The invention relates to a method for estimating fractional pitch of a code-excited linear prediction voice coder, which directly estimates a peak value position through a polynomial interpolation and obtains the fractional pitch estimation. A fractional pitch estimation value is obtained by the following steps: (1) performing an open-loop pitch estimation on an object signal, calculating an estimation value Top of an open-loop integer pitch; (2) continuing to perform the open-loop pitch estimation and calculating a close-loop correlation sequence C(k) in a given neighborhood of the open-loop integer pitch estimation value Top; (3) if a peak value point Tcl of the close-loop correlation sequence C(k) appears at two ends of the given neighborhood, then the fractional pitch estimation value Tfr being zero; if the peak point Tcl appears in the given neighborhood, then performing a polynomial fitting based on the peak point value Tcl and the values of the former point Tcl-1 and the post point Tcl+1, solving out the peak value position Tr, performing linear quantization to the peak value position Tr according to an interpolation factor D and obtaining the fractional pitch estimation value Tfr, a linear quantization formula being Tfr=round(Tr*D). Under the premise that the accurate fractional pitch estimation value is obtained and the prediction grain is improved, the invention effectively reduces the computational complexity and system overhead.