Qwen3 Voice Design: a new voice from one sentence of description, and where to run it
Apache 2.0, weights free, commercial use allowed. Self-hosting means the 1.7B Voice Design checkpoint in BF16 and Qwen's own Python package on a GPU, with FlashAttention 2 recommended to cut memory; Qwen publishes no memory figure. In your browser it runs on CNAPS Studio, included in the free plan, with no per-character cost or API key.
It invents the voice as well as reading the script. Write a sentence or two about who is speaking, such as age, tone, accent and pace, and it builds a voice to match that description before reading your text. A calm documentary narrator, a bright game-show host and a tired night-shift nurse all come from the same model.
The cost is consistency. The same description gives a similar but not identical voice on each run, and loudness varies with the voice, so the studio levels it to -16 LUFS by default. Four audio models sit in the CNAPS Studio catalog, and this is the one for a voice nobody has recorded.
What it actually is
The voice design model of Alibaba's Qwen3-TTS family, a 1.7-billion-parameter speech language model released in January 2026. It predicts discrete audio tokens from a 12.5 Hz tokenizer with 16 codebook layers, trained on more than 5 million hours of speech in ten languages, and conditions every token on your description.
Descriptions work best in English, even for a Korean script: in testing, a Korean description came out quieter and slower. One or two sentences are enough, and an empty description gives a fixed neutral adult voice.
In CNAPS Studio you connect text, and optionally a second text input that overrides the description setting. The description can run to 500 characters. The rest matches Qwen3-TTS: a language hint of Auto, Korean or English, speed from 0.5 to 2.0, up to 50 pronunciation rules, and loudness of off, -14, -16 or -20 LUFS.
Try Qwen3 Voice Design in CNAPS Studio →
The published numbers
Qwen scores voice design on InstructTTSEval in Chinese and English, across three tasks: matching exact acoustic parameters, following a descriptive style, and playing a role. Scores are percentages, higher is better.
Against eight other voice-design systems in the same table, open and commercial, it scored highest in five of the six columns. The exception was English acoustic parameters, where one commercial service scored 83.0.
The tests and the comparison come from Qwen's own report. Role-play is the hardest task for every system listed, so test character voices carefully before a long production.
How it compares to the other audio models here
These four do different jobs and share no test, so the table maps outputs. Between the two voice models, the question is whether you need a specific character or a dependable preset.
Reach for Qwen3 Voice Design for characters, ads and presenters with a particular sound. Reach for Qwen3-TTS for repeatable narration, MOSS for effects, and Text to Music (ACE-Step 1.5) for music.
What to chain it with
A script and a description in, a voice track out, so it usually feeds a video model.
The example in the LTX-2.5 Audio-to-Video post is the standard chain: Z-Image-Turbo designs a presenter, this model gives her a voice, and LTX-2.5 Audio-to-Video makes her speak. Gemma 4 31B can also turn a character brief into a one-line voice description before this step.
Open Qwen3 Voice Design in CNAPS Studio, read one line with the default description, then change a single detail, such as age or accent, and compare.
FAQ
What does Qwen3 Voice Design need to run locally?
The 1.7B Voice Design checkpoint, Qwen's qwen-tts Python package, Python 3.12 and a GPU, with FlashAttention 2 recommended to reduce memory in fp16 or bf16. Qwen publishes no memory figure or speed for this checkpoint. In CNAPS Studio it runs in the browser, included in the basic plan, with no per-character cost.
How do I keep the same voice across a series?
Reuse the description word for word, since any change in wording shifts the voice. Even then each run differs slightly, so record a series in as few runs as possible and listen to the first line of each. When a voice must never change, use a Qwen3-TTS preset instead.
Should I write the description in Korean for a Korean script?
Write it in English. The description only shapes the voice, and in testing a Korean description gave a quieter, slower read. Keep the script itself in Korean, set the language hint to Korean if names come out wrong, and describe accent, age, tone and pace in one or two English sentences.
