This paper describes an acoustic-phonetic analysis of the
unvoiced fricatives to find good features for their
classification. From a set of more than 35 features, about 15-
17 features were selected for the classification. An SVM is
trained with a limited set of phonemes and tested on all the
phonemes in the TIMIT speech database, in 3 configurations.
The procedure produces highly accurate results and
outperforms the results reported in [3], although there all the
fricatives (voiced and unvoiced) were included. The results
shown here indicate that the procedure could be utilized in
technologies for the hearing impaired, for example by
differential processing of the unvoiced fricatives, to improve
their discriminability.