Experimental evaluation of two-stage processing of blind source separation and DNN-based speech enhancement for in-car speech recognition

In this study, we present an experimental evaluation of a speech enhancement technique that combines multi-channel blind source separation (BSS) with single-channel DNN-based noise suppression in an in-car environment. Specifically, we first apply BSS to a mixture of target speech and driving noise, then select the separated signal that contains the speech and feed it into a single-channel DNN. This configuration enables speaker-independent enhancement via BSS, while the DNN can be trained without requiring knowledge of the microphone positions. To further improve performance, we fine-tune the DNN using BSS outputs for speech under driving noise conditions. We compare the proposed a two-stage approach with several baselines, including standalone BSS and multi-channel DNN-based methods. For evaluation, we use signal-to-distortion ratio improvement (SDRi) to assess noise reduction performance, and accuracy and character error rate (CER) to evaluate speech recognition performance. The experimental results highlight the effectiveness of integrating BSS with DNN-based processing to improve speech recognition in in-car environments.