Audio8-TTS-Realistic-0.6b: Zero-shot voice cloning across 23 languages with 0.6B parameters

Generates synthetic speech that reproduces a specific person's voice using only text and a short reference recording. It operates in a zero-shot manner without a separate training process and can convert text to speech using a default voice even without a reference recording. This is an improved version based on Audio8-TTS-Preview-0.6b, featuring enhanced realistic pronunciation and intonation.

The model size is 610 million parameters, making it compact, and it is provided with an audio codec at a 44.1kHz sample rate. It supports a total of 23 languages, adding 12 languages such as Arabic, Russian, and Vietnamese to the existing 11 languages including Korean, English, Japanese, and Chinese. It uses the DualAR architecture inspired by Fish Audio S2 Pro, where a slow AR transformer predicts semantic tokens and a fast AR transformer predicts the codec codebook.

Inference can be performed locally using custom code from the Transformers library. In environments requiring batch processing, GPU serving is possible via the SGLang Omni adapter, which applies paged attention and dynamic batching through an OpenAI-compatible API. It is distributed under the Apache 2.0 license, facilitating commercial use.

Suitable for developers needing high-quality TTS with a lightweight model, or content creators who must quickly replicate a specific speaker's voice. Its broad multilingual support makes it useful for building voice synthesis pipelines for global services.