Qwen3 Text to Speech

Qwen3 TTS adapts tone, pace and emotion to what your script means. Create narration and voiceovers in Chinese, English, Japanese, Korean and 6 more languages.

Create Meaning-Aware Speech with Qwen3 TTS in 10 Languages

Last verified: November 3, 2026

Qwen3 TTS adapts tone, rhythm and emotional expression to the meaning of the text it reads . Whether you write in Mandarin, Japanese, Spanish or German, each line is delivered the way its words read. The series was built by the Qwen team at Alibaba Cloud , and its technical report was submitted on January 22, 2026 . It is the first text-to-speech model in the Qwen line. It pairs a dual-track language-model architecture with two speech tokenizers and was trained on more than 5 million hours of speech in its supported languages .

You direct it in ordinary language, not markup. The series takes natural-language instructions that steer timbre, emotion and prosody . Its description-based control can also invent an entirely new voice and give fine-grained control over the speech it outputs . A dedicated VoiceDesign checkpoint builds a voice from a user's written description . The models are also significantly more robust to noisy input text .

The ten languages are Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian, and the series also has multiple dialect voice profiles . You can run two versions here: Qwen3 TTS reads your script with one of nine preset speakers and an optional Style note, and Qwen3 TTS Voice Design creates the speaker from your own description.

Speech Generation Specs and Limits

What each run takes in, the voices you can use and what you get back.

Script length

Up to 5,000 characters per run

Preset speakers (Qwen3 TTS)

9 voices, with Serena selected by default

Voice description (Qwen3 TTS Voice Design)

Free text, up to 200 characters

Style direction (Qwen3 TTS)

Optional tone, pace or emotion, up to 200 characters

Language

Auto-detect or one of 10 languages

Output

One audio clip per run

Check These Before You Generate Speech

Six checks that keep accents natural, directions clear and designed voices consistent.

1

Match the speaker to the script's language

Qwen recommends using each preset speaker in its native language for the best quality, though every speaker can read any supported language . For Japanese copy, use Ono Anna. For Korean, use Sohee, and for American English, Aiden. For Mandarin, choose one of the Chinese speakers, such as Serena, Vivian or Uncle Fu.

2

Set Language when the script uses only one

Auto works out the language from your text. If you already know the language, select it, especially for short lines or copy full of foreign product names. That way the read doesn't depend on detection.

3

Give Style one clear direction

Style holds up to 200 characters. One focused cue, such as "warm, unhurried, quietly confident", steers the read more predictably than a list of competing moods. Keep stage directions out of Text so that box holds only words meant to be spoken.

4

Make the voice description specific

Describe what a casting director would: gender, age range, pitch, pace, texture, accent and attitude. Qwen's own design examples even script a delivery arc, such as a measured start that speeds up during an emotional outburst . Fit yours into 200 characters.

5

Keep each designed voice in one run

Qwen3 TTS Voice Design builds the speaker from your description on every run. If a character must sound identical throughout, put all of that character's lines in one Text box, up to 5,000 characters. For a multi-episode series, use a Qwen3 TTS preset so every episode has the same named speaker.

6

Check Enhance Prompt suggestions word for word

Enhance Prompt suggests fuller wording from a rough idea, and here that wording becomes the spoken script. Before you accept a suggestion, check that names, figures and any legally required lines are still exactly as written.

Choose your speech engine

Qwen3 TTS or ElevenLabs v3: Choose the Voice Workflow for Your Script

Both models turn a script into speech on Vidofy, but they take different inputs and cover different sets of languages. The first five rows show the controls each generator offers here. The last two show what each company documents about its model.

Feature/Spec Qwen3 TTS ElevenLabs v3
Script length per run Qwen3 TTS and Qwen3 TTS Voice Design: Text up to 5,000 characters ElevenLabs V3: Text up to 5,000 characters
Ready-made speakers Qwen3 TTS: 9 in the Voice menu (Aiden, Dylan, Eric, Ono Anna, Ryan, Serena, Sohee, Uncle Fu, Vivian), Serena by default ElevenLabs V3: 26 in the Voice menu, Rachel by default
Create a speaker from a written description Qwen3 TTS Voice Design: Voice description field, up to 200 characters The ElevenLabs V3 generator has no such setting
Language menu Qwen3 TTS and Qwen3 TTS Voice Design: Auto plus 10 languages ElevenLabs V3: Auto plus 71 languages
Results per run Qwen3 TTS and Qwen3 TTS Voice Design: one audio clip ElevenLabs V3: one audio clip
Languages the model covers Qwen3-TTS series: 10 languages plus multiple dialect voice profiles Eleven v3: 70+ languages
How delivery is directed Qwen3-TTS series: tone, rhythm and emotion adapt to plain-language instructions and to the meaning of the text Eleven v3: inline audio tags such as [sighs] or [laughs], plus punctuation, control pacing and delivery

Matching Each Generator to Your Voice Project

Casting: pick a preset or describe the voice

ElevenLabs V3 has the larger ready-made cast here, which helps when any good stock voice will do and you just want to audition quickly. Qwen3 TTS offers a smaller roster of nine described speakers. Qwen3 TTS Voice Design works differently: instead of browsing a list, you describe the speaker you need, including age, timbre and attitude, and the run creates it. For original characters, brand mascots or a narrator with a very particular texture, describing the voice gets you there directly. For a quick, conventional read, a big preset list is the easier start.

