Smart radar sensor for speech detection and enhancement

Abstract

Speech communication with microphone requires the user to wear additional gears or have a set of kits placed at a short distance from the speaker. Such gears limit the user’ activity and cause inconvenience in certain condition. This study attempts to use a new radar sensor as a non-contact method for speech detection. To achieve this kind of non-contact detection, a smart radar sensor was developed to capture the small vibration from the larynx of a speaker. This radar sensor operates at 35.5GHz and uses two parabolic antennas. A superheterodyne receiver is also employed to reduce the DC offsets and 1/f noise. This radar sensor provides a new method for speech communication. However, more complex noise components present in the output speech signal. Harmonics of millimeter wave and electrocircuit noise are the main part of the noise source. Therefore, to reduce the noise introduced from the radar sensor, we propose a higher-order statistics (HOS) algorithm suitable for the radar speech. First, third-order cumulant is applied to identify voiced frame from unvoiced one, and then bispectrum algorithm is used to further reduce the noise because of its superiority in suppressing additive Gaussian noise. The performance of the algorithm is evaluated both by means of subjective and objective measures. Results demonstrate that the microwave radar sensor can successfully detect speech signal from the speaker. In addition, the speech quality is improved significantly by the proposed algorithm. The smart radar sensor meets the requirement of non-contact detection. It will bring new light on free communication and extend the application of radar to speech detection.