Can We Trust Explainable AI Methods on ASR? An Evaluation on Phoneme Recognition

XAI methods designed for images may not work well for speech without adaptation: The researchers found that applying LIME directly from image classification to speech recognition required modification, suggesting that explanation techniques cannot simply be transferred between different types of data without careful evaluation.

A modified LIME approach achieved 96% accuracy in identifying important audio segments: When the researchers adapted LIME to use time-partitioned audio segments rather than the original method, the top three identified segments contained the correct explanation 96% of the time, demonstrating that tailored approaches can produce trustworthy explanations for speech models.

Controlled testing with ground truth is essential for validating XAI methods: By using a phoneme recognition task with expert-annotated data, the researchers were able to objectively verify whether explanations were actually correct—highlighting that responsible AI requires rigorous validation rather than assuming explanations are trustworthy.

We still need more evidence before deploying XAI for speech recognition in real applications: The paper emphasizes that despite emerging interest in using XAI for ASR systems, there isn't yet sufficient evidence that these methods can be trusted, suggesting caution before implementing them in high-stakes scenarios like medical transcription or accessibility tools.