Uncertainty Quantification for Speech-To-Text in Spanish

Speech-to-text is a task which has been recently boosted its performance in practical applications thanks to the advent of deep neural network architectures. Moreover, recent speech-to-text models such as Whisper, have benefited from extensive pre-training with very large datasets. However, its usage in different target scenarios (in less-represented languages, speech recorded in noisy environments, etc.) might yield unsatisfactory results. Uncertainty quantification in the previously described context is important for the safe usage of speech-totext models. This study evaluates three methods for uncertainty quantification: Monte-Carlo dropout, temperature scaling, and feature density estimation. The analysis is conducted using Spanish audio datasets to assess these methods in the context of a less-represented language. A novel metric, analogous to the expected calibration error, is introduced to measure the correlation between predicted uncertainty and word error rate. We provide a detailed description of the dataset construction and experimental parameters. The findings indicate that Whisper demonstrates strong performance with Monte-Carlo dropout and temperature scaling, while the feature density estimation method shows comparatively lower efficacy. Finally, we propose enhancements to the evaluation procedures to further reduce prediction uncertainty.