# Audio8-TTS-Realistic 0.6B (slim)

> Audio8-TTS-Realistic 0.6B is Edge0's open-weight text-to-speech model for 23 languages, a 0.6B dual-autoregressive model with a bundled 44.1 kHz codec that speaks with or without a reference voice cloned from one clip and its transcript.

- Full: https://www.anchorterminal.com/tools/audio8-tts-realistic-0-6b.md (~7,200 tokens) · this version ~2,380 tokens · JSON https://www.anchorterminal.com/tools/audio8-tts-realistic-0-6b.json · canonical https://www.anchorterminal.com/tools/audio8-tts-realistic-0-6b
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-30 · pricing https://www.anchorterminal.com/tools/audio8-tts-realistic-0-6b#pricing · reviews https://www.anchorterminal.com/tools/audio8-tts-realistic-0-6b#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**D · 51.5/100 · rank #103 of 117 open-weight models · #18 in Text-to-speech · not agent-ready · confidence low**


Assessment: A small local voice-cloning cousin for 23 languages under Apache 2.0, cloning from one clip and its exact transcript. It needs about 3 GB on a CUDA GPU through Transformers with Edge0's remote code, or Edge0's SGLang Omni plugin for a speech endpoint. The main risks are remote code and a pickled codec, and no independent benchmark yet.

## Facts

- Kind: Open-weight model · vendor: Edge0 · category: Text-to-speech · legal entity: Edge0 (edge0.ai), which names no legal entity or location on its site. The technical report's authors give their affiliation as AutoArk, and commit metadata in Edge0-AI/Audio8_TTS uses an autoark.ai domain · provenance 61/100
- Packages: huggingface `Edge0/Audio8-TTS-Realistic-0.6b`
- Auth: None · pricing: Free · x402: no · licence: Apache-2.0
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/Edge0/Audio8-TTS-Realistic-0.6b, read from the model card and its 8 commits, the file list and hashes, the configs, the custom Transformers code and the LICENSE, Edge0's GitHub repository (a blobless clone of Edge0-AI/Audio8_TTS with its guides and SGLang Omni adapter), the technical report (arXiv 2610.14776) and Edge0's site on 30 October 2026. Nothing was downloaded beyond those text files or run. Edge0 has no hosted API
- Model: A dual-autoregressive model in the style of Fish Audio S2 Pro. A 24-layer slow branch (width 896, from Qwen2.5 0.5B by the report) predicts one semantic token per audio frame, and a 4-layer fast branch fills the other nine of ten codebooks of 4,096 entries. 601,159,424 BF16 parameters by the Hub's count, without the codec. A bundled 44.1 kHz codec at about 21.5 frames a second encodes the reference and decodes the waveform. Trained in three stages, supervised fine-tuning, GRPO with WER and speaker-similarity rewards, and GRPO with a realism critic
- Languages: 23 by the card, Arabic, Cantonese, Chinese, Dutch, English, Finnish, French, German, Hindi, Indonesian, Italian, Japanese, Korean, Malay, Norwegian, Polish, Portuguese, Russian, Spanish, Swedish, Thai, Turkish and Vietnamese. The card says quality varies, and its own CV3 table gives error rates above 20% for Thai, Finnish, Vietnamese, Hindi and Arabic. Chinese and English are about 30% each of the fine-tuning data, other languages at most 3%
- Files: model.safetensors (1.20 GB, BF16), codec.pth (1.35 GB, a PyTorch pickle loaded with torch.load weights_only=True), tokenizer.json (12.2 MB) and the configs, 2.57 GB in all, with a SHA-256 for the three large files on the Hub. Four Python files (configuration, modeling, codec and processing) run through trust_remote_code. No chat template; the processor builds the prompt in code. The weights and code haven't changed since 14 October; later commits are card edits
- Licence: Apache 2.0 for the code and the weights, in the card's metadata and text and a LICENSE file in the Hugging Face and GitHub repositories, with a NOTICE crediting Fish Audio S2 Pro's DualAR design. No use policy is attached. The card asks users to get permission to clone a voice and to disclose synthetic audio. The base Preview 0.6B is Apache 2.0 and Qwen2.5-0.5B is Apache 2.0. The codec's origin isn't stated; its file differs from Fish Audio's S2 Pro and S1 mini codecs
- Runtimes: Hugging Face Transformers 4.57 or later and below 5 with trust_remote_code (Transformers 5 gives all-zero codes, per the deployment guide), Edge0's command-line scripts, and Edge0's SGLang Omni adapter, a plugin pinned to SGLang Omni 0.1.0 with an OpenAI-compatible /v1/audio/speech endpoint. Not a model type in released Transformers, vLLM, SGLang, llama.cpp, Ollama or LM Studio. Community conversions run it on audio.cpp and MLX
- Input and output: The processor takes a list of texts and, for cloning, one reference clip with its exact transcript per text; without a reference it picks a voice. generate takes sampling settings and max_new_tokens, and decode_audio returns 44.1 kHz waveforms with their lengths. Text, reference and transcript share a 2,048-position context, so long passages are split. The SGLang Omni service takes input, references, sampling settings and response_format (wav, pcm, or codec codes as codes, codec or npy), with server-sent-event streaming when the server opts in. One reference per request, no style or emotion controls, no timestamps
- Recommended settings: BF16 on CUDA, temperature 0.8, top_p 0.95, top_k 50 and max_new_tokens 1024 in the card's example (generation_config.json defaults to 0.7, 0.9, 50 and 512). A clean reference clip with a transcript that matches it word for word
- Hardware: Python 3.10 or newer and a CUDA GPU recommended; the card's code falls back to float32 on CPU. Smoke-tested on an NVIDIA H20 in BF16. No memory figure is stated; the files are 2.57 GB. Edge0's SGLang Omni figure (0.691 s warm p50, RTF 0.116 on an H20) was measured on Preview with the same architecture
- Training data: Cleaned Emilia and YODAS2 plus a private collection, filtered by language, duration and ASR error, by the technical report, with the share per language. No sizes in hours and no cut-off are stated, and the card describes none

