Explanations for Automatic Speech Recognition

The paper presents a novel approach to quality assessment in neural network-based Automatic Speech Recognition (ASR) systems by generating explanations for transcriptions, which enhance understanding and trust in the models. It introduces a method that identifies a minimal and sufficient subset of audio frames responsible for a given transcription, adapting techniques such as Statistical Fault Localization (SFL) and Causal explanations, along with an adapted version of LIME for ASR. Evaluations conducted on ASR models including Google API, Sphinx, and Deepspeech using the Commonvoice dataset demonstrate the effectiveness of the proposed explanation techniques, which are crucial for practitioners seeking to improve interpretability and reliability in ASR systems.