HEAR Code
Evaluation Code
The hear-eval-kit is used to evaluate audio
embedding models on the HEAR datasets. It consists of the following functionality: 1)
compute audio embeddings and 2) evaluate audio embeddings using a shallow downstream
MLP classifier.
For more information on the downstream training regime used for evaluation, please see
Appendix B of our paper. Information to help you
get started using the hear-eval-kit is available on our github page.
Data Pre-processing Code
Individuals who would like to regenerate the HEAR benchmark tasks themselves can use hear-preprocess.
However, all the HEAR benchmark tasks are available to download
in a common pre-preprocessed format with human-readable metadata.
The hear-preprocess package was used to process HEAR datasets into
this format. It can also easily be extended to process new datasets beyond those included
in the HEAR benchmark.
Baseline Models
Several baseline models that implement the common API are
in included the hearbaseline Python package. Two baseline models on the leaderboard
are included in this repository: wav2vec2 and CREPE.
Model Validator
hear-validator provides a command-line tool to verify that an audio embedding model
conforms to the common API
required by hear-eval-kit.