- Scores: Capability pending, Reach 55, Risk 40, Agent interface 60, Documentation and upkeep 57 · total over the 7 assessed categories
- Why:  · Reach, Memory earns the full 30, since the BF16 files with the codec come to 2.57 GB, about 3.1 GB with the fifth added. · Risk, File safety gets 5, safetensors weights but a pickled codec beside them, and loading through trust_remote_code. · Agent interface, Read for a speech model, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. · Documentation and upkeep, The card covers everything but a knowledge cut-off (17).
- Sources: 18, open questions: 7, both in the full twin
- Capabilities: speech.tts, speech.languages, speech.streaming, voice.clone, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/audio8-tts-realistic-0-6b.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/audio8-tts-realistic-0-6b.svg` or a link to https://www.anchorterminal.com/tools/audio8-tts-realistic-0-6b from a page on edge0.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.
- Support: Discord https://discord.gg/Gmp3gdFsd3; https://edge0.ai/contact; https://github.com/Edge0-AI/Audio8_TTS/issues; more in the full twin

## Before you call it

1. Pin `transformers>=4.57.0,<5`; Transformers 5 can produce all-zero codes with this custom code
2. Load with `AutoProcessor.from_pretrained('Edge0/Audio8-TTS-Realistic-0.6b', trust_remote_code=True)` and `AutoModel.from_pretrained(..., trust_remote_code=True, dtype=torch.bfloat16)`; review the four .py files first
3. Clone with `processor(text=[...], reference_audio=['ref.wav'], reference_text=['exact transcript'], return_tensors='pt')`; a transcript that doesn't match the clip lowers stability
4. Decode with `model.decode_audio(output.codes)` and write at `model.config.codec_sample_rate` (44,100 Hz)
5. Split long text into short utterances; reference, transcript and text share 2,048 positions
6. For an HTTP endpoint, install Edge0's SGLang Omni adapter at the pinned SGLang Omni 0.1.0 and call `/v1/audio/speech`; start the server with `AUDIO8_TTS_STREAM_ENABLED=1` for streaming
7. Only clone a voice you have permission to use, and disclose synthetic audio

## Connect

```bash
git clone https://github.com/Edge0-AI/Audio8_TTS.git && cd Audio8_TTS
pip install -r requirements.txt
# voice cloning; drop the two reference flags to speak without a reference
python audio8_tts_infer.py --model Edge0/Audio8-TTS-Realistic-0.6b --text "Welcome to Audio8 TTS." --reference-audio reference.wav --reference-text "The exact transcript of the recording." --output outputs/realistic.wav
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| VieNeu-TTS v3 Turbo | C | 58.5 | speech.tts, speech.languages, speech.streaming, voice.clone, model.open-weights | https://www.anchorterminal.com/tools/vieneu-tts-v3-turbo.min.md |
| Fish Audio Drama 3 | C | 55.6 | speech.tts, speech.languages, speech.streaming, voice.clone, model.open-weights | https://www.anchorterminal.com/tools/fish-audio-drama-3.min.md |
| Fish Audio S2 Pro | D | 52.5 | speech.tts, speech.languages, speech.streaming, voice.clone, model.open-weights | https://www.anchorterminal.com/tools/fish-audio-s2-pro.min.md |
| HeyGen Voice | BB | 71.1 | speech.tts, speech.streaming, speech.languages, voice.clone | https://www.anchorterminal.com/tools/heygen-voice.min.md |
| Fish Audio TTS API | C | 60.8 | speech.tts, speech.streaming, speech.languages, voice.clone | https://www.anchorterminal.com/tools/fish-audio-tts.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)