Proper Error Estimation and Calibration for Attention-Based Encoder-Decoder Models

An attention-based automatic speech recognition (ASR) model generates a probability distribution of the tokens set at each time step. Recent studies have shown that calibration errors exist in the output probability distributions of attention-based ASR models trained to minimize the negative log likelihood. This study analyzes the causes of calibration errors in ASR model outputs and their impact on model performance. Based on the analysis, we argue that conventional methods for estimating calibration errors at the token level are unsuitable for ASR tasks. Accordingly, we propose a new calibration measure that estimates the calibration error at the sequence level. Moreover, we present a new post-hoc calibration function and training objective to mitigate the calibration error of the ASR model at the sequence level. Through experiments using the ASR benchmark, we show that the proposed methods effectively alleviate the calibration error of the ASR model and improve the generalization performance.