A Mel-Filterbank and MFCC-based Neural Network Approach to Train the Houston Toad Call Detection System Design

摘要

Speaker recognition or voice detection is a state-of-art in the field of signal processing which includes human as well as animal. This paper proposes a naive approach to build a predictor model to detect the Houston Toad mating call signature in an audio file which can be paraphrased as toad voice activity detection. To accomplish that, several ideal toad call voice frames of unique characteristics in audio files have been experienced. The audio file is bandpass filtered, and then preprocessed by multiplying every frame with the hamming window to break into segments. Next, the Mel-Filterbank and Mel-Frequency Spectral Coefficient (MFCC) are used for feature extraction, while the Support Vector Machine (SVM) and Multi-layer Perceptron (MLP) neural networks are utilized as classifiers to determine the best fit. This experimental result reflects the higher accuracy of the MLP neural network over SVM showing the best potential of classification.