga ditanda tarikh untuk suspek menjalani perbicaraan dengan cepat.

• Filipino

Katumbas ng kalahating oras ang pamamasyal sa makakatawag ng pansin na lugar.

• Chinese dialect (Hangzhou)

爬北高峰爬得我吃力煞，腿都抖抖战了。好在顶上风吹吹，毛落胃的。

• Chinese dialect (Shannxi)

额天天跑得气喘弹腾的，你还吃呢？赶紧把嘴管住，嫑让那肉在腰上堆着咧。

• Chinese dialect (Shanghai)

屋里向装修真个是烦人，天天敲墙壁，声音响得来！灰尘落得一塌糊涂，真想快点弄清爽。

---

Style Control In Natural Language

You can describe the delivery you want in natural language instead of hand-tuning acoustic parameters.

Simple prompts with plain language steer emotion, role, scenario, and pace without any labeling expertise.

Demos

Prompt: Say the following angrily.

• Target text：After the meeting, he organized the materials into a folder and put the laptop back in his bag.

• Reference audio:

• Generated audio:

Prompt: Large hall projection, broad pacing, slight reverberant feel, lifted intonation on the welcome — a stadium announcer welcoming the crowd.

• Target text：Good evening, everyone, and welcome back to the stadium tonight. Conditions are perfect for a great game.

• Reference audio:

• Generated audio:

---

Fine-Grained Tags For Non-Verbal Details

When you need precise control over the non-verbal details — a breath, a laugh, a shift in tone — you can embed inline tags directly in the target text, like [gasp], [giggles], or [angry].

This makes the model useful for narration, games, and dubbing where the non-verbal cues carry as much as the words.

Demos

Target text：[angry] I told you three times to lock the side door, and now the entire shipment is sitting in the rain.

• Reference audio:

• Generated audio:

Target text：[giggles] I may have replaced every office pen with one that writes in glitter, but I refuse to call it a crime.

Reference audio:

Generated audio:

---

More Robust Voice Cloning From Imperfect Audio

Reference clips from the real world are rarely studio-clean. Qwen-Audio-3.0-TTS was trained with targeted acoustic simulation so speech enhancement is built into the cloning path. The model suppresses reverb and noise while preserving timbre.

In our high-noise and high-reverb tests, this release produced noticeably cleaner output than previous versions from the same degraded references.

Demos

High-Noise Inputs:

Target text：Please keep the shared kitchen clean by washing your dishes immediately after use. Food left in the refrigerator over the weekend will be discarded every Monday morning.

Reference audio:

Generated audio:

• Qwen-Audio-3.0-TTS:

• Previous version:

High-Reverb Inputs:

Target text：The coffee shop was quiet except for the gentle hum of the machine.

Reference audio:

Generated audio:

• Qwen-Audio-3.0-TTS:

• Previous version:

---

Also In This Release

• A curated preset voice library spanning 16 supported languages, so you can ship a voice without cloning one first.

• 48 kHz audio output (coming soon).

---

Try It

Qwen-Audio-3.0-TTS is available now. Grab the model here:

• API: Model Studio

• Release Blog: Blog

If you build something with it, we’d genuinely like to hear what worked and what didn’t — the failure cases are where the next version comes from.