Edge0 has published Audio8-TTS-Realistic, an open-weight speech synthesis model that can clone a voice from a single reference clip and its transcript. The model has roughly 0.6 billion parameters and supports 23 languages, twelve more than its predecessor. For developers, the appeal is the balance between size and accuracy: a model this compact is cheap to fine-tune and to serve.
Architecture and how it works
The system uses a dual-autoregressive design. A slower branch, built on a 0.5-billion-parameter Qwen2.5 base, produces text tokens and the first audio codebook at each step. A smaller four-layer branch then fills in the remaining nine codebooks in parallel, so sequential computation stays confined to the frame level and generation time stays contained.
Image: Edge0-AI — https://github.com/Edge0-AI/Audio8_TTS
Training for realism
The team argues that word error rate and speaker similarity miss rhythm, pauses and timing. Training therefore runs in three stages: multilingual supervised fine-tuning, reinforcement learning with a reward weighted 70% on word errors and 30% on voice similarity, and a final stage guided by a learned realism critic. The critic compares each output against a human recording of the same sentence and was trained on a quality scale built from six open-source TTS systems ranked in a blind listening study. A duration penalty stops the model from slowing speech down to fool the critic.
Reported results
The figures are self-reported by Edge0, which notes that its evaluation pipelines and text normalization differ from other studies. On the CV3 benchmark across nine languages, the model averages a 4.86% error, against 7.73% for Qwen3-TTS-1.7B and 5.35% for VoxCPM2. On Seed-TTS in English and Chinese, the average error is 1.25%. Weak spots remain: on Thai the word error rate approaches 49%, and on Cantonese the character error exceeds 31% on the MiniMax benchmark. Larger models, such as the roughly 4-billion-parameter Higgs TTS 3, still lead on CV3 and MiniMax.
Running it
The model installs with PyTorch 2.5 or later and Transformers 4.57 or later, and a CUDA GPU is recommended; the team tested on an NVIDIA H20 in BF16. The paper reports peak memory of 1.1 to 1.2 GiB on an Apple M2 CPU with a 4-bit quantized build, but no public ONNX export of this Realistic version exists yet. For production serving, Edge0 points to an engine with paged attention, dynamic batching and an OpenAI-compatible API, though it notes that throughput and streaming were not benchmarked. A free demo runs on Hugging Face ZeroGPU.
The release arrives as local voice cloning spreads fast, as shown by Kyutai's models with a few million parameters running on CPU and Qwen3-TTS arriving in llama.cpp. The same ease of reproducing a voice from seconds of audio, however, keeps the consent question open.
No comments yet. Be the first!