We Can "Hear" You With Radar

Detecting speech with a radar has several benefits over a regular microphone, i.e., the radar is tolerant to ambient noise. In addition to this, a radar based microphone could locate the exact location of the speaker, which is not possible with a single microphone. In this thesis we study the implementation of detecting speech with a Frequency Modulated Continuous Wave (FMCW) radar. A spectrogram of a viseme is extracted using the Short Time Fourier Transform (STFT). These visemes are the input of a Convolutional meural Network (CNN), which can classify the visemes. Eventually, it was shown that there is uniqueness in the spectrograms of visemes. The proposed CNN has an accuracy of 70% in detecting the right visemes