Training Set Selection - Speech

What is this data challenge?

The following challenge invites participants to design novel data-centric approaches towards limited vocabulary speech recognition, i.e., keyword spotting. Keyword spotting is a speech classification task where the model can detect a limited set of keywords [1]. Familiar examples of keyword spotting (KWS) models in production include the wakeword interfaces for Google Voice Assistant, Siri, and Alexa. 

In this challenge, your task will be to design a selection algorithm which chooses the best training examples from a dataset of spoken words (the Multilingual Spoken Words Corpus (MSWC)) which maximizes the classification accuracy on the evaluation set.

Successful approaches will aid in enabling voice-based interfaces for a wide variety of use-cases across a plethora of languages spoken by over 5 billion people, in a major step towards the democratization of voice technology. This challenge is part of a larger effort to emphasize data-centric approaches to machine learning. The current challenge is the first one in a series of challenges on improving training and testing datasets.