Effectiveness of Histogram Equalization and SyDOCC Features on Speech Recognition Performance on a Real-World Noisy Speech Task

When building systems for automatic speech recognition, one often faces the challenge of dealing with speech signals containing noise. This additional noise leads to a drop in recognition performance, especially, when the acoustic environment varies during training and testing of a system. There exist several approaches to deal with noisy data or mismatched conditions. We evaluate two different approaches: Histogram Equalization (HEQ) and Synchronized Damped Oscillator Cepstral Coefficients (SyDOCC). While HEQ tries to normalize the statistical properties of the input features in an unsupervised manner without requiring a noise estimate, SyDOCCs model the acoustic properties of the human ear more accurately than Mel-Frequency Cepstral Coefficients (MFCCs). We evaluate both approaches using data with artificially added noise as well as data that contains genuine noise due to the recording conditions.