The Audio8 Falcon H1 architecture uses slow and fast autoregressive branches: the slow branch predicts semantic tokens, while the fast branch predicts codec codebooks conditioned on the slow hidden state. The bundled codec, tokenizer, processor, and remote code support Hugging Face Transformers inference and CPU deployment through ONNX INT8.
Chinese and English are the primary languages, with experimental evaluation in German, Spanish, French, Italian, Japanese, and Korean. Audio8 is useful for lightweight narration, voice prototyping, localization, and privacy-conscious local speech workflows, subject to consent and license requirements for voice cloning.
