Fish Speech
A TTS framework based on VQ-GAN and language models. The latest S2 Pro version features 4B parameters, supports 80+ languages, and is trained on 10 million hours of data.
Fish Speech is a next-generation open-source TTS framework developed by the Fish Audio team, combining VQ-GAN discretization encoding with the powerful generative capabilities of autoregressive language models. VQ-GAN compresses continuous speech signals into discrete token sequences, which are then processed by an LLM for autoregressive prediction before being decoded back into high-quality audio. This “audio tokenization + language model” approach endows the model with strong generalization abilities and multilingual adaptability. The model natively supports 80+ languages and achieves high-quality zero-shot voice cloning with just a few seconds of reference audio, running at near-real-time speeds on consumer-grade GPUs.
In 2025, Fish Audio released a major upgrade — Fish Speech S2 Pro. S2 Pro features 4B parameters trained on 10 million hours of multilingual data, achieving industry-leading results across multiple benchmarks, with significant improvements in cross-lingual voice cloning naturalness and voice similarity over its predecessor and competing solutions. S2 Pro supports 80+ languages, expanding from mainstream languages to less-commonly spoken ones and dialects, further strengthening Fish Speech’s position in multilingual TTS. The model weights are distributed under the Fish Audio Research License — note that its terms differ from Apache 2.0 and may restrict commercial use.
Fish Speech offers an online demo on Hugging Face Spaces and one-click local deployment, complemented by an active Discord community and a polished WebUI interface, greatly lowering the barrier to entry for open-source TTS technology. The project has surpassed 20,000 GitHub stars with continued strong community momentum.