Standardized documentation practices for speech data collection with emphasis on speaker diversity

What is it about?

This paper extends the "Datasheets for Datasets" framework for speech data by emphasizing documentation of often-overlooked “linguistic subpopulations”: accents, dialects, varieties, languages, and speech arising from speech disorders or pathologies. Documentation questions to answer include, e.g, why does the dataset focus on a specific language/region? Is there consensus on how you define a certain accent? Are speech types self-reported? Empty datasheet templates (in .docx and .tex) and worked examples of common speech datasets (CORAAL, CommonVoice, LibriSpeech, VoxPopuli, WHAM) are available on Github: https://github.com/SonyResearch/project_ethics_augmented_datasheets_for_speech_datasets

Featured Image

Why is it important?

Our augmented datasheets provide a standardized way to check for the linguistic make-up of speech data. Datasheets are useful to dataset creators, because linguistic subpopulations should be considered before beginning data collection. Datasheets are also useful to dataset users: when combining multiple speech datasets to build models, is there appropriate coverage across speech types to train a robust model?