Kitten TTS is an open-source, ultra-lightweight text-to-speech library from KittenML that runs high-quality voice synthesis entirely on the CPU. Built on ONNX and released under the Apache 2.0 License, it packs a full TTS system into models ranging from 15M to 80M parameters — as small as 25 MB on disk — and has quickly passed 15,000 GitHub stars. Its appeal is simple: rather than chasing ever-larger models, Kitten TTS aims to make natural-sounding speech deployable on edge devices, laptops, and servers that have no GPU at all. ## Tiny Models, Real Speech The headline is size. The v0.8 release ships three model tiers — nano (15M), micro (40M), and mini (80M) — with an int8 nano variant that fits in roughly 25 MB. Because inference runs through ONNX Runtime on the CPU, there is no CUDA requirement and no GPU memory to manage, which makes Kitten TTS practical for embedded systems, offline desktop apps, and cheap cloud instances. Output is generated at a 24 kHz sample rate, a standard quality level for conversational and narration use, and a built-in text-preprocessing pipeline handles numbers, currencies, and units so raw text turns into clean speech without manual normalization. ## A Small Cast of Voices Kitten TTS ships with eight built-in voices — Bella, Jasper, Luna, Bruno, Rosie, Hugo, Kiki, and Leo — giving developers a range of speaker options out of the box without training or cloning anything. A `speed` parameter adjusts playback rate for faster or slower delivery, which is useful for accessibility readers, voice agents, and video narration where pacing matters. The choice of a fixed voice set rather than zero-shot cloning keeps the models tiny and sidesteps some of the consent concerns that come with voice-cloning systems. ## Developer-Friendly by Design Getting started is deliberately minimal. The library installs from a released wheel via pip, and generating audio takes only a few lines of Python: load a model from the Hugging Face Hub, call `generate` with text and a voice name, and write the result to a WAV file. Convenience helpers like `generate_to_file` and an `available_voices` list round out the API, and an optional CUDA backend is available for teams that do want GPU acceleration. A hosted demo on Hugging Face Spaces lets users hear the voices before installing anything. ## Trade-offs and Limitations Kitten TTS is explicitly labeled a developer preview, so its APIs may change between releases and some builds — notably the int8 nano model — have known issues that users are asked to report. The lightweight design trades away features found in heavier systems: there is no zero-shot voice cloning, the voice set is fixed, and multilingual coverage is limited compared with larger multilingual TTS stacks. As a small model optimized for footprint, its prosody and expressiveness, while strong for the size, will not match the largest GPU-bound models. Commercial support and enterprise licensing are offered separately for teams that need custom voices or integration help. ## Who Should Use This Kitten TTS is a natural fit for developers who need speech on constrained hardware: offline desktop tools, IoT and embedded devices, low-cost servers, accessibility features, and voice agents where a GPU is impractical or too expensive. The Apache 2.0 license, tiny footprint, and dead-simple Python API lower the barrier to shipping speech in places where running a large model was never an option — making it one of the more compelling entries in the growing on-device TTS space.