Why size matters now: making zero-shot voice cloning practical outside huge server farms requires trading parameter scale for portability and cost. This preview checkpoint demonstrates that a compact generative stack (~170M main + ~120M codec) can still produce usable cloned speech in realistic settings—especially for Chinese and English—while keeping inference and deployment demands far lower than multi-billion-parameter systems.
Key Capabilities
- Zero-shot voice cloning: accepts a short reference audio plus its transcript to condition speech synthesis, producing voice-preserving outputs without fine-tuning. This enables quick prototype cloning or personalized TTS where collecting large corpora is impractical.
- Compact dual-branch design: a slow autoregressive branch predicts semantics and a fast branch predicts codec codebooks; this architecture balances expressiveness and generation speed at small scale, with 10 acoustic codebooks and up to 2,048-context positions.
- Bundled neural codec: includes an integrated codec decoder (44.1 kHz) so inference needs no separate model download, simplifying deployment for experiments and demos.
- Multilingual but targeted: optimized for Chinese and English; German/Spanish/French/Italian/Japanese/Korean are supported experimentally but quality is more variable.
Who it's for and trade-offs
Great fit if you need a low-parameter, zero-shot TTS baseline for research, demos, edge or cost-sensitive deployment, or rapid voice-prototyping where obtaining large adaptation datasets is hard. Look elsewhere if you require state-of-the-art multi-language fidelity or maximum similarity for challenging references—larger checkpoints (0.6B+) and multi-billion models produce higher similarity and lower error rates. Also, noisy/long/mistranscribed reference clips reduce stability and speaker similarity, so data quality matters.
