Natural-sounding speech is harder to achieve than intelligible speech: low recognition error rates do not guarantee convincing prosody. This release targets that gap while keeping the main model at 601 million parameters, with a separate bundled audio codec.
Key Capabilities
- It supports zero-shot voice cloning from a reference recording and its transcript, as well as generation without a reference. This makes it useful for adapting a voice to new text without training a speaker-specific model.
- Its DualAR design separates semantic-token prediction from codec-codebook prediction. The two-stage approach lets the model represent speech content and acoustic detail through distinct autoregressive components.
- Coverage spans 23 languages, up from 11 in the Preview release. Benchmark results are uneven: the newly added languages include notably higher error rates in some cases, so language support should not be read as equivalent quality across languages.
- Training includes multilingual supervised fine-tuning, optimization with intelligibility and speaker-similarity rewards, and a final stage guided by a realism critic. The reported benchmark metrics measure recognition error and speaker similarity, not perceived realism directly.
Who It Suits—and the Trade-Off
A good fit for researchers and developers who need a compact multilingual TTS model, reference-based voice cloning, or a practical baseline for speech-generation experiments. Look elsewhere if you need consistently high quality across all 23 languages, guaranteed artifact-free output, or verified high-throughput/streaming performance: the reported serving checks were runtime tests, not throughput benchmarks. Use reference voices with permission and disclose synthetic speech where appropriate.
