Qwen3-TTS: choosing CustomVoice, VoiceDesign or Base

Start with the material you have

Several entries in a model library can share the name Qwen3-TTS. Their size is only part of the choice. CustomVoice, VoiceDesign and Base take different inputs: a preset speaker, a description of a voice, or a reference recording.

For a first narration, start with CustomVoice. To explore a voice for a character, try VoiceDesign. When you have an authorized recording and want a similar voice to read new text, use the Base cloning route.

What each variant needs

The upstream cloning example uses reference audio with its transcript. It also offers a speaker-embedding-only mode. Use the inputs exposed by Sonano’s current interface; individual Python options upstream are not necessarily app controls.

Describe something you can hear

“Beautiful and engaging” leaves a lot open. Try a lower pitch, a steady pace, or sentences that settle rather than rise at the end. A useful first description might be:

An adult woman with a warm, clear voice, speaking at a moderate pace as if explaining a new idea to a friend. Natural pauses, without an advertising delivery.

Preview a short passage containing a statement, a question and a number. If the voice fits but a word sounds wrong, check the text first. If the character feels wrong throughout, revise the voice description. Change one kind of input at a time so you can hear what helped.

Use the voice across a project

In Sonano, save a designed or cloned voice to the Voice Library when you need to reuse it. Saved designed and cloned voices can be shared across Qwen3-TTS Base, VoxCPM2, OmniVoice and Chatterbox when the language matches. Kokoro and CustomVoice use their own presets. See the voice and model pairing rules.

For dialogue, assign voices to the characters and test a short exchange. For longer narration, check names, abbreviations and numbers before generating chapters. These small previews can catch selection and text problems before a long job.

Choose language and size with a short test

Qwen3-TTS lists ten languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. Coverage does not mean every speaker sounds equally natural in every language.

Upstream offers some variants at 0.6B and 1.7B; VoiceDesign is 1.7B. Check Sonano’s Model Library for the variants available in the app. Compare the same passage for delivery, waiting time and resource use on your Mac before choosing. Server benchmark speeds do not predict your local results.

Continue with the Qwen3-TTS model overview or the Voice Library documentation.