Unimodal aggregation (UMA) enhances feature representations in non-autoregressive speech recognition, reducing errors and complexity.
