What it does
Audio8 TTS is a family of small open text-to-speech models from Edge0. The current release, Audio8-TTS-Realistic-0.6B, has 601 million parameters plus a bundled 44.1 kHz audio codec. It clones a voice from a reference recording and its transcript (zero-shot, no training), or speaks without a reference. It supports 23 languages, 12 more than the earlier Preview model, and Edge0 says it sounds substantially more realistic than Preview. The earlier Preview 0.6B model stays available, and a smaller 0.1B model for Chinese and English runs on a CPU through ONNX. The code and the 0.6B weights are under Apache 2.0.
Official source 1 ↗ · Official source 2 ↗Features
- Zero-shot voice cloning. Give it a clean reference recording and the exact transcript of what is said in it, and it speaks new text in that voice without any training. Without a reference it uses a voice of its own. Official source 1 ↗
- 23 languages. Chinese, Cantonese, English, Japanese, Korean, Dutch, French, German, Italian, Polish and Spanish carried over from Preview, plus Arabic, Finnish, Hindi, Indonesian, Malay, Norwegian, Portuguese, Russian, Swedish, Thai, Turkish and Vietnamese, new in Realistic. Official source 1 ↗
- Small model. The main model has 601 million parameters (a two-stage DualAR design inspired by Fish Audio S2 Pro); the 44.1 kHz codec that encodes the reference and decodes the waveform is bundled, so no extra checkpoint is needed. Official source 1 ↗
- Benchmarks against larger models. Edge0 reports a 1.25% English/Chinese error rate on Seed-TTS, 4.86% across nine CV3 languages and 5.89% across 24 MiniMax languages. Its comparison table lists Higgs Audio v3 (4B) at 1.11 / 4.41 / 2.74 and Qwen3-TTS-1.7B at 1.01 / 7.73 / 27.41; Edge0 calls it a reference comparison, not a matched ranking. Official source 1 ↗
- Self-hosted speech API. The SGLang Omni plugin serves the model with paged attention and dynamic batching behind an OpenAI-compatible speech endpoint, with or without voice cloning. Official source 1 ↗
- Listening demo and fine-tuning. A demo page plays 24 English and Chinese clips of the same text and reference voice rendered by Audio8 and four other models. The repository includes command-line and batch inference and supervised fine-tuning. Official source 1 ↗ · Official source 2 ↗
- Realistic or Preview. Realistic 0.6B is the current release: same size as Preview, with more realistic speech and 23 languages instead of 11. Preview keeps its own checkpoint and listening page, and it is the one with an ONNX INT4 export; Realistic has none yet. On Seed-TTS English/Chinese, Edge0 reports 1.23% error for Preview and 1.25% for Realistic, with speaker similarity up from 0.682 to 0.704, from separate evaluations. Official source 1 ↗
Pricing & free limits
Download the Realistic 0.6B or Preview checkpoints from Hugging Face and run them on your own GPU. The 0.1B checkpoint has its own model license. Your only cost is the hardware. Official source 1 ↗ · Official source 2 ↗
Try speech generation and voice cloning without installing anything. Availability depends on the queue and your daily Hugging Face GPU allowance. Official source 1 ↗
DeepInfra hosts the Preview 0.6B model (11 languages, zero-shot voice cloning) as an API; you need a DeepInfra account. This is DeepInfra's price, not Edge0's, and it does not cover Realistic. Official source 1 ↗
Billing details
Billing details were not established by the checked official sources.
Free access & limits
The Playground runs on Hugging Face ZeroGPU, so how much you can generate depends on the queue and your daily GPU allowance. Official source 1 ↗
Best for & limitations
Developers who want a small, Apache-licensed voice-cloning TTS model they can host themselves, including behind an OpenAI-compatible speech API. Official source 1 ↗
Narration and voice-overs in one of 23 languages in a cloned voice, after getting the speaker's permission. Official source 1 ↗
On-device or CPU-only speech for Chinese and English with the 0.1B ONNX model. Official source 1 ↗
- Quality differs by language. Edge0 says the language list is not a claim of equal quality. On the CV3 test set, error rates are under 4% for Chinese, English, German and Spanish, but about 21% for Arabic and Hindi, 32% for Finnish and 49% for Thai; Cantonese scores poorly on the MiniMax set. Official source 1 ↗
- Needs a GPU and setup for the best model. Realistic 0.6B is meant for a CUDA GPU and has no ONNX export, so CPU use means the smaller Chinese/English 0.1B model. Loading it runs custom code (trust_remote_code), which Edge0 asks you to review first. Official source 1 ↗ · Official source 2 ↗
- Reference audio must be clean and transcribed. Cloning needs a clean recording and an exact transcript; noise or a mismatched transcript lowers stability and voice similarity. Long texts must be split into shorter utterances. Official source 1 ↗
- Can mispronounce, repeat or stop early. Edge0 warns that output may contain pronunciation errors, repetitions, omissions or early stopping, and asks you to listen before using it. Official source 1 ↗
- Vendor-run benchmarks. All scores are Edge0's own; they measure intelligibility and speaker similarity, not realism, and serving speed and streaming were not benchmarked. Official source 1 ↗ · Official source 2 ↗
- Voice consent is on you. Edge0 asks users to get permission before cloning a voice and to disclose synthetic audio where appropriate. Official source 1 ↗
- Use Edge0's checkpoints. The official weights are under the Edge0 account on Hugging Face (for example Edge0/Audio8-TTS-Realistic-0.6b), and the model runs custom code that Edge0 asks you to review before loading it with trust_remote_code. Re-uploads under other accounts are not Edge0's, so prefer the official repositories. Official source 1 ↗ · Official source 2 ↗
Submit a tool
