Speech models can score well on transcription accuracy and speaker similarity yet still sound noticeably synthetic. This release targets that gap with a compact model, while its reported intelligibility and similarity benchmarks should not be mistaken for a direct measurement of realism.
Key Capabilities
- A 601-million-parameter DualAR model predicts semantic and acoustic tokens in separate stages; the codec is bundled separately. This keeps the main model relatively small without removing the voice-cloning workflow.
- It supports zero-shot cloning from a reference recording and its transcript, as well as speech generation without a reference. That makes it useful for prototyping voices without first training a speaker-specific model.
- Coverage spans 23 languages, expanded from the Preview model's 11. However, the benchmark results vary widely across languages, so language support should not be read as uniform quality.
- The accompanying report describes a realism critic trained using matched human recordings and a quality ladder of synthesized speech. Its preference signal is relative to the same prompt, not an absolute naturalness score.
Who It Suits—and the Trade-offs
A good fit for researchers and developers who need a relatively compact multilingual TTS model, voice cloning, or a basis for testing speech-generation systems. Look elsewhere if consistent quality across all 23 languages, guaranteed pronunciation accuracy, or a validated objective score for perceived realism is essential. Noisy references and inaccurate transcripts can also reduce cloning stability; cloned voices should be used with permission and synthetic speech disclosed where appropriate.
Where It Fits
Compared with much larger speech-generation systems, its main point of distinction is the combination of a 0.6B core, multilingual support, and zero-shot cloning. Its reported error rates are competitive on some evaluated sets, but cross-model comparisons are not fully matched, and the paper's realism objective addresses a different question than intelligibility benchmarks.
