Real-world speech rarely matches clean training data. This work targets the train–test distribution mismatch that degrades ASR in the field: background noise, channel artifacts, and far-field/reverberant capture.

• Cut WER 9.21% (LibriSpeech) and PER 5.92% (TIMIT) via residual-noise artifact removal, and improved far-field ASR on CHiME-5 with a TDNN denoising autoencoder (Joshi et al., 2018).