Weak Speech Supervision: A case study of Dysarthria Severity Classification

Mirali Purohit, Mihir Parmar, Maitreya Patel, Harshit Malaviya, Hemant A. Patil

Abstract: Machine Learning methodologies are making a remarkable contribution, and yielding state-of-the-art results in different speech domains. With this exceptionally significant achievement, a large amount of labeled data is the largest bottleneck in the deployment of these speech systems. To generate massive data, hand-labeling training data is an intensively laborious task. This is problematic for clinical applications where obtaining such data labeled by speech pathologists is expensive and time-consuming. To overcome these problems, we introduce a new paradigm called Weak Speech Supervision (WSS), a first-of-its-kind system that helps users to train state-of-the-art classification models without hand-labeling training data. Users can write labeling functions (i.e., weak rules) to generate weak data from the unlabeled training set. In this paper, we provide the efficiency of this methodology via showing the case study of the severity-based binary classification of dysarthric speech. In WSS, we train a classifier on trusted data (labeled with 100% accuracy) via utilizing the weak data (labeled using weak supervision) to make our classifier model more efficient. Analysis of the proposed methodology is performed on Universal Access (UA) corpus. We got on an average 35.68% and 43.83% relative improvement in terms of accuracy and F1-score w.r.t. baselines, respectively.