TensorFlow Lite for Microcontrollers is an open-source machine learning framework in which a TensorFlow model is built and trained on a host computer. That model is then reduced in size and computational complexity by an exporter that converts it to the TensorFlow Lite format. For the tiniest of compute platforms — microcontrollers — that model is then converted to a C array containing the model structure and any trained parameters, like weights and biases. On the microcontroller, an interpreter parses the C array to extract operations and data to run inferences against new input data.

Given that TF Lite for Microcontrollers runs on some heavily resource-constrained devices, I got to wondering whether or not I could run inferences against these models on a Commodore 64.

An application that uses AI and ML to intelligently strip the vocals out of music tracks. Written in Python, uses Tensorflow. The pre-trained model is included in the distribution. Use as a CLI tool or a library for your own code.

Can be used with audio files and probably a hot mic to transcribe speech into text for later processing. Uses git Large File Storage for the neural network objects. GPU acceleration enabled. Includes trained models as well as source code. Available in PyPy as deepspeech and deepspeech-gpu. Supports the RasPi explicitly as a platform, interestingly.

Looking at the releases page is a good way to keep up with the project: https://github.com/mozilla/DeepSpeech/releases