One main problem in adapatation of ML models from Python to Rust is feature extraction. I can run ONNX models, but some specific feature extractors can not be implemented in the Rust ecosystem. Ways to implement them: 1) In the AI era we can fastly prototype…

Hugging Face gave me access to their new Xet storage It's an example of how fast it is. Btw, under the hood is Rust. #hf #rust

Done #rust #speech

Also, made FP16 inference with https://crates.io/crates/half

CUDA execution

Idk why, FP16 execution is slower than FP32 shown on this picture

Speed on TensorRT even faster

Google Colab has been updated