FishAudio S1-mini

Fish Audio distilled this 0.5B text-to-speech model from its proprietary 4B S1, which was trained on over two million hours of speech. The open checkpoint uses a dual autoregressive design that predicts codec tokens and decodes them to audio, speaks 13 languages including English, Chinese, Japanese, Korean, and Arabic, and clones a voice from a short reference clip without fine-tuning. Parenthesized markers in the input text such as (excited), (whispering), or (laughing) steer emotion, tone, and nonverbal sounds, and the model was tuned with online RLHF. Weights sit in a gated Hugging Face repo that asks you to accept usage terms. Until the March 2026 S2 release, the fish-speech codebase was built around it, with command-line inference, a WebUI, an HTTP API server, and LoRA fine-tuning on Linux or WSL; those docs listed 12 GB of GPU memory, half what the newer 4B S2 Pro needs, and current docs center on S2 Pro. Weights are CC-BY-NC-SA-4.0 and the code falls under the Fish Audio Research License, so any commercial use needs a separate deal with Fish Audio. It targets hobbyists and researchers.