Speech description through MINERS: Model Invariant to Noise and Environment Robust for Speech

This paper presents a new speech descriptor called MINERS. It uses image analysis, Mel-Frequency Cepstrum Coefficients (MFCC), and a combination of Wavelet denoising and Power Normalized Cepstral Coefficients (PNCC) Temporal Masking. The purpose of this descriptor is to make easier feature extraction for speech. The descriptor is invariant to noise, environment, and speaker. For evaluation, NOIZEUS database was used for speech recognition through two classifiers: Support Vector Machine (SVM) and Hidden Markov Model (HMM). MINERS has better results among all other evaluated descriptors. The most successful approach was obtained using MINERS with a SVM classifier.