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
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
However, all the HEAR benchmark tasks are available to download
in a common pre-preprocessed format with human-readable metadata.
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.
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:
hear-validator provides a command-line tool to verify that an audio embedding model
conforms to the common API