Audio8-TTS-Realistic: clone any voice with a 0.6B model

Audio8-TTS-Realistic is an open-weight text-to-speech model that clones any voice from just a few seconds of clean reference audio and an exact transcript.

Released by Edge0 under the permissive Apache 2.0 license, the model packs 601M main parameters and weighs only 2.6GB. It supports 23 languages, adding 12 new languages including Arabic, Hindi, Russian, and Vietnamese.

How it works under the hood

Instead of burning compute with a massive architecture, Audio8 uses a dual-autoregressive design inspired by Fish Audio S2 Pro.

The slow branch runs on Qwen2.5 0.5B at around 21.5 frames per second, predicting text tokens and the first audio codebook. A shallow 4-layer fast branch then fills in the remaining nine codebooks in parallel. The repo bundles a 44.1 kHz neural audio codec and can also generate speech from scratch without reference audio.

To make voices sound human rather than robotic, Edge0 trained the model with GRPO guided by a learned realism critic. The critic scores generations against actual human recordings on a quality ladder, penalizing rushed syllables and unnatural pauses.

Benchmarks and deployment

On the 9-language CV3 benchmark, Audio8 hits a 4.86% average error rate, outperforming larger rivals like Qwen3-TTS-1.7B (7.73%) and VoxCPM2 (5.35%).

It still lags behind 4B-class models like Higgs TTS 3 on complex benchmarks, and performance degrades on low-resource languages, reaching a 48.57% word error rate on Thai.

Running it locally is remarkably cheap. Any modern consumer CUDA GPU handles it easily in BF16. Research tests showed INT4 ONNX peaking at just 1.1 to 1.2 GiB of memory on an Apple M2 CPU, though that export is not public yet.

You can test it right now on a free Hugging Face ZeroGPU playground. Local deployment requires PyTorch 2.5+ and Transformers 4.57+, with an SGLang Omni adapter ready for OpenAI-compatible API serving.

Why it matters

For indie hackers and developers building voice agents, Audio8-TTS-Realistic hits an unusually sweet spot. Getting solid zero-shot voice cloning typically requires renting beefy 4B-parameter setups, but this model delivers production-grade English and Chinese prosody on consumer hardware under a commercial-friendly license.

Edge0 showed that training against human rhythm matters more than raw parameter bloat. You will want to watch out for accuracy drops in rarer languages, but for mainstream voice cloning on a budget, it is easily one of the best open tools available right now.