This paper is concerned with the influence of different parametric representations of the speech signal in speaker-independent isolated-word recognition using clean and telephone quality speech.
Four parameters sets were considered: MFCC (Mel-Frequency Cepstrum Coefficients), PLP (Perceptual Linear Prediction Coefficients), PCC (Cepstrum Coefficients from PLP), LPCC (Cepstrum Coefficients from LPC).
Two recognition systems, one based on Dynamic Time Warping (DTW) and the other based on Hidden Markov Models (HMM), were used to estimate the behaviour of the acoustic parameters.
The experiments were conducted by using two different versions of the same database: the first one (clean) just limited to a telephone bandwidth and the second (noisy) transmitted over a telephone channel. The creation of these databases was realized under controlled conditions, in agreement with the specifications of the ESPRIT Project 2589 SAM. The results show comparable performances among the cepstral parameters sets in all cases in which training and test were realized in the same conditions.
Cite as: Brancaccio, A., Ceglie, F., D'Acunzo, G., Pelaez, C., Riccio, A., Rigosi, F. (1992) A comparative study of the influence of parameter processing on two different approaches for speech recognition in adverse environment. Proc. ETRW on Speech Processing in Adverse Conditions, 93-96
@inproceedings{brancaccio92_spac,
author={A. Brancaccio and F. Ceglie and G. D'Acunzo and C. Pelaez and A. Riccio and F. Rigosi},
title={{A comparative study of the influence of parameter processing on two different approaches for speech recognition in adverse environment}},
year=1992,
booktitle={Proc. ETRW on Speech Processing in Adverse Conditions},
pages={93--96}
}