Abstract: The classical frequency transform can describe perception characteristics of human auditory system, but can not relatively enhance speaker's individuality in short-time spectrum of speech. A non-linear frequency transform and feature detection algorithm are proposed based on analyzing contribution of short-time spectrum in different frequency sub-bands and using of polynomial curve fitting. The experimental results show that the proposed non-linear frequency transform can improve the performance effectively in comparison with classical non-linear frequency transform such as Mel, Bark and ERB. In the same condition, the average error rate falls about 70.5%, 60.8% and 70.5% respectively. The proposed frequency scale and feature detection algorithm can strengthen the individual personality and improve the recognition performance.