Summary: IEEE SIGNAL PROCESSING LETTERS, VOL. 9, NO. 9, SEPTEMBER 2002 275
The Effect of Additive Noise on Speech Amplitude
Spectra: A Quantitative Analysis
Qifeng Zhu, Member, IEEE, and Abeer Alwan, Senior Member, IEEE
Abstract--This letter analyzes the effect of additive noise on
speech amplitude spectra, and introduces a method to estimate
speech spectra from noisy observations. Estimated spectra are used
to compute the Mel-Frequency Cepstral Coefficients as a recogni-
tion front-end. Compared to linear spectral subtraction, this tech-
nique improves the performance of digit recognition in noise.
Index Terms--Robust speech recognition, spectral subtraction.
NOISE ROBUSTNESS is an important challenge for au-
tomatic speech recognition (ASR). It is typically handled
by the acoustic model [often a hidden Markov model (HMM)],
and/or at the front-end (feature extraction).
When the noise is additive and stationary, and if one can es-
timate the average noise spectrum, a widely used technique for
noise removal is linear spectral subtraction (SS) , . SS at-
tempts to remove noise effects by subtracting the average mag-