Edge0 launches Audio8-TTS-Realistic, a 0.6B-Parameter voice cloning model

Edge0 has launched Audio8-TTS-Realistic, a 0.6B-parameter text-to-speech model capable of zero-shot voice cloning. The model supports 23 languages, including Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, and Vietnamese. It uses a dual-autoregressive design inspired by Fish Audio S2 Pro, with a slow branch based on Qwen2.5 0.5B and a fast branch with 4 layers. The model runs at about 21.5 frames per second and can generate speech without a reference voice. Training involves three stages, with a critic evaluating outputs against human recordings. The model improved error rates on 13 lower-resource languages, reducing word error rate from 22.53% to 20.77% and increasing speaker similarity from 0.637 to 0.679. Audio8-TTS-Realistic outperformed several competitors on benchmark tests, including VoxCPM2, CosyVoice3-1.5B, and Qwen3-TTS-1.7B. It is available under Apache 2.0, with code and checkpoints provided in a GitHub repository.