What is Audio8 TTS?
Audio8 TTS is a family of compact text-to-speech models from Edge0 that clone a voice from a short reference recording and its transcript. The current release, Audio8-TTS-Realistic-Air, has 601 million parameters plus an audio codec and supports 23 languages, 12 more than the Preview 0.6B it builds on. It was uploaded to Hugging Face on 14 October 2026 alongside a paper describing training with a "realism critic" that rewards human-like prosody.
The architecture follows Fish Audio S2 Pro's DualAR design: a 24-layer slow model predicts one semantic token per frame and a 4-layer fast model predicts 10 codec codebooks, output at 44.1 kHz. Weights are BF16 safetensors under Apache 2.0, as is the code.
A smaller Preview 0.1B (about 170 million parameters plus a 120 million parameter codec) targets Chinese and English and runs on CPU as an ONNX INT8 export in about 0.4 GiB. It carries a separate community licence that requires a paid licence for companies with US$2 million or more in yearly revenue.
Is Audio8 TTS worth it?
Audio8 TTS is notable for size rather than outright quality. On the self-reported Seed-TTS English and Chinese test, Realistic Air's average error rate is 1.25%, against 1.11% for the 4B Higgs Audio v3, 1.32% for the 4.6B Fish S2 Pro and 1.67% for CosyVoice3. A model of 601 million parameters is about 1.2 GB of weights in BF16 by our arithmetic, small enough for a modest GPU.
The multilingual numbers are less flattering. On the MiniMax set of 24 languages its average error is 5.89%, against 4.40% for Fish S2 Pro and 2.74% for Higgs Audio v3, and Qwen3-TTS's 0.6B model scores 1.12% on Seed-TTS, slightly better. The card calls its table a reference comparison rather than a matched ranking and notes that quality varies by language; all figures come from Edge0, and none measures realism directly.
Serving is better documented than for most small labs. An SGLang Omni adapter offers an OpenAI-compatible speech API with batching; on one NVIDIA H20 it measured 0.691 s warm latency for about 5.9 s of audio, a real-time factor of 0.116, though that test ran on the Preview 0.6B with the same architecture. The card says Realistic Air's own throughput and streaming were not benchmarked, and there is no ONNX export for it yet, so CPU users are limited to the older Preview models.
The catch is provenance and safeguards. The maker's identity is thin: GitHub says Edge0-AI, Hugging Face says Edge0, the 0.1B card's code refers to Audio8, and no location or company details are published. The 0.1B card asks for consent before cloning, but neither card mentions watermarking. The repository is active, with 2,361 stars and a push on 2 November 2026, and Realistic Air drew 11,349 downloads in the last month.
Where it earns its place
A two-person studio localising indie games into Thai, Vietnamese and Indonesian
Voice the same characters across languages without per-character fees
Clone each actor from a short reference clip with its transcript and batch-render lines on one consumer GPU; Apache 2.0 allows commercial use. Check each language by ear, since the card warns quality varies.
A backend engineer at a voice-agent start-up
Serve cloned voices to many callers at once from one GPU
Run the SGLang Omni adapter, which batches up to 32 running requests by default behind an OpenAI-compatible /v1/audio/speech endpoint; the measured RTF of 0.116 on an H20 suggests headroom, but Realistic Air's own throughput is unmeasured.
A hobbyist building an offline reading assistant on a laptop
Speak Chinese and English text without a GPU
Use the Preview 0.1B ONNX INT8 runtime on CPU in about 0.4 GiB, with streaming PCM and a local web server. Fine for personal use; a company over US$2 million revenue needs a licence.
Before you commit
- Setup
- Python 3.10+ and a CUDA GPU; clone the repo, pip install -r requirements.txt and run audio8_tts_infer.py with the model ID, adding a reference clip and transcript to clone. SGLang Omni must be pinned to a tested commit.
- Integrations
- Transformers (4.57 to below 5; 5.x can produce silent output), SGLang Omni with an OpenAI-compatible speech API, ONNX Runtime for the Preview exports, supervised fine-tuning scripts.
- Data and privacy
- Self-hosted; nothing leaves your machines. Realistic Air and Preview 0.6B are Apache 2.0; Preview 0.1B's community licence requires a written commercial licence at US$2 million or more yearly revenue, counting affiliates.
- Cost at scale
- About 1.2 GB of BF16 weights for the 0.6B model; the SGLang server defaults to a 0.2 static memory fraction on one GPU. No hosted provider currently serves it.
- Limits
- 2,048 packed text and audio positions; quality uneven across languages; Realistic Air speed and streaming not benchmarked; long, noisy or mistranscribed reference clips reduce stability and similarity.
What’s good
- Apache 2.0 weights and code for the 0.6B models
- Seed-TTS error rate near models six to eight times larger
- 23 languages, 44.1 kHz output
- SGLang Omni server with OpenAI-compatible API, RTF 0.116 on an H20
- 0.1B ONNX INT8 runs on CPU in about 0.4 GiB
What’s not
- Trails Fish S2 Pro and Higgs Audio v3 on 24-language test
- All benchmarks self-reported; Realistic Air speed not measured
- No ONNX or CPU path for Realistic Air
- No watermarking; maker's details not published
- 0.1B licence charges companies over US$2 million revenue
Best for
Developers who need permissively licensed multilingual voice cloning on a small GPU budget and can test quality in their own languages.
Audio8 TTS alternatives
Audio8 TTS: common questions
- Is Audio8 TTS worth it?
- Audio8's 0.6B model clones voices in 23 languages at error rates near 4B-parameter rivals, a strong deal on size that still trails them on the hardest multilingual test. We rate it 3.8 out of 5.
- How much does Audio8 TTS cost?
- Free. Realistic Air and Preview 0.6B weights are Apache 2.0; Preview 0.1B is free for non-commercial use and for companies under US$2 million yearly revenue. No hosted API.
- What are the downsides of Audio8 TTS?
- Trails Fish S2 Pro and Higgs Audio v3 on 24-language test; all benchmarks self-reported; Realistic Air speed not measured; no ONNX or CPU path for Realistic Air; no watermarking; maker's details not published; 0.1B licence charges companies over US$2 million revenue.
- Who is Audio8 TTS best for?
- Developers who need permissively licensed multilingual voice cloning on a small GPU budget and can test quality in their own languages.
- How much does Audio8 TTS cost for a team or in production?
- About 1.2 GB of BF16 weights for the 0.6B model; the SGLang server defaults to a 0.2 static memory fraction on one GPU. No hosted provider currently serves it.
- Is Audio8 TTS safe for company data?
- Self-hosted; nothing leaves your machines. Realistic Air and Preview 0.6B are Apache 2.0; Preview 0.1B's community licence requires a written commercial licence at US$2 million or more yearly revenue, counting affiliates.
- What are the best Audio8 TTS alternatives?
- The closest alternatives we have reviewed are Qwen3-TTS, Chatterbox, OmniVoice and Kokoro-82M.
Answers checked Nov 5, 2026.