Impress: A Machine Learning Approach to Soundscape Affect Classification for a Music Performance Environment

Soundscape composition in improvisation and performance contexts involves many processes that can become overwhelming for a performer, impacting on the quality of the composition. One important task is evaluating the mood of a composition for evoking accurate associations and memories of a soundscape. We present a new system called Impress that uses supervised machine learning for the acquisition and realtime feedback of soundscape a↵ect. We used an audio features vector of audio descriptors to represent an audio signal for fitting multiple regression models to predict soundscape a↵ect. A model of soundscape a↵ect is created by users entering evaluations of audio environments using a mobile device. The same device then provides feedback to the user of the predicted mood of other audio environments. The evaluation of the Impress system suggests the tool is e↵ective in predicting soundscape a↵ect.