Discriminative feature domains for reverberant acoustic environments

Abstract

Several speech processing and audio data-mining applications rely on a description of the acoustic environment as a feature vector for classification. The discriminative properties of the feature domain play a crucial role in the effectiveness of these methods. In this work, we consider three environment identification tasks and the task of acoustic model selection for speech recognition. A set of acoustic parameters and Machine Learning algorithms for feature selection are used and an analysis is performed on the resulting feature domains for each task. In our experiments, a classification accuracy of 100% is achieved for the majority of tasks and the Word Error Rate is reduced by 20.73 percentage points for Automatic Speech Recognition when using the resulting domains. Experimental results indicate a significant dissimilarity in the parameter choices for the composition of the domains, which highlights the importance of the feature selection process for individual applications.

Speech Classification using Logical ART Deep Mechanism of Machine Learning

“…A. Morsy, 2017) propose a sub-vocal recognition system using deep learning approach suing surface electromyogram and a closely placed microphone. The system is tested with a small corpus of words and works at 9.44-word error rate [7].There are two methods of evaluating sound, 1) instrumental measure and 2) perceptual measure, out of which perceptual measure is more powerful but it is very much time exhaustive process as well as not very cost effective, therefore, measuring these is an open research challenge. The authors, (S. Safavi, A. Pearce, W. Wang and M. Plumbley, 2018) have developed a method to perform prediction of sound from the signals related to the perceptions from a leaning model.…”

End-to-End Classification of Reverberant Rooms Using DNNs

“…Acoustic environments shape and define aspects of the sounds we hear and through this process, we experience the world around us from an audible perspective. At the same time, the environments provide listeners with cues that enable the understanding of their properties [1]. Humans infer through reverberant sounds characteristics of the environment such as the size of a room [2].…”

“…A similar approach was later taken in [3] but with Frequency-Dependent Reverberation Times (FDRTs) as the features. In [1], a set of acoustic parameters was proposed as a feature domain able to discriminate between different rooms. The experiments showed that spectral and energy-decay features, such as the FDRTs, are discriminative information for the task.…”

Room Identification Using Frequency Dependence of Spectral Decay Statistics

“…The aim is to be able to identify the specific room in which a speech recording was made given a limited number of previously seen rooms. A variety of features may be helpful including estimating the geometry [3,4,5,6,7], the background noise [2] and reverberation characteristics [1,8].…”