KittenML/KittenTTS ? reverse-engineered prompt

Build me a small Python text to speech library that runs locally on CPU and does not need a GPU.

I want it to load a model from Hugging Face, turn text into spoken audio, and let me choose from a handful of built in voices. It should support changing the speaking speed, saving straight to a WAV file, and returning audio data I can use in Python. Please also include text cleanup so things like numbers, dates, times, currencies, and abbreviations sound natural before speech is generated.

Make it lightweight and easy to install, with a simple example script or two showing basic use and a streaming or GPU example if that is supported. Include tests for the main features, and keep the audio output at 24 kHz. If you need to check current library docs or model details online, go ahead and do that.