We address the fundamental inconsistencies of attention-based encoder-decoder (AED) models with long-form acoustic encoding. AED models trained for segmented pronunciation learn to encode frame positions entirely by using the acoustic context measured over segment boundaries, but fail to generalize when recording long-form segments where these cues disappear. The model loses the ability to order the acoustic encoding due to the variation of key permutations and values in cross-talk. We propose four changes: (1) to include a complete clear coding in focusing the attention of each extracted segment, (2) training of a long form with an extended acoustic context to eliminate the unclear local coding, (3) segment integration to combine the various stages required during the training of the didemantic segment and 4 (4) training stages. We show these changes bridge the accuracy gap between continuous and discrete acoustic encoding, allowing for automatic retrospective use of the attention recorder.
