AI BriefingAudio8-TTS-Realistic-Air: Zero-shot TTS for 23 languages with 0.6B parameters
Edge0/Audio8-TTS-Realistic-Air
About the project
Audio8-TTS-Realistic-Air is a compact model with 610 million parameters that provides natural prosody and zero-shot voice cloning. It significantly improves the realism of prosody compared to the existing Preview version and bundles a 44.1kHz audio codec, allowing immediate use without a separate codec checkpoint.
It supports a total of 23 languages by adding 12 languages, including Arabic, Russian, and Vietnamese, to the 11 languages such as Chinese, English, Korean, and Japanese. It features a DualAR architecture inspired by Fish Audio S2 Pro, designed so that a slow AR transformer predicts semantic tokens while a fast AR transformer predicts the codec codebook.
It can clone a specific voice with just reference audio and text, and can synthesize without a reference. Through the SGLang Omni adapter, it supports paged attention, dynamic batching, and an OpenAI-compatible speech API, making it suitable for high-throughput GPU serving environments.
Released under the Apache 2.0 license, it can be loaded directly or called as a pipeline via the Transformers library. It has been verified with BF16 precision in CUDA-based GPU environments, making it easy for developers to experiment locally or integrate into services.
Edge0/Audio8-TTS-Realistic-Air
The original page has no description.
text-to-speech
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.