KittenTTS

KittenTTS pushes text-to-speech toward the smallest viable footprint: the nano model is 15M parameters at about 25 MB quantized, with 40M micro and 80M mini variants topping out around 80 MB, all running through ONNX inference on CPU with no GPU anywhere in the loop. Eight built-in English voices ship with the release at 24 kHz output with adjustable speed, and multilingual support sits on the public roadmap. Installation is a pip install of the wheel published on GitHub releases, after which synthesis takes only a few lines of Python, making the models practical for single-board computers, offline appliances, and privacy-sensitive applications where audio must never leave the device. The project went viral in mid 2025 as a developer preview, and the maintainers still flag that APIs may change between releases, so production users should pin versions. KittenML publishes the models and code under the Apache-2.0 license, which permits commercial use, the repository has climbed past 15,000 GitHub stars, and a community ecosystem of wrapper servers and integrations has grown up around the models.