Content-Based Audio Retrieval by using Elitism GA-KNN Approach

Abstract

The digital audio became very popular and important in a computer user experience. The increasing amounts of audio data requires improvement, new methods and algorithms for processing this information. In this paper, our approach proposed the GA-KNN approach (Genetic Algorithm with K-Nearest Neighbor as fitness function) for content-based audio retrieval. The input is an audio file (query) and the output is a list of audio files ranked by their similarity. The system first extracts the features from an audio database and audio query. The query pattern is considered as a boundary for comparison. Then, the initial population in a genetic algorithm is constructed from a database containing all audio features. To improve the results, this paper uses Cosine measure in the genetic algorithm, and an improved selection method selection method to prevent the fittest chromosomes from being wasted in the new population by adding an Elitism feature, using 4% Elitism count. Furthermore, we proposed a new crossover method (Close Up-Feature Crossover) to create a new offspring by comparison between two audio patterns to query pattern. Finally, we evaluated our approach using a well-known audio database, which contains 409 sound samples of 16 classes to give 0.71475 as a precision of the audio retrieval.