RED-ACE

Data and code for the paper RED-ACE: Robust Error Detection for ASR using Confidence Embeddings (EMNLP 2022).

RED-ACE is an ASR Error Detection (AED) model. Our approach is based on a modified BERT encoder with an additional embedding layer, which jointly encodes the textual input and the word-level confidence scores into a contextualized representation.

This is not an officially supported Google product.

Code

Please follow run.sh for an example how to run the code.

Data

Our dataset is placed in a public Google Cloud Storage Bucket and can be downloaded from this link.

Additional details and data description can be found in the paper.

BibTeX

If you find this useful for your work, please use the following citation:

@misc{2203.07172,
Author = {Zorik Gekhman and Dina Zverinski and Jonathan Mallinson and Genady Beryozkin},
Title = {RED-ACE: Robust Error Detection for ASR using Confidence Embeddings},
Year = {2022},
Eprint = {arXiv:2203.07172},
}