Producing novel machine learning research for those with motor, neurological and articulatory speech disorders
In my "ASR for Dysarthria" project I outlined an interface solution to the problems that quadriplegics face.
Please navigate to that project from the drop-down as it was the preliminary work for this one.
This solution was limited to current state-of-art which is ineffective.
Dysarthric speech has more diffuse regions than non-disordered speech, so generic voice assistants like Alexa and Siri do not recognize voice commands
Voice assistants use transformers trained on massive datasets, but the amount of training data for dysarthric speech is insufficient
Proposed a two-stage training process (fine-tuning a speech-based Large Language Model like ChatGPT) and wake-word detection process
Collected the largest dataset of dysarthric speech and used SpeechBrain to train on the latest transformers.
Training revealed that the tranformer could not converge on keyword spotting.
Alternatively, tried merging samples to create longer tokens, but this reduced the number of observations preventing learning.
Literature Review conducted summarizing research and surverying state-of-the-art for Automatic Speech Recognition.
Academic publication disseminating my research-- presented at the Young Researchers Consoritum (ICCHP-AATE) 2022.