Language reach and how you direct the read

Language coverage is the clearest difference. Eleven v3 is documented for more than seventy languages and lists 71 in its menu here, so scripts in Hindi, Arabic, Polish or Swahili belong there. Within the ten languages Qwen3-TTS covers, the choice comes down to how you give direction. Qwen3-TTS takes direction as ordinary sentences kept apart from the script. Eleven v3 expects bracketed audio tags written into the script itself. If your copy has to stay clean for approval or translation, keeping directions outside the text is easier to manage. If you want a cue tied to one exact word, inline tags do that.

Which Generator Fits Your Script?

How to choose: Choose Qwen3 TTS Voice Design when you need to invent the speaker from a written description. It takes a Voice description, a setting the ElevenLabs V3 generator doesn't have. Choose ElevenLabs V3 when you want a bigger preset cast (26 voices, compared with 9 in Qwen3 TTS) or a language outside the ten that Qwen3-TTS covers: its menu lists 71, and Eleven v3 is documented for 70+. Script length and output are the same for both: up to 5,000 characters in and one audio clip out per run.

From Script to Spoken Audio in Five Steps

Five steps, from choosing a version to playing back your clip.

1

Choose a version in the model list

Pick Qwen3 TTS to use one of nine preset speakers, or Qwen3 TTS Voice Design to create the speaker from a written description.
2

Paste your script into Text

Enter up to 5,000 characters of exactly what should be spoken. Enhance Prompt, under the box, suggests fuller versions of a rough draft. Your text changes only if you pick one.
3

Pick a Voice or fill in Voice description

In Qwen3 TTS, open Voice and choose a speaker (Serena is selected by default). In Qwen3 TTS Voice Design, use Voice description (up to 200 characters) to set the speaker's age, timbre, accent and attitude.
4

Set Language and Style

Leave Language on Auto to detect it from the text, or choose one of the ten listed. With Qwen3 TTS, you can add a Style note of up to 200 characters for tone, pace or emotion.
5

Click Generate

Each run returns one audio clip. Listen back, change the script, speaker or direction if needed, and generate again until the read fits.

Frequently Asked Questions

What is Qwen3 TTS best at?

Qwen3 TTS fits delivery to the meaning of a script and to plain-language direction: Qwen3-TTS adapts tone, rhythm and emotional expression to both instructions and text semantics, across 10 languages . Qwen3 TTS Voice Design applies the same idea to the voice itself, building a speaker from a written description . That makes it a strong fit for directed narration, original character voices and voiceovers in Chinese, English, Japanese, Korean and six European languages.

What do the nine Qwen3 TTS voices sound like?

Qwen describes them like this: Vivian is a bright young female voice, Serena warm and gentle, Uncle Fu a seasoned male with a mellow timbre, Dylan a youthful Beijing male, Eric a lively Chengdu male, Ryan a dynamic, rhythmic male voice, Aiden a sunny American male, Ono Anna a playful Japanese female and Sohee a warm Korean female . Choose one from the Voice menu in Qwen3 TTS (Serena is selected by default). Dylan and Eric add a regional Chinese flavor that suits local Mandarin content.

Is Qwen3 TTS open source?

Yes. The Qwen team released both the speech tokenizers and the models under the Apache 2.0 license , with code on GitHub and checkpoints on Hugging Face. On this page, you work with it through the Text, Voice or Voice description, Language and Style controls.

Can Qwen3 TTS generate speech in real time?

The series uses a dual-track architecture designed for real-time synthesis, and its 12Hz speech tokenizer sends the first audio packet within 97 ms . On this page, each run turns your script into one audio clip that you can play back and use in your edit.

How does Qwen3 TTS perform on speech benchmarks?

In Qwen's technical report, the series reaches state-of-the-art results on objective and subjective benchmarks, including a multilingual TTS test set, InstructTTSEval and the team's own long-speech test set . InstructTTSEval matters most here, because it tests how closely speech follows natural-language instructions: the same kind of direction you write in Style or Voice description.

How is the cost of a Qwen3 TTS run shown?

Each version in the model list shows the credits a run starts from, and any visitor can see it. Once you're signed in, the cost of the options you've chosen is calculated and shown above the Generate button. For both Qwen3 TTS and Qwen3 TTS Voice Design, that cost changes with prompt length.