Identifying Sentiment from Crowd Audio

Oftentimes the voices of a few drown out the voices of many. We interpret this problem literally as we investigate the ability of supervised machine learning models to predict sentiment directly from crowd audio in which multiple speakers are speaking simultaneously. Using a dataset of one second audio recordings of individuals saying “yes” or “no” (n=4753), we mixed together different voices speaking at the same time to create a large dataset (n=150k) of crowd audio responses. Each audio mixture was annotated with the specific proportion of the crowd response. On different mixture sizes ranging from 1 to 10 constituent voices, we trained BLSTM models using raw audio to predict the average sentiment of the crowd response. We were able to predict the average crowd sentiment for up to 10 speakers speaking simultaneously (Pearson’s r=0.70), suggesting the feasibility of a deep-learning approach to approximate simple sentiment directly from audio recordings with overlapping speakers. We discuss our results in the context of the need for machine listening systems that make can make equitable real-time determinations about crowd sentiment, so that no voice goes unheard.