Reverberation degrades both perceptual speech quality and downstream Automatic Speech Recognition (ASR). This paper extends our prior work on Deep Internal Learning (DIL) for single-channel speech dereverberation to a downstream ASR setting. The proposed method trains an utterance-specific multi-scale frequency-band Convolutional Neural Network (CNN) on pseudo pairs derived solely from the reverberant input, requiring neither external data nor pretrained models. We evaluate on LibriSpeech test-clean convolved with measured Room Impulse Responses (RIRs) from three rooms with a wide T60 range, assessing both signal-level and recognition metrics under matched and mismatched degradation-prior configurations. The findings characterise the conditions under which DIL provides reliable ASR benefit and identify prior accuracy as a key determinant of recognition performance.