The study of beamforming in sound classification optimization
Date of Issue2016-05-20
School of Electrical and Electronic Engineering
Sound Classification is the process by which we assign the audio signals to one of a number of classes, based on their features we extract from them. During cross validation process, a sample of data is partitioned into complementary subsets. One subset (called training set) are used to estimate boundaries, distributions or1class-membership among different classes. Another subset (called validation set) is used to perform the analysis. New data can be classifies based on these estimations. Multiple rounds of cross-validation with different partitions are performed in order to reduce variability. The validation results are averaged over the rounds. In order to improve the current sound classification system to be able to recognize desired sound signal at plural sound source case, Filter and Sum (FAS) beamforming and the Direction-Informed Speech Extraction algorithm via time-frequency masking (DISE) are applied before sound classification to reduce the interference. In this report, a detailed discussion of the effect of these two interference-reduction methods on sound classification is given based on the performance. The database used for the experiment are generated by adding two interference sources to the desired sound source in Matlab simulation environment in order to imitate the noisy environment in reality. As a result, fixed beamforming is able to help improve the performance of sound classification in noisy environment. Compared with fixed beamforming, DISE algorithm achieves higher interference suppression. But it causes more distortion and hurts the features at the same time, which is not good for sound classification.
DRNTU::Library and information science::Libraries::Cataloguing and classification
Final Year Project (FYP)
Nanyang Technological University