Contribution
Using machine learning to identify acoustic fingerprints of concert halls in classical audio recordings
* Presenting author
Abstract:
Classical music is commonly performed and recorded in concert halls with distinct acoustical properties. However, the question remains to what extent these properties are still audible in stereo recordings that have been submitted to professional post-processing. This study therefore investigated the capability of machine learning models to identify such 'acoustical fingerprints' of concert halls in classical music recordings. To that end, recordings from renowned venues such as the Golden Hall of the Wiener Musikverein, the Concertgebouw in Amsterdam, and the Berliner Philharmonie, were used as a database to detect free decay regions in which the room was clearly audible. The isolated raw audio sequences were then utilised to classify the recording locations with Convolutional Neural Networks (CNNs). Despite challenges related to dataset size, recording quality, and post-processing, the study showed promising initial results, suggesting that machine-learning models can indeed detect room-specific acoustic characteristics in stereo recordings. Further research with larger datasets and refined methods are being conducted to improve accuracy and generalisation.