A multilingual text-to-speech system that generates speech in over 80 languages and clones voices from short 10-30 second audio samples. It offers inline emotional control using tags and supports multi-speaker generation, with Docker, Python, and source deployment options.
Project overview
Achieves the lowest WER among all evaluated models including closed-source systems on Seed-TTS Eval.
Project type
Model Runtime · Audio & Speech
Use cases
Audio & Speech
Deployment
Refer to project documentation
License
License pending
Best for
Developers, creators, and AI engineers who need to generate speech in over 80 languages or clone voices from short reference samples.
Teams building multi-speaker or multi-turn dialogue applications that require inline emotional control via [tag] syntax.
Key capabilities
Generates natural, realistic, and emotionally rich speech in over 80 languages without requiring phonemes or language-specific preprocessing.
Clones voices using short reference samples (typically 10-30 seconds) capturing timbre, speaking style, and emotional tendencies without additional fine-tuning.
Embeds emotional instructions at any position in the text using simple [tag] syntax (e.g., [whisper], [excited], [angry]).
Processes reference audio containing multiple speakers and controls performance via speaker ID tokens, enabling a single generation to include multiple speakers.
Leverages previous information to improve the expressiveness of subsequent generated content, increasing the naturalness of the dialogue.
Limitations and risks
Released under the FISH AUDIO RESEARCH LICENSE, which may restrict commercial usage.
The developers do not hold responsibility for illegal usage of the codebase; users should refer to local laws about DMCA.
Getting started
Setup is rated medium difficulty because it requires Python setup, downloading large model weights, and configuring the inference server or WebUI. Users should follow the official installation instructions and run command line or WebUI inference. Interaction is available via web GUI, CLI, and API.
Alternatives and comparisons
A text-to-speech generation library with pretrained models in over 1100 languages, training tools, and voice cloning capabilities.
A text-to-speech model optimized for dialogue scenarios, providing fine-grained prosodic control and multi-speaker support.
A tokenizer-free text-to-speech model supporting 30 languages and highly accurate voice cloning.
README: State-of-the-art multilingual text-to-speech (TTS) system, redefining the boundaries of voice generation.
README: On Seed-TTS Eval, S2 achieves the lowest WER among all evaluated models including closed-source systems
README: S2 Pro supports over 80 languages without requiring phonemes or language-specific preprocessing
README: Fish Audio S2 supports accurate voice cloning using short reference samples (typically 10-30 seconds). The model captures timbre, speaking style, and emotional tendencies, produci…