To provide code book search technology which lightens the load on arithmetic operations and has accuracy close to that of a conventional coercive technique.

The speech encoding system which encodes a speech parameter in plural temporary successive frames composes multipulse vectors from the respective frames and serves as a remaining signal specifier. The multipulse vectors specify the temporary relation of plural pulses corresponding to a specific frame and includes plural sample positions. At each sample position, a value showing whether or not there is a pulse at a sample position in a vector and/or the sign is stored. Plural pulse positions in the specific multipulse vector are so optimized as to minimize the square mean error of a source signal and a quantized pulse train represented with the multipulse vector. Instead of the optimization, the pulse positions can be optimized so as to minimize the perceptually weighted square mean error of the source signal and quantized pulse train.