Experimental Evaluation of Speech Enhancement for In-Car Environment Using Blind Source Separation and DNN-based Noise Suppression

In this study, we evaluate speech enhancement using multi-channel blind source separation (BSS) and single-channel DNN-based noise suppression in an in-car environment. Specifically, blind source separation is applied to a mixture of target speech and driving noise, and one of the outputs is fed into the DNN. This combination allows for speaker position-independent enhancement through blind source separation, while the single-channel DNN can be trained regardless of the microphone position. To further improve performance, we fine-tune the pre-trained model using the BSS outputs, including driving noise. Our evaluation criteria include signal-to-distortion ratio improvement (SDRi) for noise reduction performance, and accuracy and character error rate (CER) for speech recognition performance. Results demonstrate that SDRi for the proposed method reached 28.37 dB. Additionally, the average speech recognition accuracy improved from 0.785 to 0.826, and CER improved from 15.3% to 11.9%. These findings demonstrate the potential of combining BSS and DNN to enhance speech recognition capabilities in in-car environments.