همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Beyond Semantics: A Perception-Based CNN Model for Quantifying Phonetic Rhythm in Speech

نویسندگان :

Mohammad Mahdi Peyravi¹ Alireza Talebpour² Zeynab Hajimohammadi³

1- دانشگاه شهید بهشتی 2- دانشگاه شهید بهشتی 3- دانشگاه شهید بهشتی

کلمات کلیدی :

Computational Phonetics،Speech Rhythm،Convolutional Neural Network (CNN)،Perception-Based Learning،Semantic Bias،Natural Language Processing

چکیده :

Quantifying subjective, prosodic features of speech, such as rhythm, remains a significant challenge in Natural Language Processing. A primary obstacle is "semantic bias," where a listener's perception of sound is heavily influenced by the meaning of the word. This paper proposes a novel methodology to overcome this challenge and develop a robust computational model for objectively quantifying phonetic rhythm. Our method involves three key steps: 1) Operationally defining rhythm within a continuous two-dimensional space (Percussiveness/Softness and Tempo/Slowness) based on phonetics. 2) Collecting perceptual ground-truth data from native speakers via an online survey. 3) Critically, to eliminate semantic bias, we designed and utilized a corpus of phonotactically valid meaningless sentences. These sentences were generated using a heuristic function to ensure full coverage of the rhythmic spectrum. Finally, a Convolutional Neural Network (CNN) was trained on this perception data, learning to predict rhythmic scores directly from phoneme sequences. Experimental results demonstrate that the CNN architecture significantly outperforms a baseline LSTM model, confirming our hypothesis that phonetic rhythm is a local phenomenon best captured by CNNs. This work provides a foundational tool for computational stylistics and advanced Text-to-Speech (TTS) systems.