Combined PNCC feature extractor for robust speech recognition

Abstract

Recently, two major types of Power-Normalized Cepstral Coefficients (PNCCs) were proposed as noise robust Automatic Speech Recognition (ASR) front-end. All the literatures for these two PNCCs assume clean training data and clean or noisy test data. However, we find that one PNCC method has good performance for the clean training/noisy test scenario, but degrades when test data is cleaner than the training data. The other PNCC method performs relatively better for noisy training/clean test conditions, but is not very robust for the clean training/noisy test conditions. We propose Combined PNCC (C-PNCC) algorithm, which is superior to both previous PNCCs for clean training/noisy test cases, and which also has reasonably good performance for noisy training/clean test conditions.