Apache-2.0audioRuns locally0.6B (601,159,424; codec excluded)
Audio8-TTS-Realistic-0.6B: Specs, License & How to Run
Edge0 Apache-2.0 0.6B DualAR TTS. 23 languages, zero-shot cloning. Not Qwen3-TTS.
- License
- Apache-2.0
- Parameters
- 0.6B (601,159,424; codec excluded)
- HF likes
- 207 (2026-11-02)
- First seen
- 2026-10-14
TL;DR. Audio8-TTS-Realistic-0.6B is a audio model from Edge0. Parameters: 0.6B (601,159,424; codec excluded).License: Apache-2.0. It is listed under ElevenLabs.
| Developer | Edge0 |
|---|---|
| Type | audio |
| Architecture | DualAR TTS (HF model_type arktts; pipeline_tag text-to-speech). Slow AR 24 layers, width 896, 14 attention heads, 2 KV heads; Fast AR 4 layers, same width. 10 codec codebooks, 4,096 entries each. Card: inspired by Fish Audio S2 Pro DualAR. Fine-tune of Edge0/Audio8-TTS-Preview-0.6b (HF base_model). Custom Transformers code; load with trust_remote_code=True. Distinct from Qwen3-TTS (12Hz CustomVoice/VoiceDesign/Base family) and from Drama 3 (35B MoE). |
| Parameters | 0.6B (601,159,424; codec excluded) |
| Modalities | audio, speech, text |
| Languages | Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Spanish, Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, Vietnamese |
| Context | Up to 2,048 packed text/audio positions (card) |
| License | Apache-2.0 — LICENSE on Edge0/Audio8-TTS-Realistic-0.6b is Apache License 2.0 (cardData.license apache-2.0). The README says code and weights are Apache-2.0 and points at NOTICE for attribution. GitHub Edge0-AI/Audio8_TTS is also Apache-2.0. Card responsible-use text still applies: obtain permission to clone a voice. Not legal advice. |
| Hardware | Not yet verified — Card: Python 3.10+ and a CUDA-capable GPU recommended. Transformers package smoke-tested with PyTorch 2.8.0 and Transformers 4.57.3 on an NVIDIA H20 using BF16. HF usedStorage is 2,564,417,959 bytes (~2.39 GiB) including codec.pth. Peak VRAM and tok/s are not independently measured. Serving path named on the card is an SGLang Omni adapter; throughput was not benchmarked there. (not independently tested) |
| Hugging Face likes | 207 (as of 2026-11-02) |
| Model card | https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b |
| First seen | 2026-10-14 |
Audio8-TTS-Realistic-0.6B is Edge0’s 0.6B DualAR text-to-speech checkpoint. Official hosts we opened on 2026-11-02: the ungated card Edge0/Audio8-TTS-Realistic-0.6b (created 2026-10-14T03:02:49Z; last modified 2026-11-01T14:38:29Z; 207 likes / 4,613 downloads), LICENSE (Apache License 2.0), the Preview parent Edge0/Audio8-TTS-Preview-0.6b (424 likes / 3,983 downloads), github.com/Edge0-AI/Audio8_TTS (2,238 stars / 272 forks), arXiv 2610.14776, and the Gradio Space Edge0/Audio8-TTS-Realistic-0.6b (18 likes; runtime zero-a10g when opened). API gated was false. Hugging Face pipeline_tag is text-to-speech. library_name is transformers.
This identification page hangs off open ElevenLabs alternatives. That mapping is editorial for downloadable TTS weights. We have no sourced quality comparison against ElevenLabs.
It is not Qwen3-TTS (Qwen’s 12Hz CustomVoice / VoiceDesign / Base family). Head-to-head facts: Audio8-TTS-Realistic-0.6B vs Qwen3-TTS. It is also not Drama 3 (Fish Audio 35B MoE) and not LFM3-Audio-2B (Liquid any-to-any audio).
The card’s “substantially improved speech realism” sentence stays vendor. The same card says language quality varies and that the support list is not a claim of equal quality.
What we verified (2026-11-02)
| Field | Value | Source |
|---|---|---|
| Host | huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b (not gated) | HF API |
| Created / last modified | 2026-10-14T03:02:49Z / 2026-11-01T14:38:29Z | same |
| Likes / downloads | 207 / 4,613 | same |
| License tag + file | apache-2.0 + Apache License 2.0 | HF cardData + LICENSE |
| Pipeline tag | text-to-speech | HF API |
| Params (card) | 601,159,424 excluding the codec | README “Model Details” |
| safetensors.parameters | 601,159,424 BF16 | HF API |
| usedStorage | 2,564,417,959 bytes (~2.39 GiB) | same |
| Codec | 44.1 kHz, 10 codebooks, ~21.5 frames/s, bundled as codec.pth |
README |
| Context | Up to 2,048 packed text/audio positions | README |
| Languages | 23 (11 from Preview + 12 added) | README |
trust_remote_code |
Required | README |
| GitHub stars / forks | 2,238 / 272 | GitHub API Edge0-AI/Audio8_TTS |
| Preview parent likes / downloads | 424 / 3,983 | HF API |
Preview languages (11): Cantonese, Chinese, Dutch, English, French, German, Italian, Japanese, Korean, Polish, Spanish. Added in Realistic (12): Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish, Vietnamese.
