Qwen3-TTS-VoiceDesign

Qwen3-TTS-VoiceDesign

Describe a voice. Get that voice. No recording required.

Design, Not Cloning

The distinction that defines this model, and it is worth being precise about.

Voice cloning needs a sample. You supply a recording, the model reproduces that speaker. It works well, and it requires a real person whose voice you have and are permitted to use.

Voice design needs a sentence. You describe the speaker — age, timbre, accent, emotional state, pacing — and the model synthesises someone matching the description. No reference audio exists at any point.

That changes what is possible in two directions.

Where no sample exists. A character in a game, a narrator for a brand, a voice for an interface — none of these have a real person behind them to record.

Where using a real voice is not appropriate. Consent, licensing, and likeness rights are questions that do not arise when nobody's voice was copied.

What You Can Describe

The control surface is a natural-language string. Community documentation names the dimensions that work:

Age. "80 years old", "in her 30s", "a young child".

Acoustic attributes. Hoarse, shrill, croaky, breathy, resonant, nasal.

Emotional register. Grumpy, warm, confident, frustrated, calm.

Pacing and emphasis. Deliberate, rapid, hesitant.

Accent and regional character.

A worked example from published usage, which shows how specific a description can get:

A very old, cranky, and croaky African-American grandma. 80 years old.
Very hoarse, grumpy, shrill, and frustrated.

And a simpler one:

A warm, confident female narrator in her 30s with a clear mid-range voice.

Both work, and the difference matters. The second produces a competent generic narrator. The first produces a character. Specificity is the lever — vague descriptions get you the middle of the distribution, which is a voice nobody remembers.

Architecture

The tokenizer is the foundation, and Qwen built it themselves. Operating at 12 Hz with multiple codebooks, it compresses speech efficiently while — in Qwen's description — fully preserving paralinguistic information and acoustic environmental features.

Paralinguistic information is everything in speech that is not the words: breath, hesitation, emphasis, emotional colour. A tokenizer that discards it produces correct speech that sounds wrong, which is the failure mode of most compression-first approaches.

Non-DiT is the other notable choice. Most current high-quality speech synthesis runs a diffusion transformer. This family uses a lightweight discrete language-model architecture instead — bypassing traditional information bottlenecks, per Qwen, and enabling high-speed, high-fidelity reconstruction without a diffusion stage in the path.

Benchmark Results

Qwen report that this model outperformed MiniMax-Voice-Design — a closed-source model — on the InstructTTS-Eval benchmark, in both instruction-following capability and generative expressiveness, while leading other open-source models significantly.

Two axes, and they pull apart. Instruction-following measures whether the voice matches what you described. Expressiveness measures whether it sounds like a person rather than a synthesiser. A model can do either without the other — a precisely-matched robot, or a lively voice that ignored half your brief.

Leading on both is the claim worth checking on your own descriptions.

Ten Languages

Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish, Italian — plus multiple dialectal voice profiles.

The dialect support is worth noting separately. A model that speaks a language and a model that speaks a regional variety of it are different capabilities, and the second is what makes synthesised speech sound local rather than translated.

Specifications

Capabilities

Using Qwen3-TTS-VoiceDesign on DEVUP AI

Endpoint: POST https://api.devupai.com/v1/audio/speech

The response is raw audio bytes, not JSON. Write response.content to a file directly.

Python

import requests

DEVUP_API_KEY = "$DEVUP_API_KEY"
MODEL = "Qwen/Qwen3-TTS-VoiceDesign"

response = requests.post(
    "https://api.devupai.com/v1/audio/speech",
    headers={
        "Authorization": f"Bearer {DEVUP_API_KEY}",
        "Content-Type": "application/json",
    },
    json={
        "model": MODEL,
        "input": "Hello! This is a text-to-speech example using DEVUP AI.",
    },
)

# Save the returned audio
with open("output.wav", "wb") as f:
    f.write(response.content)

cURL

curl -X POST "https://api.devupai.com/v1/audio/speech" \
  -H "Authorization: Bearer $DEVUP_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen3-TTS-VoiceDesign",
    "input": "Hello! This is a text-to-speech example using DEVUP AI."
  }' \
  --output output.wav

⚠️ The voice description is this model's defining parameter, and the shape above carries model and input only. How the description is passed on your path — an instruct field, a voice field, or something else — is worth confirming with a test request before you build.

Writing a Voice Description

The description is the product. Three habits separate a usable voice from a generic one.

Be specific about age. "Old" is a range spanning forty years. "80 years old" is a voice.

Stack acoustic attributes. Hoarse and shrill and croaky describe a particular throat. Any one alone describes a category.

Name the emotional state, not just the tone. "Grumpy and frustrated" produces different prosody from "calm and measured" even at identical timbre — and prosody is most of what makes speech sound human.

Match the description to the script. A frustrated voice reading cheerful copy sounds wrong in a way neither element explains. The two are one design decision.

And write the text for speech, not for reading. Shorter sentences, natural contractions, punctuation that marks breath rather than grammar. Text written for the eye reads badly aloud regardless of the voice.

Where It Fits

Character voices — games, animation, audio drama, where the speaker never existed.

Brand and product voices, designed to specification rather than cast from auditions.

Interface and assistant speech, tuned to a defined personality.

Localisation, with dialectal profiles across ten languages.

Anywhere consent and likeness are a concern. No real voice is copied, which removes a category of question that voice cloning raises.

Not for reproducing a specific person. That is voice cloning, and it is a sibling model in this family.

Not for reproducible output without verifying determinism on your path.

The Family

Qwen released this alongside several siblings at 0.6B and 1.7B, and they solve different problems.

The Instruct sibling is the one most easily confused with this model. It changes how a fixed voice speaks — style, emotion, emphasis — while keeping the timbre. This one creates the timbre. If you want a consistent speaker delivering lines differently, that is Instruct; if you want a new speaker, it is this.

A 0.6B variant exists for cases where the footprint matters more than the fidelity.

Practical Notes

Confirm how the voice description is passed on your path before building.

The response is audio bytes — write them directly, do not parse JSON.

Be specific. Vague descriptions produce forgettable voices.

Match the emotional register of the voice to the content of the script.

Write the script for speaking rather than for reading.

Test a description once and save the output if you like it — verify whether repeat calls reproduce it before assuming they do.

Use the Instruct sibling for style control on a fixed voice; use this one to create the voice.

Limitations

No reference audio means no exact target. You get a voice matching a description, not a specific voice you had in mind. Iteration is description refinement.

Ten languages. Others are outside the documented capability.

Audio output only. Text in, speech out — no transcription, no understanding.

Determinism is unverified. Whether the same description and text reproduce the same voice on repeat calls is worth testing rather than assuming, and it decides whether you can regenerate a character or must archive every clip.

Specific-person reproduction is a different model. This one designs; cloning clones.

1.7 billion parameters. Capable for its class, and not unlimited — very unusual vocal characteristics may sit outside what it produces.

Apply your own moderation. Synthetic speech that sounds human carries obvious misuse potential, and a designed voice can still be described in a way that approximates a real person.