Qwen Audio 3.0 TTS is Alibaba's multilingual text-to-speech model, built for creators who need both emotion control and low latency. It handles long scripts efficiently across 16 languages with fine-grained emotion tags and one TTS route.

Why creators choose Qwen Audio 3.0 TTS

Qwen Audio 3.0 TTS wins on flexibility. It gives you one TTS route, fine-grained emotion control, and support for 16 languages, so one model covers most narration use cases.

Emotion tags for precise delivery

Embed tags like (whisper), (angry), (laughs), and (breaths) inline in your script. The model generates those emotional and non-verbal cues as part of the narration, not as separate stems. Your script controls the performance.

One route for narration on Craze

Craze exposes one Qwen text-to-speech route. Write the script and regenerate a take when you want to try a different delivery.

Handles long scripts in one pass

Generate up to 2,000 characters of script in a single call. No splicing, no continuity breaks. That's practical for documentaries, audiobooks, and tutorials where you want the full narration at once.

Text-first narration

The Craze route accepts a script and an optional voice ID. It does not take a reference-audio upload.

How to use Qwen Audio 3.0 TTS in Craze

1. Write or paste your script. Add emotion tags like (warmly) or (urgent) to steer the delivery.
2. Optionally choose a voice ID, then name Qwen Audio 3.0 TTS in chat or select it in the Models panel.
3. Craze generates the narration and places it in your project.
4. Trim, adjust timing, or regenerate a section if you want a different emotional take.

Qwen Audio 3.0 TTS pricing

Pricing starts at $0.053 per 1,000 characters and varies with script length. The calculator on this page shows the total before you generate.

Who Qwen Audio 3.0 TTS is for

Emotional storytelling and character narration

When tone and emotion are central to your story, Qwen's inline tags let you direct the performance. A character can deliver the same line warm, angry, or afraid (all in your script).

Long-form content and documentation

Audiobooks, courses, and technical documentation can be split into scripts of up to 2,000 characters per generation and assembled in the timeline.

Rapid iteration and A/B testing

Regenerate takes to try different emotional readings of the same script.

Multilingual projects with consistent voice

Narrate a global campaign in 16 languages with the same voice and emotional consistency across all versions. One voice, many markets.

Qwen Audio 3.0 TTS vs Other Models

Gemini 3.1 Flash TTS is Google's expressive model with steerable prompts optimized for English narration. Qwen Audio 3.0 TTS is built for multilingual work with inline emotion tags and one TTS route. Pick Gemini for directed English narration with fine pacing control. Pick Qwen for multilingual scripts and emotion-driven delivery across languages.

Seed Speech from ByteDance offers text narration with optional voice selection, while Qwen focuses on emotion control and multilingual support. MiniMax Speech 2.8 HD includes non-speech sounds like laughs and sighs, while Qwen is optimized for pure dialogue and narration with emotion tags. Seed Audio 1.0 also accepts audio or image references for speech.

Emotion Tags and Delivery Control

Qwen Audio 3.0 TTS recognizes standard emotion tags like (whisper), (angry), (happy), (sad), (laughs), (coughs), and (breaths). You can also use natural language descriptions in the prompt: "speak this line like a news anchor" or "sound nervous and hesitant here." The model combines both approaches, so mix inline tags with prompt direction for fine control.

Stacked emotions work on Qwen. Mark a line as "(angry but controlled)" or "(excited but whispering)" for nuanced delivery. The model balances competing emotional directions intelligently, so you get a complex read rather than one emotion overwriting another. This is valuable for character work where a speaker feels conflicted or shifting emotions within a single line.

Directing a single TTS route

For different emotional reads, adjust the inline tags and regenerate. Listen to each take before placing it in a longer sequence.

On Craze, you can refine line delivery by changing the script and tags, then regenerate with the same route for a final take.

Multilingual Support and Voice Consistency

Qwen handles 16 languages including English, Spanish, French, German, Italian, Portuguese, Dutch, Polish, Russian, Turkish, Arabic, Chinese (Simplified and Traditional), Japanese, Korean, and Vietnamese. An optional voice ID lets you select a voice for different scripts. Code-switching (mixing languages mid-script) works within reason.

If a language needs a different accent, try a different available voice ID and review the generated take.

Script Limits and Performance

Qwen Audio 3.0 TTS handles up to 2,000 characters of script per request on Craze. Longer scripts need to be split and combined in your timeline. The model works best with punctuation-heavy scripts because it uses pauses to infer pacing. Run-on sentences sound rushed. Add periods, commas, and dashes to control timing. Ellipses (...) create pauses for dramatic effect.

For longer narration, split your script into requests of at most 2,000 characters and assemble the results in your timeline.