A methodology, which is oriented to voice classification, is proposed for selecting acoustic features. The raw voice characteristic assemble is preprocessed by means of statistical techniques and thereafter its reduction up to the lowest assemble dimension of representative voice parameters is accomplished, yet preserving enough discriminating properties of voice classes. The methodology introduced shows an important reduction in initial assemble dimension of voice characteristics. In addition, a method of background noise reduction for quality improvement of acoustic voice analysis is developed. The method accomplishes a spectral subtraction technique.