kittenml/kittentts is a family of open-source text-to-speech models from KittenML that makes speech generation usable on CPUs and consumer GPUs through a single Python library. It supports zero-shot voice cloning, 20 languages, and real-time CPU inference.
Project overview
It runs real-time speech generation on CPU alone, including faster-than-realtime generation on a single smartphone CPU thread with only 40–50 MB peak memory for the smallest model, removing the GPU requirement that typically gates high-quality TTS.
Project type
Model Runtime · Audio & Speech
Deployment
Refer to project documentation
License
Apache-2.0
Best for
Developers who need to generate speech locally on CPUs or consumer GPUs from Python code.
AI engineers needing multilingual speech (20 languages) with expression control and optional zero-shot voice cloning.
Key capabilities
Runs real time on a CPU without a GPU, with faster-than-realtime generation on a single smartphone CPU thread and only 40–50 MB peak memory for the smallest model.
Clone any speaker from 5–30 seconds of reference audio, no fine-tuning; 115 built-in voices are also listed via m.available_voices.
Generates speech in 20 languages, with expression control via [emotion] tags, inline <event> tags and (((emphasis))) spans, and 24 kHz audio output.
Limitations and risks
Kitten-TTS 2 C++ inference is in beta.
The models are licensed separately from the Apache-2.0 code and their terms may differ; review the model terms before adoption.
Usage is via Python code; the repository documents no no-code interface (a hosted demo exists separately).
Getting started
Install with pip install kittenml, instantiate KittenTTS with a model name, then call m.generate(text, voice=...) and save the output. Setup difficulty is easy.
Evidence and sources
GitHub project description: Open-source State-of-the-art TTS model which runs on a CPU 😻
README: we make high-quality speech generation accessible on CPUs and consumer GPUs
README: Pass `reference=` instead of `voice=` — same method, the recording just replaces the built-in speaker. Give it 5-30 seconds of a single speaker.
README: **Faster-than-realtime speech generation on a single smartphone CPU thread**, with only **40–50 MB peak memory usage**