Fig. 1. Schematic diagram of network training and prediction

Fig. 2. ResUnet network structure

Fig. 3. Network structure of generator

Fig. 4. Discriminators in the time and frequency domains

Fig. 5. Network structure of time domain and frequency domain discriminator

Fig. 6. Schematic diagram of experimental apparatus

Fig. 7. The four objects used in the experiment (A4 sheet, A4 box, corrugated box, PET plastic bottle)

Fig. 8. The frequency response of the target (yellow: A4 sheet; purple: A4 box; green: Corrugated box; orange: PET plastic bottle)

Fig. 9. Spectrogram. (a) Clear original voice saved in SD card; (b) Laser speech collected by laser microphone; (c) Nonlinear function harmonic reconstruction method; (d) DNN+nonlinear harmonic reconstruction method; (e) Method of this article; 1-4: A4 sheets, A4 box, corrugated box, PET plastic bottles

Fig. 10. Calculation diagram of PESQ

Fig. 11. Objective evaluation trsults of speech quality. (a) PESQ method; (b) SNRseg method. Degraded: laser speech collected by laser microphone; harm_rec: nonlinear function harmonic reconstru-ction method; DNN+harm_rec: DNN+harmonic reconstruction method; ours: the method proposed in this paper