Introduction

Edge0 just

, a small text-to-speech model that clones a voice from one short reference clip. It has 0.6 billion parameters, speaks 23 languages, and the weights are on Hugging Face under Apache 2.0, so you can use it commercially and fine-tune it.

The launch post says it's currently #1 on the Hugging Face TTS Arena, the leaderboard where people vote blind on which voice sounds more human. Edge0's own model card is more modest and says it has "received encouraging ratings," so treat the arena spot as a moving target. The part I care about is simpler: a voice cloner this small, this permissive and this multilingual is rare.

What Shipped

From the model card and GitHub repo:

  • Model: Audio8-TTS-Realistic-0.6b, the follow-up to Audio8-TTS-Preview, tuned for more natural prosody and expressive delivery.
  • Size: 601 million parameters in the main model. The 44.1 kHz audio codec ships in the same repo and isn't counted in that number.
  • Languages: 23, up from 11 in Preview. New this time are Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish and Vietnamese.
  • Modes: zero-shot voice cloning from a reference recording plus its exact transcript, or plain speech generation with no reference at all.
  • Architecture: a two-stage "DualAR" design borrowed from Fish Audio S2 Pro. One transformer predicts a sound token per audio frame, and a smaller one fills in the codec detail for that frame.
  • License: Apache 2.0 for both the code and the weights.
  • Serving: an SGLang Omni adapter with batching and an OpenAI-compatible /v1/audio/speech endpoint.
  • Paper: Training a Compact, Strong, and Realistic Text-to-Speech Model on arXiv.

The Numbers

Edge0 reports error rates (lower is better) and speaker similarity on a 0 to 1 scale (higher is better). These measure how clearly the words come through and how close the clone sounds to the reference, not how "real" it feels.

  • Seed-TTS, English and Chinese: 1.25% error and 0.704 similarity, about the same clarity as Preview but a closer clone (Preview scored 0.682).
  • CV3, 9 core languages: 4.86% error, ahead of CosyVoice3 (6.89%), VoxCPM2 (5.35%) and Qwen3-TTS-1.7B (7.73%) in Edge0's comparison table.
  • MiniMax suite, 24 languages: 5.89% error and 0.754 similarity. Bigger models like Higgs Audio v3 (2.74%) and Fish S2 Pro (4.40%) still win here.

So it isn't beating 4B+ models everywhere, but it lands in the same neighbourhood at a fraction of the size. Edge0 also notes these comparisons mix different evaluation pipelines, so read them as a rough guide.

Quality isn't even across languages, and the card is honest about it. Thai, Finnish and Vietnamese show high error rates on the CV3 set, and Cantonese struggles on the MiniMax set. English, Chinese, German, Spanish and Portuguese look strongest.

How to Try It Today

The fastest way in is the free Hugging Face playground. It runs on ZeroGPU, so expect a queue at busy times. Edge0 also has a listening demo page if you just want to hear samples first.

To run it locally, Edge0 recommends Python 3.10+ and a CUDA GPU:

pip install "torch>=2.5.0" "torchaudio>=2.5.0" \
  "transformers>=4.57.0,<5" "soundfile>=0.12" "safetensors>=0.4"

Then load it through Transformers with your reference clip and its transcript:

from transformers import AutoModel, AutoProcessor

model_id = "Edge0/Audio8-TTS-Realistic-0.6b"
processor = AutoProcessor.from_pretrained(model_id, trust_remote_code=True)
model = AutoModel.from_pretrained(model_id, trust_remote_code=True).eval().to("cuda")

inputs = processor(
    text=["Welcome back, traveler. The night market opens at dusk."],
    reference_audio=["my_voice.wav"],
    reference_text=["The exact words spoken in my_voice.wav."],
    return_tensors="pt",
)

From there, call model.generate(...) and model.decode_audio(...) as shown on the model card, and write the waveform out at the codec's 44.1 kHz rate. The model uses custom code, so look over the repo files before you set trust_remote_code=True.

Pro Tip

Your reference transcript has to match the clip word for word. A clean 5 to 10 second recording with an exact transcript beats a long, noisy one every time. The reference, its transcript and your target line share one context budget, so split long scripts into sentence-sized chunks and stitch the audio afterwards.

Why It Matters for Audio Creators

Last week we covered Tesslate's Lure TTS, another open 0.6B cloner. Audio8 pushes that same small-model idea much wider on languages and pairs it with a ready-made OpenAI-style server. That makes it a practical drop-in for indie games, narrated shorts, podcasts and dubbing tests where you want your own voices on your own hardware without per-character fees.

The license matters too. Apache 2.0 means you can ship it inside a product, which isn't true for a lot of open-weight voice models. Just remember Edge0's own rule: only clone voices you have permission to use, and label synthetic audio where it counts.

Conclusion

Audio8-TTS-Realistic is Edge0's new open-weight, Apache-2.0 text-to-speech model: 0.6B parameters, zero-shot voice cloning from a short clip, and 23 languages. You can try it free in the Hugging Face playground today, or run it locally with Transformers. Start with English or Spanish to hear it at its best, then test your target language before you commit.

—Anabel ♡