How Does a Deep Neural Network Look at Lexical Stress in English Words?

A new study reveals how deep neural networks process English lexical stress, achieving up to 92% accuracy in predicting stress patterns from speech. Using interpretability techniques, researchers found the model focuses on specific acoustic features—primarily the first and second formants of stressed vowels—to make its decisions. This demonstrates AI's ability to learn complex, distributed phonetic cues directly from natural speech data, moving beyond highly controlled laboratory stimuli.