Audio8-TTS-Preview-0.1b: Zero-Shot Voice Cloning in 8 Languages with 170M Parameters

Provide only reference audio and text to clone the voice exactly and synthesize new sentences. It works immediately without any training process, using a zero-shot approach. It primarily supports Chinese and English, with experimental support for German, Spanish, French, Italian, Japanese, and Korean.

The main model has approximately 170M parameters and is much lighter than modern multilingual TTS systems, even when including a separate codec decoder (approximately 120M). Compared to large models such as Fish S2 Pro (4.6B) or MOSS-TTS (8.5B), it focuses on implementing practical zero-shot TTS by significantly reducing the footprint of language and audio models.

Based on the Audio8 Falcon H1 architecture, it adopts a dual autoregressive structure where the slow branch predicts semantic tokens and the fast branch predicts codec codebooks. An audio codec with a 44.1kHz sample rate is included in the repository, allowing it to run immediately without additional checkpoints.

It is suitable for developers who want to test multilingual speech synthesis in limited GPU environments or build voice cloning pipelines based on small models. It is loaded via the Transformers library and requires the trust_remote_code option because it uses custom code.