Edge0 has released Audio8-TTS-Realistic, an open-weight text-to-speech model capable of cloning a voice from a single reference recording and its transcript. With 601 million parameters for the main model, it supports 23 languages, 12 more than its predecessor. The weights and code are distributed under an Apache 2.0 license.
The model is based on Audio8-TTS-Preview-0.6b and retains the same Transformers inference interface. The repository includes the checkpoint, a 44.1 kHz neural audio codec, the tokenizer, and specific code. It can also generate speech without a reference voice. The new languages are Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, and Vietnamese.
A dual-branch architecture
The model relies on a so-called Dual AutoRegressive (DualAR) architecture, inspired by Fish Audio S2 Pro. According to the technical report published by Edge0, the slow branch is built on Qwen2.5 0.5B. It has 24 layers with a width of 896 and operates at approximately 21.5 frames per second. At each step, it predicts text tokens as well as the first audio codebook, i.e., the first of the code dictionaries that represent the sound.
A fast branch of 4 layers then predicts the remaining 9 codebooks in parallel. Sequential computation thus only occurs at the level of each audio frame. The codec, which reconstructs the 44.1 kHz waveform, is excluded from the 0.6 billion parameter count.
For training, Edge0 estimates that Word Error Rate (WER) and Speaker Similarity (SIM) do not account for rhythm, pauses, and timing. A model can achieve good scores on these two criteria and still sound artificial. Training takes place in three stages: a supervised multilingual fine-tuning, followed by two phases of GRPO, a reinforcement learning method. The final phase relies on a “realism critic” trained from six open-source text-to-speech systems, ranked during a blind listening test.
Edge0 indicates that this final step used only Chinese and English prompts, while improving both metrics for the 13 less-resourced languages. The error rate there drops from 22.53% to 20.77%, and SIM from 0.637 to 0.679.
Performance figures announced by the publisher
These figures are self-reported by Edge0, which notes that evaluation pipelines and text normalization differ from one model to another. On the CV3 benchmark, covering 9 languages, Audio8-TTS-Realistic shows an average error rate of 4.86%, ahead of VoxCPM2 (5.35%), CosyVoice3-1.5B (6.89%), and Qwen3-TTS-1.7B (7.73%). On Seed-TTS in English and Chinese, the average error is 1.25%. On MiniMax, which covers 24 languages, it reaches 5.89%, while Qwen3-TTS-1.7B scores 27.41%, a model that officially supports only 10 of these languages.
Larger models remain ahead on some tests. Higgs TTS 3 reports 4.41% on CV3 and 2.74% on MiniMax. The least-covered languages remain fragile: the error rate reaches 48.57% for Thai on CV3, and 31.35% for Cantonese on MiniMax.
For deployment, Edge0 recommends a CUDA GPU, with tests conducted on an NVIDIA H20 in BF16. The report mentions a peak memory of 1.1 to 1.2 GiB on an Apple M2 processor, in INT4 ONNX version, but no ONNX export of the Realistic version is yet public. A free Playground, running on Hugging Face ZeroGPU, allows testing the model.
