Qwen3-TTS Online

Qwen3-TTS is a speech generation model family developed by Alibaba’s Qwen team. The official releases include 0.6B and 1.7B variants and cover ten languages. Base supports cloning from reference audio; CustomVoice provides preset voices; VoiceDesign creates voices from text descriptions. These names describe different model capabilities, rather than three steps every user must complete.

qwentts.app is an independent browser workspace, with no official affiliation to the Qwen team. It offers preset voices and reference-based cloning. VoiceDesign and a style-instruction control are not exposed in the current workspace. Use the official sources below for model details and the visible workspace controls for your browser task.

Start with a preset when you need narration and have no particular speaker to match. This removes reference preparation from the first trial and lets you focus on whether the text sounds clear. Choose a sentence with the names, numbers, or phrasing that matter in your actual project.

Choose reference-based cloning when the speaker’s vocal identity matters. Prepare a recording with one clear speaker and little background interference, using audio you have permission to use. The cloning interface accepts a reference up to 2 MB. Judge resemblance and intelligibility separately: a recognizable tone can still come with an awkward word or pause.

The preset workspace language selector offers English, Chinese, Spanish, Japanese, Korean, German, French, Russian, Portuguese, Italian, and auto. For a script written mainly in one language, select it explicitly so you can keep that choice consistent during evaluation. Auto is another selectable option; listen carefully when your sentence mixes languages.

Write text as you want it spoken. Expand an abbreviation if the full phrase is intended, and decide whether a number should sound like a date, an amount, or individual digits. Use ordinary punctuation to clarify sentence boundaries. Keep pronunciation experiments short so that one difficult term does not obscure the rest of your assessment.

Public recorded demos need no login. Listen first for an example relevant to your language or use case, remembering that a recording is an illustration rather than a prediction of your script. Generating your own text requires login. A signed-in free account has one generation per day with up to 120 characters per request; paid requests allow up to 1,000 characters.

Prepare the trial before generating: choose the mode, finalize a short passage, and check the selected voice or reference. After generation, listen once for the meaning and again for trouble spots. Download the WAV when you want to inspect it in your editing workflow. Note the text and settings yourself to support later comparisons.

Use the same sentence when comparing voices, then change only the voice. When testing punctuation, retain the voice and language. This makes it easier to connect a listening difference to a decision. Compare the opening, a difficult phrase, and the ending, rather than judging only the first few words.

• Check whether every word is understandable without reading along.
• Listen for pauses that split a name or separate closely related words.
• Consider whether the delivery suits the surrounding video, lesson, or narration.

Stay in the browser for a short audition and a downloadable result. Local setup is a separate technical project: you manage model files, dependencies, execution, and updates. Consult the official repository and model card if controlling the runtime or exploring a different released variant is your objective. Browser availability does not mean every upstream feature has a matching control here.

Use the demos page for listening, the free guide to plan an allowance-limited trial, and the cloning guide for reference preparation. Return to the workspace when the text and voice approach are ready. For longer scripts, plan meaningful passage boundaries before working within the request limit.