End-to-End Classification of Reverberant Rooms using DNNs
share
Reverberation is present in our workplaces, our homes and even in places designed as auditoria, such as concert halls and theatres. This work investigates how deep learning can use the effect of reverberation on speech to classify a recording in terms of the room in which it was recorded in. Approaches previously taken in the literature for the task relied on handpicked acoustic parameters as features used by classifiers. Estimating the values of these parameters from reverberant speech involves estimation errors, inevitably impacting the classification accuracy. This paper shows how DNNs can perform the classification in an end-to-end fashion, therefore by operating directly on reverberant speech. Based on the above, a method for the training of generalisable DNN classifiers and a DNN architecture for the task are proposed. A study is also made on the relationship between feature-maps derived by DNNs and acoustic parameters that describe known properties of reverberation. In the experiments shown, AIRs are used that were measured in 7 real rooms. The classification accuracy of DNNs is compared between the case of having access to the AIRs and the case of having access only to the reverberant speech recorded in the same rooms. The experiments show that with access to the AIRs a DNN achieves an accuracy of 99.1 the proposed DNN achieves an accuracy of 86.9 testing procedure used in previous work, which relied on handpicked acoustic parameters, allowing the direct evaluation of the benefit of using deep learning.
READ FULL TEXT