Audio8-TTS-Realistic-Air is a small open text-to-speech model that clones a voice from a short reference clip. It has 0.6 billion parameters, supports 23 languages and uses the Apache 2.0 license. The Hugging Face model page shows 240 likes and about 7,000 downloads as of November 2, 2026, which is why it is trending.
The model comes from Edge0 and builds on Audio8-TTS-Preview-0.6b. Its paper on arXiv, "Training a Compact, Strong, and Realistic Text-to-Speech Model," was submitted on October 10, 2026 by Runyuan Cai, Yu Lin, Yiming Wang and Xiaodong Zeng. Every number below comes from the model card or the paper, so it is self-reported. No independent listening test has been published that we could find.
What you get #
The card lists these highlights:
- Improved realism: more natural prosody and expressive delivery than Preview.
- 23 languages: 12 added to Preview's 11.
- Compact: 601,159,424 parameters in the main model, with the 44.1 kHz audio codec bundled separately.
- Two modes: zero-shot voice cloning from a reference recording and its transcript, or generation without a reference.
- Serving: an SGLang Omni adapter with paged attention, dynamic batching and an OpenAI-compatible
/v1/audio/speechAPI.
The repository includes the complete checkpoint, the codec, the tokenizer, the processor and custom Transformers code. The model file is about 1.2 GB in BF16, small enough for a single consumer GPU.
How it works #
Audio8 uses a DualAR design that the card says was inspired by Fish Audio S2 Pro. A slow autoregressive transformer predicts one semantic token per audio frame. A fast autoregressive transformer then predicts the frame's codec codebooks, conditioned on the slow model's hidden state and the earlier codebooks.
| Component | Configuration |
|---|---|
| Main model | 601,159,424 parameters, excluding the codec |
| Slow AR | 24 layers, width 896, 14 attention heads, 2 KV heads |
| Fast AR | 4 layers, width 896, 14 attention heads, 2 KV heads |
| Acoustic tokens | 10 codebooks, 4,096 entries each |
| Codec | 44.1 kHz, 2,048 samples per frame (about 21.5 frames per second) |
| Context | Up to 2,048 packed text and audio positions |
For readers who follow open voice models, the Fish Audio lineage matters. We covered the lab's commercial side in Fish Audio's $52M seed and S2.1 Pro and its open release in Fish Audio Drama 3.
The training idea: a realism critic #
The paper's argument is simple. TTS systems are usually tuned for intelligibility (word error rate) and speaker similarity. Neither captures the prosody that separates human performances from synthetic speech. The authors say there is a "realism gap" that bigger models and cleaner data narrow but do not close.
Their fix has three parts, as the abstract describes:
- A quality ladder. They build a ladder of outputs from widely used open-source TTS systems.
- A realism critic. It is trained to judge generated speech against matched human recordings of the same sentences. An identity-erasure objective discourages the critic from keying on a generator's fingerprints.
- Three training stages. Multilingual supervised fine-tuning, then GRPO with word error rate and speaker similarity rewards, then GRPO guided by the realism critic.
According to the abstract, the last stage "improves perceived prosodic naturalness and speaker similarity while preserving WER performance." This is the interesting part: reinforcement learning with a learned critic for something as fuzzy as realism. It is also the hardest claim to check, because the card says plainly that its metrics "do not directly measure perceived speech realism."
Benchmarks, read carefully #
Error rates are percentages, lower is better. Similarity (SIM) is on a 0 to 1 scale, higher is better. The card's summary:
| Benchmark | Coverage | Preview error | Realistic Air error | Preview SIM | Realistic Air SIM |
|---|---|---|---|---|---|
| Seed-TTS | English and Chinese | 1.23 | 1.25 | 0.682 | 0.704 |
| CV3 | 8 shared languages | 4.80 | 4.56 | not reported | 0.663 |
| CV3 | 9 core languages | not reported | 4.86 | not reported | 0.661 |
| MiniMax | 24 evaluated languages | not reported | 5.89 | not reported | 0.754 |
The honest reading: intelligibility is about the same as Preview on Seed-TTS (1.25 versus 1.23), while speaker similarity improves from 0.682 to 0.704. The headline gain, realism, is not in this table.
Against other open models
The card also gives a reference comparison. It warns that "different evaluation pipelines and text normalization make this a reference comparison rather than a strictly matched ranking," and that some figures are the authors' own re-evaluations.
| Model | Params (B) | Seed-TTS error | CV3 (9) error | MiniMax (24) error |
|---|---|---|---|---|
| Audio8-TTS-Realistic-Air | 0.6 | 1.25 | 4.86 | 5.89 |
| Qwen3-TTS-12Hz-0.6B | 0.6 | 1.12 | not listed | not listed | | Qwen3-TTS-1.7B | 1.7 | 1.01 | 7.73 | 27.41 | | Higgs Audio v3 | 4 | 1.11 | 4.41 | 2.74 | | Fish S2 Pro | 4.6 | 1.32 | 4.49 | 4.40 | | VoxCPM2 | 2.3 | 1.41 | 5.35 | not listed |
Audio8 is not the best on any single column. Its case is size and breadth: it matches or beats several larger models on Seed-TTS and CV3 while being 0.6B. On the 24-language MiniMax set, Higgs Audio v3 (2.74) and Fish S2 Pro (4.40) beat it. Parameter counts use different conventions across systems, and Audio8's 0.6B excludes its codec.
