Statistical spectrogram modeling and analysis for blind estimation of room acoustics from musical recordings.

Abstract

An algorithm employing Markov random field modeling has been applied to spectrographic representations of musical recordings to uncover acoustical features of the recording environment. For recorded music, the reverberation pattern is most visible at the onset/offset edges of harmonic components in the spectrogram. Edge features extracted through image analysis algorithms can be mapped to the acoustical features of the recording space, but the correspondence is complicated by the inherent randomness of the musical “test signal.” This is exacerbated in the media production process because a final mix can be generated from separate recordings from multiple acoustic spaces. The Markov random field modeling algorithm utilizes data-modeling techniques to estimate the probabilistic links between edge analysis results and room acoustical features and further to identify the probabilistic nature of any time-variation in latent room acoustics. Our algorithm obtained enhanced room acoustics feature extraction performance by allowing a gradual refinement of the room acoustics feature vectors through dynamic fusion of prior estimates and knowledge from current musical segments. This on-line approach is also more computationally efficient compared to batch processing. Multiple variants of the proposed algorithm are demonstrated and compared to existing blind room acoustics parameter estimation methods.