ONNX models

This release contains all capacities of CREPE models in the ONNX format.

Overview

There are totally 5 model weights of different capacities: full, large, medium, small and tiny.

• The 'full' and 'tiny' models were converted from the torchcrepe repository.
• The 'large', 'medium' and 'small' models were converted from the original TensorFlow implementation using scripts in this repository.

Usage

1. Put these models into the onnxcrepe/assets/ directory.
2. Edit model in the configuration file(s) in onnxcrepe/configs/ directory to switch among different model capacities.

Speed and performance

As the 'full' capacity refers to the model in the original paper, smaller models provide a trade-off between the computation time versus the time resolution or slightly lower accuracy. More details of the speed comparison can be found here.

Notes: higher rates of octave errors were observed when predicting with smaller models. Carefully tuning fmin and fmax may help.