How to run (from the official README, not tested)
Copied names only. forkalt has not loaded a checkpoint.
The card 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"Load with AutoProcessor.from_pretrained("Edge0/Audio8-TTS-Realistic-0.6b", trust_remote_code=True) and AutoModel.from_pretrained(..., trust_remote_code=True). Zero-shot cloning takes text, reference_audio, and reference_text (the transcript must match the recording). Generation without a reference omits those two fields. Sampling copied: temperature=0.8, top_p=0.95, top_k=50. Decode with model.decode_audio at model.config.codec_sample_rate (44.1 kHz on the card).
The same README names an SGLang Omni adapter with an OpenAI-compatible /v1/audio/speech API. Validation note on the card: Realistic was tested on an NVIDIA H20 in BF16; throughput and streaming were not benchmarked.
The Hugging Face Space is ZeroGPU-hosted. Queue and daily GPU allowance are Hugging Face product limits, not the weight grant.
Vendor benches (self-reported)
Copied Realistic-column rows only. The card says error rates measure intelligibility, not perceived realism, and that the Qwen/Higgs/Fish comparison is a reference comparison, not a matched ranking.
| Bench (card) | Realistic | Note |
|---|---|---|
| Seed-TTS EN/ZH error | 1.25% | Unweighted EN WER + ZH CER; hard Chinese excluded |
| Seed-TTS English WER | 1.464% | Same table |
| Seed-TTS Chinese CER | 1.029% | Same table |
| CV3 9-language error | 4.86% | Russian added vs the 8-language shared subset |
| MiniMax 24-language error | 5.89% | Includes languages outside the support list |
CV3 detail on the same card includes difficult newly supported languages (Thai WER 48.574%, Finnish 32.279%, Vietnamese 28.111%). That is why this page does not treat “23 languages” as equal quality.
License and responsible use
LICENSE is Apache License 2.0. The README points at NOTICE for attribution and thanks Fish Audio for DualAR. Custom Transformers files run under trust_remote_code=True — review them. The card’s responsible-use section: obtain permission to clone a voice; disclose synthetic audio where appropriate; noisy references or mismatched transcripts can reduce stability. Not legal advice.
Related on this site: Qwen3-TTS — Audio8-TTS-Realistic-0.6B vs Qwen3-TTS — ElevenLabs alternatives — Drama 3 — Phonon — LFM3-Audio-2B — LFM3-Omni-2B. Transcription (the other direction): Nemotron 3.5 Transcribe on Descript alternatives. Still images beside a voiceover: Fooocus, Iris-3B vs Qwen-Image-2.1-Turbo, PhotoCraft vs GIMP. Chat weights: Nemotron 3.5 Bolt, LFM3-1.7B. Style rewrite: humanizer. Catalog: home. Corrections: contact.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| Seed-TTS EN/ZH error (card, Realistic; unweighted EN WER + ZH CER) | 1.25% (self-reported) | https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b |
| Seed-TTS English WER (card, Realistic) | 1.464% (self-reported) | https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b |
| CV3 9-language error (card, Realistic) | 4.86% (self-reported) | https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b |
Deploy options
Comparisons
FAQ
Are Audio8-TTS-Realistic-0.6B weights public?
Yes. huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b is ungated (created 2026-10-14T03:02:49Z; last modified 2026-11-01T14:38:29Z; 207 likes / 4,613 downloads in the 2026-11-02 API check; pipeline_tag text-to-speech; card license apache-2.0). Code lives at github.com/Edge0-AI/Audio8_TTS (2,238 stars / 272 forks; GitHub license API Apache-2.0). It builds on Edge0/Audio8-TTS-Preview-0.6b (424 likes / 3,983 downloads that day).
Can I use it commercially?
The LICENSE file is Apache License 2.0. The README says code and model weights are Apache-2.0. The card still asks you to obtain permission to clone a voice and to disclose synthetic audio where appropriate. That is responsible-use text, not a second license. Not legal advice.
Does it beat ElevenLabs or Qwen3-TTS?
No sourced bake-off. The Realistic card prints vendor Seed-TTS / CV3 / MiniMax tables and a reference comparison that cites Qwen3-TTS report numbers. Those rows are self-reported or compiled from other reports. Language quality varies. See ElevenLabs alternatives and Audio8 vs Qwen3-TTS.
Do I need trust_remote_code?
Yes. The card says the model uses custom Transformers code (modeling_arktts.py and related files in the repo). Load with AutoProcessor / AutoModel and trust_remote_code=True. Review those files before you run them.
Sources & verification
- https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b
- https://huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b/raw/main/LICENSE
- https://huggingface.co/Edge0/Audio8-TTS-Preview-0.6b
- https://github.com/Edge0-AI/Audio8_TTS
- https://arxiv.org/abs/2610.14776
- https://huggingface.co/spaces/Edge0/Audio8-TTS-Realistic-0.6b
- https://github.com/QwenLM/Qwen3-TTS
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.