AIMpower

One of the focal points of AImpower.org is studying the role of technology in the marginalization of people who stutter. The organization has collected a dataset of stuttered speech recordings (72 people and a total of 50 hours) containing both conversational and voice command recitation speech, and have manually created verbatim transcriptions that include disfluent utterances. Furthermore, they’ve labeled the recordings on the different range of stuttering (mild to a lot), different kinds of stuttering (prolongation, filler words, long silence, repetition), and whether it’s freeform/casual conversation vs reading of voice commands. The goal of this project is to build off of Summer 2024, where they benchmarked these recordings against popular speech recognition models provided by OpenAI, Google, Facebook, Microsoft, and Otter.AI. You will quantify fluency bias in each of these speech models and visualize how the results change with different types of stuttering and across each model. *knowing chinese is not required, but knowing them may be beneficial