Robust Phoneme Recognition Using High Resolution Temporal Envelopes

Frequency domain linear prediction (FDLP) is a technique for auto-regressive (AR) modeling of Hilbert envelopes of the signal. The model is derived by the application of linear prediction on the discrete cosine transform (DCT) of the signal. In this paper, we propose modifications of the basic FDLP approach for deriving high resolution envelopes. We determine various factors which affect temporal resolution in FDLP such as the location of the input peaks within the analysis segment, type of window applied in the DCT of the signal, and order of the FDLP model. This analysis enables us to improve the resolution of temporal envelopes derived from FDLP. The features extracted from high resolution envelopes outperform MFCC features in noisy phoneme recognition experiments (relative improvements of 10 %) and phoneme recognition in conversational telephone speech (relative improvements of 5 %).