Edge0 has released Audio8-TTS-Realistic, an open-weight text-to-speech model that clones a voice from a single reference clip and its transcript — no fine-tuning, no dataset of the target speaker, just one sample. It’s a 601-million-parameter model (the audio codec is separate), supports 23 languages — twelve more than its predecessor, Audio8-TTS-Preview-0.6b — and ships under Apache 2.0, weights and code both.
What’s actually new over the preview release: the language list grows to add Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, and Swedish, among others, while keeping the same Transformers inference interface so existing integrations mostly carry over. The repository bundles the checkpoint, a 44.1 kHz neural audio codec, tokenizer, processor, and custom code — and it can generate speech with no reference voice at all, not just clone one.
The accuracy picture is genuinely mixed, and Edge0’s own numbers say so. On the CommonVoice 3 benchmark across nine languages, Audio8-TTS-Realistic averages a 4.86% error rate — ahead of Qwen3-TTS-1.7B’s 7.73%, despite being under half that model’s size. But it’s a 0.6B model competing with 4B-class systems on harder multilingual benchmarks, and it shows: 48.57% word-error-rate on Thai (CommonVoice 3) and 31.35% character-error-rate on Cantonese (MiniMax) are the model’s weak points, named explicitly in its own documentation rather than buried.
Hardware needs are modest. A CUDA GPU is recommended, and Edge0 smoke-tested it on an NVIDIA H20 in BF16. For edge deployment, the model’s paper reports 1.1 to 1.2 GiB peak memory on an Apple M2 CPU using an INT4 ONNX export — though that export covers the earlier Preview model; no Realistic-specific ONNX build is public yet, so anyone planning CPU-only deployment should test before committing.
Who this is actually for: anyone building narration, voiceover, or dubbing tooling who needs multilingual coverage without the latency or cost of a bigger model, and can tolerate (or route around) weak spots on specific languages — Thai and Cantonese being the clearest ones flagged. It’s the second 0.6B release in the Audio8 TTS family, and the one-clip cloning plus Apache 2.0 licensing make it a realistic base for a self-hosted narration pipeline rather than a research curiosity.
Related Reading
- Edge0 Releases Audio8-TTS-Realistic, a 0.6B Voice Cloning TTS Model — MarkTechPost
- Audio8-TTS-Realistic-0.6b — Hugging Face
- Audio8 Transcribes Forever at Constant Memory — our earlier coverage of Edge0’s ASR model
- Getting Started With Local Text-to-Speech for Narration and Voiceover — our tutorial
- Getting Started: Cloning Your Own Voice With AI — our tutorial