We deal with the basic incompatibility of attention-based encoder-decoder (AED) fashions with long-form acoustic encodings. AED fashions educated on segmented utterances be taught to encode absolute body positions by exploiting restricted acoustic context past section boundaries, however fail to generalize when decoding long-form segments the place these cues vanish. The mannequin loses capacity to order acoustic encodings on account of permutation invariance of keys and values in cross-attention. We suggest 4 modifications: (1) injecting express absolute positional encodings into cross-attention for every decoded section, (2) long-form coaching with prolonged acoustic context to get rid of implicit absolute place encoding, (3) section concatenation to cowl numerous segmentations wanted throughout coaching, and (4) semantic segmentation to align AED-decoded segments with coaching segments. We present these modifications shut the accuracy hole between steady and segmented acoustic encodings, enabling auto-regressive use of the eye decoder.
