Introducing Audio8-TTS-Realistic, edge0(official)’s new open-weight speech model: realistic, zero-shot voice cloning in just 0.6B parameters. • Human-like prosody and expressive delivery • Clones a voice from one short clip, in 23 languages • 601M params: small enough to fine-tune and serve affordably Currently #1 on the Hugging Face TTS... Arena, where people vote blind on which voice sounds more human. Open weights are available now on Hugging Face. Try it free.show more

Audio8-TTS-Realistic is now open source. 🤗 Weights: 📄 Paper: 🎧 Samples: 💻 Code: ▶️ Playground:

How it learns to sound human: TTS is usually trained on two things: the right words (WER) and the right voice (similarity). But neither captures rhythm, pauses, or timing. So we trained a realism critic that scores each output against a real person saying the same sentence in the same voice—ranked via a blind listening study. Then: SFT → GRPO on accuracy + similarity → GRPO on the critic. That last stage only saw Chinese and English prompts, yet it improved all 13 lower-resource languages.

How it stacks up at 0.6B: • CV3 (9 languages): 4.86% error — ahead of CosyVoice3-1.5B (6.89%), Qwen3-TTS-1.7B (7.73%), and VoxCPM2 (5.35%) • MiniMax (24 languages): 5.89% error — vs. 27.41% for Qwen3-TTS-1.7B • Runs in ~1.1 GB RAM on an Apple M2 CPU (INT4) Larger 4–8B models still lead on some sets, but we’re within reach at a fraction of the size.

@Edge0_AI Is this just a weight update, or have the architecture and inference logic changed as well? I'm wondering whether audio.cpp's current Audio8 TTS Preview implementation can run the new model out of the box.

@Edge0_AI sub-1B local TTS hitting decent prosody matters way more than another 7B model nobody can afford to run locally

@Edge0_AI tf y'all trained the french side of the model on Emmanuel Macron's voice?? 😂😂😂

@Edge0_AI No voice direction? 😕

@Edge0_AI Looks very promissing

@Edge0_AI just added audio8 realistic to my tts-bench. sounds way better than the preview. samples and scores here:

@Edge0_AI 0.6B with open weights is a great size for on device voice. Small and steerable is where we see the most real use.

@Edge0_AI Excellent! Thank you!

@Edge0_AI Getting #1 on the TTS Arena with such a compact model is a solid achievement.

@Edge0_AI Congrats!

@Edge0_AI @Presidentlin

@Edge0_AI Cool demo. But 90% of agencies don't need voice cloning - they need a process that doesn't touch 4 tools.

@Edge0_AI 0.6B 就能做到零样本克隆，还支持 23 种语言，确实挺轻量的

@Edge0_AI Congratulations on the release 🎉🎉🎉

@Edge0_AI @grok как пойдет на 16 vram 32 ram?

@Edge0_AI Link to the TTS Arena you mention? Its paused so idk how you can claim first place.

@Edge0_AI how short is the short clip, like a voicemail greeting?

@Edge0_AI Wtf pourquoi Macron est la

@Edge0_AI why is the Japanese example in the video so low quality and horrible? I'm sure it can do better. It's as though its voice cloning a low quality tts.

@Edge0_AI there's a founder running a whole wake-word to tts pipeline on device on cheap hardware. couldn't get a dm through to you but i've got an intro worth taking.

@Edge0_AI Sorry but I knew it was B

@Edge0_AI Español como una de las lenguas más habladas, siempre son de los peores. Muy interesante la razón de no dar esta especial importancia.

@Edge0_AI the realism critic is what caught me: rhythm and pauses are what a word-error score misses, and it runs in about 1.1 GB on a CPU.

@Edge0_AI pourquoi c'est la voix à Macron?

@Edge0_AI It seems like it's extremely sensitive to the reference voice' quality. low quality inputs (which are most of the examples) results in high frequency noises in output. interestingly it can copy the audio noise in background too, which may be useful in some case.

@Edge0_AI this is great, i think i played with porting an earlier version of this to apple silicon before. i would suggest to use different voice samples, for example some of normal contents creation, vlog, review, podcast etc, instead of exclamations only.