Edge0 has released Audio8-TTS-Realistic, an open-weight text-to-speech model featuring zero-shot voice cloning capabilities in a compact 0.6B-parameter size. The model requires only a single reference clip and transcript to clone voices across 23 supported languages, marking an expansion of 12 new languages compared to its predecessor. Distributed under the Apache 2.0 license, the package includes model checkpoints, a 44.1 kHz neural audio codec, tokenizer, and processor designed for affordable fine-tuning and deployment.
Built on a dual-autoregressive design inspired by Fish Audio S2 Pro, the architecture uses a slow branch powered by Qwen2.5 0.5B running at 21.5 frames per second, alongside a shallow 4-layer fast branch that predicts remaining audio codebooks in parallel. To improve output quality beyond traditional word error rate and speaker similarity metrics, Edge0 implemented a three-stage training process involving supervised fine-tuning and GRPO guided by a learned realism critic. This critic integrates a quality ladder and identity-erasure loss to penalize synthetic fingerprints and unnatural pacing.
On benchmark evaluations, Audio8-TTS-Realistic achieved a 4.86% average error rate on CV3 across nine languages, outperforming several larger open models such as Qwen3-TTS-1.7B and CosyVoice3-1.5B. Developers can run the model locally using PyTorch and Transformers, with NVIDIA CUDA GPUs recommended for optimal performance alongside low-memory options for Apple silicon. Why it matters: This release lowers the resource barrier for deploying multilingual, high-fidelity voice cloning applications by packing competitive accuracy into an efficient 0.6B architecture.