Fish Speech: audio generator — what it does and what you need to run it

Fish Speech is an open-source speech generator from Fish Audio that tackles the text-to-speech task: turning text into voiceover. The project publishes both the code and ready-to-use model weights, so you can run it yourself or study how a modern speech synthesis system is built.

What it does

The authors describe Fish Speech as SOTA Open Source TTS, i.e., they claim it is a state-of-the-art open-source speech synthesis system. It is based on the Dual-AR architecture, which, as described by the developer, is structurally isomorphic to standard large language models. Thanks to this, the model supports SGLang inference acceleration mechanisms: Continuous Batching, Paged KV Cache, CUDA Graph, and RadixAttention-based Prefix Caching.

The repository lists the tags llama, transformer, tts, valle, vits, vqgan, and vqvae — this gives an idea of the technology stack: transformer architecture, vector quantization, and approaches typical of neural speech synthesis.

What you need to run it

The code is written in Python. The repository was created in October 2023, with the last code change on August 3, 2026. The latest release is v1.5.1.

The model weights are published on Hugging Face. The largest weight file takes up 1.2 GB, and all weight files total 1.4 GB. The weights are licensed under cc-by-nc-sa-4.0: you can use them with attribution, for non-commercial purposes, and under the same license for derivative works. The code license is non-standard — the authors do not state it explicitly, so you should check the terms manually before using the code.

The platform, as described by the author, is CUDA. This is a developer statement, not the result of independent testing: the description states that the model supports CUDA acceleration via SGLang mechanisms.

Who it suits

Fish Speech suits those looking for an open-source speech synthesis system with public weights and the ability to run it yourself. The project will be of interest to developers who want to understand the architecture of modern TTS models or integrate speech generation into their pipeline, provided they comply with the non-commercial weights license.

The project is unlikely to suit those who need unrestricted commercial voiceover: the cc-by-nc-sa-4.0 weights license prohibits commercial use. It is also worth noting that, as described by the author, running it requires CUDA — this limits the range of suitable hardware. The non-standard code license will require separate review before use in your own projects.

Fish Speech is an open-source speech synthesis system with public weights and a transparent architecture. The project has been in development since 2023, has a current release v1.5.1, and relies on approaches typical of large language models. Before use, it is worth carefully reviewing the license terms: for weights, they restrict commercial use, and the authors ask you to check the code license manually.