In feature extracting hearing aids that aim at providing optimum pitch
information to the profoundly hearing impaired listener, special care
must be given to the robustness of pitch extraction. The performance
of three pitch tracking algorithms was studied as a function of signal
-to-noise ratio: 1) cross-correlation of the time signal followed by d
ynamic programming (CCF-DP) as implemented in the formant-extraction p
rogram of Entropic [5]; 2) Subharmonic Summation (SHS; [4]) and 3) an
artificial neural net consisting of a multi-layer perceptron (MLP; [3]
). The CCF-DP algorithm provided the highest accuracy in pitch estimat
ion, with SHS and MLP about equal when the effective time window of ML
P was extended to 40 ms. Voicing classification was most robust in MLP
, followed by SHS and CCF-DP. Classification by the latter algorithm a
ppeared to be poor at signal-to-noise ratios of 0 and -5 dB S/N.