Machine Audition of Acoustics

Propagation of sound from a source to a receiver in an enclosure can be modeled as an acoustic transmission channel. Objective room acoustic parameters are routinely used to quantify properties of such channels in the design and assessment of acoustically critical spaces such as concert halls, theatres and recording studios. Traditionally, room acoustic parameters are measured using artificial probe stimuli such as pseudo random sequences, white noise or sine sweeps. The noisy test signal hinders occupied in-situ measurements. On the other hand, virtually all audio signals acquired by a microphone have undergone a process of acoustic transmission in the first place. Properties of acoustic transmission channels are essential for the design of suitable equalizers to facilitate subsequent machine audition. Motivated by these needs, a number of new methods and algorithms have been developed recently to determine room acoustic parameters using machine audition of naturally occurring sound sources, i.e. speech and music. In particular, reverberation time, early decay time and speech transmission index can be estimated from received speech or music signals using statistical machine learning or maximum likelihood estimation in a semi-blind or blind fashion. Some of these estimation methods can achieve accuracies similar to those of traditional instrument measurements.