A new kind of nonacoustic sensor speech detecting system is developed in our laboratory. Since the special attribute of the millimeter wave, this method may provide some exciting opportunity for wide applications. However, due to the special feature of the nonacoustic sensor speech detecting principle, the additive combined noises corrupted the nonacoustic sensor speech greatly. Therefore, an improved spectrum subtract algorithm, which is based on the wavelet packet entropy analysis is proposed in order to enhance the nonacoustic sensor speech. The results are also compared to the traditional spectral subtraction algorithm, suggest that this method achieves a better reduction of the whole-frequency noise, and yields good speech quality.