When to choose Qwen3-TTS
Choose CustomVoice for preset speakers in multiple languages. Check the conversion name carefully: Base and VoiceDesign are different models.
- Model variant
- CustomVoice
- Model sizes upstream
- 0.6B and 1.7B
- Preset voices
- 9
- Languages upstream
- 10
- Model licence
- Apache 2.0
Where shown, download sizes are rounded from the app’s pinned packages. They exclude the shared speech engine and do not measure runtime memory.
CustomVoice is the version to look for
Qwen publishes CustomVoice, Base and VoiceDesign variants. Outmake’s voice browser offers compatible CustomVoice conversions and excludes Base and VoiceDesign.
CustomVoice uses preset speakers. A page or demo describing voice cloning with Base, or voice creation with VoiceDesign, does not establish that those features are available in Outmake.
Upstream CustomVoice comes in 0.6B and 1.7B sizes. The complete MLX download depends on the conversion and includes more than the main model weights. Read its repository before downloading.
Ten supported languages
The upstream model card lists Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian. It provides nine preset voice timbres.
The list describes language support, not equal pronunciation quality for every voice and text. Test the speaker with your own names, numbers and sentences. The official Qwen demo is linked below if you want to hear examples before installing.
Use the voice controls available in the app
The upstream 1.7B CustomVoice release documents instruction control. The smaller model’s feature table differs. Outmake’s available controls also depend on its speech runtime.
This guide does not promise every control from Qwen’s Python examples in the app. In Models, open Voices, choose a compatible CustomVoice conversion and inspect its speaker list after downloading.
Run Qwen3-TTS in Outmake
Outmake is free and requires macOS 26 or later on Apple silicon. These voice models run locally after their files have downloaded. You can also follow the upstream project links below to use another runtime.
- Open Models, then Voices.
- In the Hugging Face section, find a Qwen3-TTS CustomVoice conversion. Read its model information and download size.
- Choose Get. Once downloaded, play a sentence in the voice’s page.
- Choose Use to make it the local voice. To hear a reply, open its menu and choose Read Aloud.
Voice settings and read-aloud help explain how to choose voices and stop playback. This page is about generating speech. For speech-to-text, use dictation or transcription instead.
Before you rely on it
- The ten-language specification comes from Qwen’s upstream model card. No ten-language evaluation was performed in Outmake for this guide.
- Do not assume a Qwen chat model can speak: Qwen3-TTS is a separate speech model download.
Try a short passage before committing to a long one. Local speech keeps synthesis on your Mac; it does not make an online chat model local. Outmake’s privacy policy explains the difference.
Sources
Model cards and runtime documentation checked 11 November 2026. App-specific details were checked against Outmake’s voice catalog and runtime configuration. No new listening test or speed benchmark was performed for this guide.