Audio8-TTS-Realistic-Air: Edge0's 0.6B Multilingual Text-to-Speech Model Explained

Audio8-TTS-Realistic-Air is a 0.6 billion parameter multilingual text-to-speech model developed by Edge0. It can generate speech from text in 23 different languages, or clone a voice using a reference audio recording and matching transcript. The model can also be deployed through an SGLang Omni adapter with an OpenAI-compatible speech API.

The model uses a custom DualAR architecture, consisting of a slow autoregressive transformer and a fast autoregressive transformer. The main model has 601 million parameters, while the codec is a separate component. The slow transformer has 24 layers, while the fast transformer has 4 layers. The model uses 10 codebooks with 4,096 entries each and processes audio at 44.1 kHz with 2,048 samples per frame.

To use the model, text input is passed to a processor, which can also handle reference audio and transcript for voice cloning. Generation parameters include max_new_tokens, temperature, top_p, top_k, and do_sample. The model outputs audio codes that are decoded into waveforms, which can be saved as a WAV file. The SGLang Omni adapter supports a non-streaming OpenAI-compatible speech API.

Compared to a previous Preview release, Realistic Air adds 12 supported languages, with a slightly lower error rate and speaker similarity metric. The model requires Python 3.10 or newer and a CUDA-capable GPU, with PyTorch 2.8.0, Transformers 4.57.3, and BF16 recommended for optimal performance.

The model is licensed under Apache-2.0. For more details, visit the Audio8-TTS-Realistic-Air model page or follow Edge0 on Twitter.