Qwen3-TTS: narration in nine preset voices with pronunciation rules, and where to run it
Apache 2.0, weights free, commercial use allowed. Self-hosting means Qwen's own Python package on a GPU, with Python 3.12 and FlashAttention 2 recommended to cut memory; Qwen publishes no memory figure, and the family comes in 0.6B and 1.7B sizes. In your browser it runs on cnaps studio, included in the basic plan, with no per-character cost or API key.
It reads a script aloud in one of nine preset voices, each with a native language: Sohee for Korean, Ryan and Aiden for English, Ono Anna for Japanese, and five Chinese voices including two regional dialects. Pronunciation rules fix brand names and acronyms, and a loudness setting levels the result for a platform.
The cost is range. The voices are fixed, the studio's language hint covers only Korean and English, and a voice reading outside its native language keeps an accent. Four audio models sit in the cnaps studio catalog, and this is the one for quick narration in a ready-made voice.
What it actually is
The preset-voice model of Alibaba's Qwen3-TTS family, released in January 2026. It is a speech language model that 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. It streams, and Qwen's own serving tests put the first audio about a tenth of a second after the request.
Default picks Sohee for Hangul text and Ryan for English, so most scripts need no setting at all. Speed is changed after synthesis with the pitch kept, so a faster read does not raise the voice.
In cnaps studio you connect text. Settings cover the voice, a language hint of Auto, Korean or English, speed from 0.5 to 2.0, up to 50 pronunciation rules in the form written=spoken, and loudness of off, -14, -16 or -20 LUFS. Loudness normalization runs in one pass with the true peak held at -1.5 dBTP.
Try Qwen3-TTS in cnaps studio →
The published numbers
Qwen scores the 1.7B preset-voice checkpoint by word error rate: a speech recognizer transcribes the generated audio and the result is compared with the script. Lower is better, and the figures are percentages.
Korean and English are both under 2%, meaning fewer than two words in a hundred came back wrong. Japanese is the weak spot of the four, and long passages of 200 to 2,000 words stay under 3%.
All of it is Qwen's own testing on its 1.7B checkpoint, and the 0.6B version scores differently. Listen to a long script in your own language before relying on it.
How it compares to the other audio models here
These four do different jobs and share no test, so the table maps outputs. Among the two voice models, the question is whether a preset voice is enough.
Reach for Qwen3-TTS for fast, repeatable narration in Korean or English. Reach for Qwen3 Voice Design when you need a specific character, MOSS for effects, and Text to Music (ACE-Step 1.5) for music.
What to chain it with
Text in, a voice track out, so it usually feeds a video model or a mix.
Two chains suit it. Before it, Gemma 4 31B can tighten a script to the length of a clip. After it, the voice can drive LTX-2.5 Audio-to-Video from a still portrait, or re-voice existing footage through LatentSync 1.6, with an ACE-Step score mixed underneath.
Open Qwen3-TTS in cnaps studio, read one line with the Default voice, then add a pronunciation rule for your brand name and compare.
FAQ
What does Qwen3-TTS need to run locally?
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 for either the 0.6B or 1.7B checkpoint. In cnaps studio it runs in the browser, included in the basic plan, with no per-character cost or API key.
How do I fix a mispronounced name?
Add a pronunciation rule in the form written=spoken, such as cnaps=see naps, one per line or separated by semicolons. Matching ignores case and works on whole words, Korean particles after a match still count, and the longest form wins. Up to 50 rules fit, and a malformed rule stops the run.
Which voice should I use for Korean?
Leave the voice on Default, which picks Sohee, the native Korean voice, whenever the text is in Hangul. The other eight voices can read Korean but keep their own accent. For a Korean voice with a different age, tone or gender, describe it in Qwen3 Voice Design instead.
