Edge0 Releases Audio8-TTS-Realistic, a 0.6B Voice Cloning TTS Model
​
Edge0 released Audio8-TTS-Realistic, an open-weight TTS model that clones a voice from 1 reference clip and its transcript. It supports 23 languages, 12 more than its predecessor. Weights and code ship under Apache 2.0.
What it does:
- 601M parameters (codec excluded), dual-autoregressive design with a Qwen2.5 0.5B slow branch and a 4-layer fast branch
- 3-stage training: multilingual SFT, GRPO on a WER/SIM reward, then GRPO guided by a learned realism critic that scores outputs against a real human recording of the same sentence
- 44.1 kHz codec, can also generate speech without a reference voice
- Transformers inference, CLI and SFT scripts on GitHub, SGLang Omni adapter with an OpenAI-compatible /v1/audio/speech API
Key numbers (self-reported, lower is better):
- CV3 (9 languages): 4.86% avg error vs VoxCPM2 5.35%, CosyVoice3-1.5B 6.89%, Qwen3-TTS-1.7B 7.73%
- Seed-TTS (EN/ZH): 1.25% avg error, SIM 0.704
- MiniMax (24 languages): 5.89% avg error, SIM 0.754
- Final RL stage used only Chinese and English prompts but cut error on 13 lower-resource languages from 22.53% to 20.77%
- Apple M2 CPU (INT4 ONNX, per the paper): 1.1 to 1.2 GiB peak memory
The catch: Higgs TTS 3 (~4B) still leads at 4.41% on CV3 and 2.74% on MiniMax. Lower-resource languages are weak, with 48.57% WER on Thai (CV3) and 31.35% CER on Cantonese (MiniMax). Throughput and streaming were not benchmarked, and no Realistic ONNX export is public yet.
Full analysis: marktechpost.com
Model: huggingface.co
Repo: Edge0-AI/Audio8_TTS
Paper: arxiv.org
29 points
Mentions
The SFT training code is also included in this version. Enjoy! github.com/Edge0-AI/Audio8_TTS/blob/master/audio8…
@aisearchio Thanks for the mention! For anyone interested, here's our github: Edge0-AI/Audio8_TTS And here's our discord: discord.com
@an_le28347 @SamuelZengML It does! Try now: Edge0-AI/Audio8_TTS
Edge0 Releases Audio8-TTS-Realistic, a 0.6B Voice Cloning TTS Model ​ Edge0 released Audio8-TTS-Realistic, an open-weight TTS model that clones a voice from 1 reference clip and its transcript. It supports 23 languages, 12 more than its predecessor. Weights and code ship under Apache 2.0. What it does: - 601M parameters (codec excluded), dual-autoregressive design with a Qwen2.5 0.5B slow branch and a 4-layer fast branch - 3-stage training: multilingual SFT, GRPO on a WER/SIM reward, then GRPO guided by a learned realism critic that scores outputs against a real human recording of the same sentence - 44.1 kHz codec, can also generate speech without a reference voice - Transformers inference, CLI and SFT scripts on GitHub, SGLang Omni adapter with an OpenAI-compatible /v1/audio/speech API Key numbers (self-reported, lower is better): - CV3 (9 languages): 4.86% avg error vs VoxCPM2 5.35%, CosyVoice3-1.5B 6.89%, Qwen3-TTS-1.7B 7.73% - Seed-TTS (EN/ZH): 1.25% avg error, SIM 0.704 - MiniMax (24 languages): 5.89% avg error, SIM 0.754 - Final RL stage used only Chinese and English prompts but cut error on 13 lower-resource languages from 22.53% to 20.77% - Apple M2 CPU (INT4 ONNX, per the paper): 1.1 to 1.2 GiB peak memory The catch: Higgs TTS 3 (~4B) still leads at 4.41% on CV3 and 2.74% on MiniMax. Lower-resource languages are weak, with 48.57% WER on Thai (CV3) and 31.35% CER on Cantonese (MiniMax). Throughput and streaming were not benchmarked, and no Realistic ONNX export is public yet. Full analysis: marktechpost.com Model: huggingface.co Repo: Edge0-AI/Audio8_TTS Paper: arxiv.org
@Raizo8a Yes, it supports Spanish! You can try it out for yourself: Edge0-AI/Audio8_TTS
Audio8-TTS-Realistic is now open source. 🤗 Weights: huggingface.co 📄 Paper: arxiv.org 🎧 Samples: edge0-ai.github.io 💻 Code: Edge0-AI/Audio8_TTS ▶️ Playground: huggingface.co
Audio8 TTS Preview 0.1B 0.1 B 的文本转语音模型 - 支持中文、英文、日语、粤语等共十一种语言 GitHub: github.com/Edge0-AI/Audio8_TTS/tree/master Hugging Face: huggingface.co huggingface.co