Edge0 Releases Audio8-TTS-Realistic, a 0.6B Voice Cloning TTS Model

Edge0 has released Audio8-TTS-Realistic, a roughly 0.6B-parameter text-to-speech model that clones a voice from a single reference clip and its transcript. It supports 23 languages, 12 more than its predecessor, and ships with weights and code under Apache 2.0, which permits commercial use.

The company's self-reported numbers put it ahead of larger open models on one multilingual benchmark, with a 4.86% average error across nine languages on CV3. The paper reports it running in about 1.1 to 1.2 GiB of memory on an Apple M2 CPU in a quantized build. Small, cheap, permissive, and realistic is a powerful combination, and it cuts both ways.

Let's start with what is genuinely good. Accessible speech in two dozen languages, at a size that is cheap to fine-tune and serve, is useful for screen readers, dubbing, language learning, customer service, and people who have lost their own voice. The Apache license means small teams can build on it without negotiating terms. We want that kind of tool to exist.

The technical idea we find most interesting is the push past the usual metrics. Edge0 argues that word error rate and speaker similarity miss rhythm, pauses, and timing, so a voice can score well and still sound fake. Its final training stage uses a learned realism critic that compares output to a matched human recording, plus guards against gaming it, including a penalty for stretching speech too long. Even though that stage used only Chinese and English prompts, it improved results on the 13 lower-resource languages too.

Now the harm, and I don't want to soften it. The explicit goal is speech that is harder to tell apart from a real person, cloned from one clip. The same capability that dubs a lecture can impersonate a parent on a scam call, fake a CEO's approval of a wire, or put words in a candidate's mouth. Open weights cannot be recalled or rate-limited. The release notes we read do not describe watermarking or consent checks, so we are treating those as absent until shown otherwise.

If this becomes normal, and it is heading that way, voice stops working as proof of identity. Banks that use voiceprints, families that trust a familiar voice, and newsrooms that verify audio all need to adjust now, not after the next viral fake. Our position: build with it, but pair it with consent records, disclosure, and provenance, and push phone and banking systems to stop treating a voice as a password.

One more honest note on quality. It is not uniformly strong. Thai error sits near 49% and Cantonese is weak, and 4B-class Higgs TTS 3 still leads on two benchmarks. Edge0 itself calls its comparison table a reference, since evaluation setups differ.

Audio8-TTS-Realistic uses a dual-autoregressive design inspired by Fish Audio S2 Pro, with a slow branch built on Qwen2.5 0.5B and a shallow fast branch predicting the remaining audio codebooks. It bundles a 44.1 kHz codec and can also generate speech without a reference voice. An SGLang Omni adapter offers an OpenAI-compatible speech endpoint, though Edge0 says throughput and streaming were not benchmarked.