Signal samples X in a floating-point format, each of which is composed of 1 bit of sign S, 8 bits of exponent E and 23 bits of mantissa M, are converted through truncation by an integer formatting part 12 into signal samples Y in a 24-bit integer format, the integer-value signal samples Y are coded by a compressing part 13 into a code sequence Ca, and the code sequence Ca is output. According to the number of digits n following the most significant “1” in the integer-value signal sample Y, a difference producing part 14 extracts the least significant (23−n) bits from the mantissa M of the input signal sample X to form a difference signal Z, and a compressing part 17 performs entropy coding of the difference signal Z to produce a code sequence Cb and outputs the code sequence Cb. Alternatively, the difference signal Z may be output as it is, rather than being compressed.