For other open voice options, see our guides on VoxCPM2 and the VoiceStudio local ElevenLabs alternative, and for a commercial yardstick, Cartesia Sonic 3.6 on Artificial Analysis.
Per-language results: where it is strong and where it is weak #
The averages hide a wide spread, and the card publishes per-language rows. That is the most useful part for anyone shipping outside English and Chinese. On the CV3 test (500 items per language), error rates in percent were:
| Language | Error % (lower is better) | Similarity |
|---|---|---|
| Chinese (CER) | 1.814 | 0.6898 |
| English (WER) | 2.907 | 0.6295 |
| German | 3.200 | 0.6669 | | Spanish | 3.389 | 0.6680 | | Portuguese | 6.871 | 0.6772 | | Russian | 7.258 | 0.6451 | | French | 7.938 | 0.6355 | | Arabic | 21.098 | 0.6122 | | Finnish | 32.279 | 0.7028 | | Vietnamese | 28.111 | 0.6790 | | Thai | 48.574 | 0.6691 |
Major European and East Asian languages look solid. Several of the 12 newly added languages do not: Thai at 48.6% word error, Finnish at 32.3% and Vietnamese at 28.1% on CV3 are far from production quality. The 24-language MiniMax set, with only 100 items per language, is kinder to most of them, but it shows Cantonese at 31.3% character error. The card itself says "the support list is not a claim of equal quality across languages."
So "23 languages" is a coverage claim, not a quality claim. If your product serves Thai or Finnish speakers, run your own test set first. Note too that the card lists known failure modes: pronunciation errors, repetitions, omissions and early stopping. Noisy reference audio or a mismatched transcript lowers stability, and the reference audio, transcript and target text share one finite context budget, so split long passages into shorter utterances.
How to run it #
The card recommends Python 3.10 or newer and a CUDA GPU. Install:
pip install "torch>=2.5.0" "torchaudio>=2.5.0" "transformers>=4.57.0,<5" "soundfile>=0.12" "safetensors>=0.4"
The authors smoke-tested it with PyTorch 2.8.0 and Transformers 4.57.3 on an NVIDIA H20 in BF16. Checks covered , English and Chinese generation with and without reference audio, and generation in Arabic, Cantonese, Portuguese and Russian. They note these are runtime checks, separate from the benchmarks.
Usage follows the standard Transformers pattern. Load the processor and model with trust_remote_code=True, pass text, plus reference_audio and reference_text for cloning, then call generate with sampling settings such as temperature 0.8, top-p 0.95 and top-k 50, and decode the codes to a waveform with decode_audio. The reference transcript must match the spoken content exactly, and you should use a clean recording. A hosted Playground on Hugging Face ZeroGPU lets you try it before installing, subject to queue and daily GPU limits.
For production serving, the SGLang Omni adapter installs as an independent plugin and does not overwrite core files. The authors validated non-streaming speech API generation on an H20 in BF16 with Torch compilation and CUDA graphs disabled. They state that throughput and streaming were not benchmarked, so measure latency yourself before building a voice agent on it.
Risks and caveats #
-
Custom code.
trust_remote_code=Trueruns code from the repository. Read the files first, and pin a revision. -
Voice cloning consent. The model can imitate a speaker from a short sample. Use only voices you have rights to, and label synthetic audio. Ourdeepfake and consent coverage shows how sensitive synthetic voices have become.
-
Language quality varies. The card says the supported-language list and the benchmark coverage are distinct, and that per-language results differ. Test your language with your speakers.
-
Self-reported results. Some comparison figures are the authors' re-evaluations, and the realism claim rests on the paper's own critic.
-
Naming. Do not confuse Edge0's Audio8 with other Edge0 releases such as the35B on-device model .
Who should try it #
If you run a multilingual product that needs voices on your own hardware, this is worth an afternoon: Apache 2.0 license, 1.2 GB of weights, an OpenAI-compatible serving path and a documented training recipe. If you need the best possible quality in one major language, larger models still post lower error rates in some columns, and a commercial API may still sound better. The sensible test is a blind listening comparison on your own sentences, plus a check for any language you care about beyond English and Chinese. Treat the realism claim as a hypothesis until someone outside the team runs that test.