mariateleki / comparing-asr-systems

Code for our INTERSPEECH 2024 paper: Comparing ASR Systems in the Context of Speech Disfluencies.

README.txt

Code Design

The code is structured as two pipelines of scripts. The following diagrams capture the dependency structure of the scripts (the following script depends on the output of the previous script):

Modules

Install the listed dependencies for each of these modules -- following the instructions on each of their pages.

evaluate by HuggingFace

Link: https://github.com/huggingface/evaluate Library Version: 0.4.0 Python Version: 3.8

torcheval by PyTorch

Link: https://github.com/pytorch/torcheval Library Version: 0.0.7 Python Version: 3.8

WhisperX

Link: https://github.com/m-bain/whisperX Version: 3.1.1

OpenAI API

Link: https://platform.openai.com/docs/introduction Version: 1.9.0

english-fisher-annotations

Link: https://github.com/pariajm/english-fisher-annotations

We had to modify this code, so we provide the code here as a subdirectory.

Spotify Podcast Dataset

Link: https://podcastsdataset.byspotify.com/

This dataset is maintained by Spotify, and access to the dataset is determined by Spotify.

Additional Dependencies

• Pandas (Link: https://pandas.pydata.org/)
• tqdm (Link: https://github.com/tqdm/tqdm)