Qwen3-TTS Voice Design Text to Speech: Voices You Describe
Last verified: October 29, 2026
Qwen3-TTS Voice Design creates an entirely new voice from a written description instead of reading your script with a stock speaker. You type the lines into the Text box, say who should speak them in the Voice description field (age, timbre, pitch, pace, mood) and pick a Language. The model then reads the script in a voice that didn't exist until you described it. That makes Qwen3-TTS Voice Design text-to-speech as much about casting as about narration: you write the voice instead of choosing it.
The description controls more than timbre. The model takes natural-language instructions covering several acoustic attributes at once, including timbre, emotion and prosody. Your script, the description and the language are the only three things you give the generator, so rewording the description is the main way to change the result.
Use it when the voice is the creative problem: a cast of side characters for a game, a narrator who has to fit a brand instead of a catalogue, or a voiceover for a one-off promo where no stock speaker sounds right.
Explore more text to speech
Qwen3-TTS Voice Design at a Glance
What goes in, which languages you can set and what comes out of each run.
Input
A text prompt of up to 5,000 characters
Language
11 options
Output
1 audio per run
Writing Voice Descriptions and Scripts That Work
Describe What the Listener Hears, Not the Backstory
Let the Script's Punctuation Carry the Performance
How to Create Speech With a Designed Voice
Five steps take you from a written script to a downloadable audio file.
Write your script in the "Text" box
Type or paste the exact words to be spoken into the "Text" box. It is required and holds up to 5,000 characters, enough for a full explainer script or a scene of dialogue in one run. The "Enhance Prompt" button under the box suggests fuller versions of a rough idea, and nothing changes until you pick one. Use it while drafting, then leave it alone once the wording is final, because everything in the box gets read aloud.
Describe the speaker in "Voice description"
"Voice description" is required and takes up to 200 characters, about two short sentences. Use them on how the voice should sound rather than on the character's life story. The tips further down show a structure that works.
Choose the "Language"
The "Language" menu offers Auto, Chinese, English, French, German, Italian, Japanese, Korean, Portuguese, Russian and Spanish. Auto is the default and detects the language from your text. Auto is fine for a quick test, but select the script's language for the final render, because the Qwen team advises setting the language explicitly whenever you know it.
Click "Generate"
Make sure both "Text" and "Voice description" are filled in, then click "Generate". For a long script, try two or three lines first. A short test tells you whether the description works, and rewording it at that stage is quicker than redoing the whole piece.
Download the result
Each run returns one audio file. Listen to the whole thing before you download it. If the voice is close but not quite right, change one phrase in "Voice description" at a time and run again, so you can hear which change made the difference.
Designing a Voice vs Starting From a Preset Speaker
How Qwen3-TTS Voice Design compares with picking one of the Qwen3-TTS CustomVoice preset speakers.
| Criterion | Qwen3-TTS Voice Design | Alternatives | Best For |
|---|---|---|---|
| Where the voice comes from | Every voice is built from the description you write, so the speaker can be any voice you can put into words. | Qwen3-TTS CustomVoice works from nine preset voices that cover different combinations of gender, age, language and dialect. | If you need a voice that fits a specific character or brand, design it here. If one of the nine preset voices already sounds right, CustomVoice lets you start from it without writing a description. |
| Steering tone and emotion | Instruction control is built in, and the same voice description that defines the speaker can also set the mood of the read. | CustomVoice also takes written style instructions on top of a preset speaker, such as asking a named voice to sound especially angry. | Both accept written style instructions, so this won't decide it for you. Choose based on where the voice comes from, and expect emotional control either way. |
Common Voice Design Problems and Fixes
Two problems that come up most once a project grows past a single line.
The voice sounds different every time I generate with the same description
My script is too long to fit in one run
Frequently Asked Questions
Do I need a voice sample or reference audio to create a voice?
No. The model works from three inputs: the script in Text, the Voice description and the Language setting. If you have a real voice in mind, describe how it sounds rather than whose voice it is, because the description is all the model has to go on.
What makes one generation cost more than another?
Once you're signed in, the credit cost of a run is worked out from the options you've chosen and shown above the Generate button. It changes with prompt length, meaning how much script is in the Text box. Asking Enhance Prompt for suggestions doesn't use credits.
Should the voice description be in the same language as the script?
Qwen's own voice design examples pair them that way: a Chinese description for a Chinese line and an English description for an English line. If you're unsure, follow that pattern and write the description in the script's language.
How closely does Qwen3-TTS Voice Design follow a detailed voice description?
The Qwen team reports that on the InstructTTS-Eval benchmark, its voice design model beat MiniMax's closed-source voice design model on both instruction following and expressiveness, and clearly led the other open-source models. These are the developer's own results, so treat them as a guide and judge by how your own lines sound.
Can it read scripts full of numbers, abbreviations or messy formatting?
Qwen3-TTS is documented as much more robust to noisy input text, and its technical report lists handling complex texts among the tasks it supports. How symbols get spoken still isn't guaranteed. For a final render, spell out anything that could be read more than one way, for example writing "2.1%" as "two point one percent".