Compensation of Chann Spectrum Freq

Line Spectrum Frequencies (LSFs) is an effective and efficient representation for low bit-rate (LBR) speech coding. It is also appealing to use LSFs in speech or speaker recognition within a digital communication based system. However, the channel effect on LSFs degrades the recognition performance. This paper attempts to treat the problem of channel effect in LSF domain so that the recognition performance can be improved. A formulation describing the channel effect in channel phase is derived. Based on this formulation, the methods of channel estimation and channel mean subtraction are developed for the reduction of the channel effect. Experiments on connected digits recognition, where the speech signal is affected by channel or microphone, are conducted to show the effectiveness of the proposed methods.