Edge0 has released Audio8-TTS-Realistic, an open-weight text-to-speech model for zero-shot voice cloning. The 0.6B-parameter model clones a voice from a single reference clip and transcript, supports 23 languages — 12 more than its predecessor — and ships its weights and code under Apache 2.0.
Audio8-TTS-Realistic (Credits: MarkTechPost)
Audio8-TTS-Realistic: Tiny Size, Competitive Multilingual Accuracy
The catch for developers: competitive multilingual accuracy at a size that is cheap to fine-tune and serve. The model has 601M main-model parameters, handles up to 2,048 packed text/audio positions of context, and is smoke-tested on an NVIDIA H20 in BF16, with 1.1–1.2 GiB peak memory on an Apple M2 CPU (INT4 ONNX).
Performance is ahead of larger open models on the CV3 benchmark, with a 4.86% average CV3 error across 9 languages versus 7.73% for Qwen3-TTS-1.7B, though it trails 4B-class models on MiniMax. Thai and Cantonese remain its weakest languages. New languages include Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish and Vietnamese.
DualAR Architecture And GRPO Trained On Human-Like Realism
The model uses a dual-autoregressive (DualAR) design inspired by Fish Audio S2 Pro. The slow branch, built on Qwen2.5 0.5B, predicts text tokens plus the first audio codebook, while a shallow 4-layer fast branch predicts the remaining 9 codebooks in parallel.
The team argues that word error rate and speaker similarity miss rhythm, pauses and timing — a model can score well on both and still sound synthetic. So the final training stage uses GRPO driven by a learned realism critic: 0.7 weight on WER, 0.3 on speaker similarity, with the critic trained on a quality ladder built from 6 open-source TTS systems ranked in a private blind listening study.
It is the second 0.6B release in Edge0’s Audio8 TTS family, building on Audio8-TTS-Preview-0.6b while keeping the same Transformers inference interface. The repository bundles the checkpoint, a 44.1 kHz neural audio codec, tokenizer and processor, and it can also generate speech without a reference voice.
Also Read: Granola Adds Free Unlimited Dictation to Take On Wispr — Voice AI Apps Are All Converging | Mistral Releases Weights for Trillion-Parameter ‘Le Chonk’ AI Model Targeting Cybersecurity
Source: MarkTechPost — read the full story there.
