Qwen3-TTS vs SuperClonica for local text-to-speech
Qwen3-TTS is the Qwen team's open speech model family, released in January 2026 in 0.6B and 1.7B sizes under Apache 2.0. It clones a voice from three seconds of audio, designs voices from a text description and streams audio with low latency. SuperClonica is narrower: ready-made and trained Supertonic 3 voices that run fast on a CPU.
Both run on your own hardware. They differ in speed, languages and in how you get a voice.
What Qwen3-TTS offers
The official examples run on an NVIDIA GPU in bfloat16, and FlashAttention 2 is recommended to reduce GPU memory use.
- 10 languages: Chinese, English, Japanese, Korean, German, French, Russian, Portuguese, Spanish and Italian, plus dialect voice profiles.
- 3-second voice cloning with the Base models, without a training step.
- Voice design from a natural-language description (the 1.7B VoiceDesign model).
- Instructions that control tone, speaking rate and emotion.
- Streaming generation; the authors report end-to-end latency as low as 97 ms.
Qwen3-TTS speed on CPU
Our estimate for the 0.6B model on a desktop 8-core CPU is 1–2× slower than real time: 10 seconds of speech in 10–20 seconds. It rests on a community engine written in C that reports a real-time factor of 2.02 for the 0.6B model with INT4 weights on a Ryzen 7 6800H laptop. The 1.7B model is slower; we put it at 2–5× slower than real time. Plain PyTorch can be considerably slower than that C engine.
On Apple Silicon the numbers favor Qwen more: the same C engine reports 0.52 on an M1, about twice as fast as real time.
SuperClonica on a desktop 8-core CPU runs 6.4–9.7× faster than real time by our estimate, so the same 10 seconds take about 1.0–1.6 seconds.
Quality, stability and languages
In our ratings Qwen3-TTS gets 4 out of 5 for quality and 4 for timbre stability; SuperClonica gets 4.5 and 5.
The language count favors Supertonic 3, 31 against 10, but the overlap is uneven. Qwen3-TTS speaks Chinese and Supertonic 3 doesn't. Supertonic 3 adds Arabic, Hindi, Turkish, Vietnamese, Ukrainian, Polish and most other European languages.
Which one to use
Both models are open and both run on your hardware, so the choice comes down to the job.
- Qwen3-TTS: you need Chinese, instant cloning or voice design, or you are building a voice assistant on a GPU and care about first-packet latency.
- SuperClonica: you generate a lot of speech on CPUs, work in one of the 31 Supertonic 3 languages, or want one trained voice to keep the same timbre across thousands of lines.