Decoding Speech: AI-Powered Acoustic Detection of Stuttering

Speech dysfluency, or stuttering, is a widely recognized speech fluency impairment that severely impacts communication and psy-chosocial well-being. Standard clinical assessment of stuttering is predominantly manual, making it subjective, labor-intensive, and prone to inter-clinician variability. To address these limitations, this paper presents an end-to-end intelligent framework for the automated acoustic detection of stuttering. The proposed solution leverages hybrid acoustic feature fusion, integrating Mel-Frequency Cepstral Coefficients (MFCCs), Linear Predic-tive Coding (LPC), pitch, formants, jitter, shimmer, short-time energy, and zero-crossing rate (ZCR). These multi-feature rep-resentations are modeled using state-of-the-art Deep Learning networks, comparing Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and Transformer architectures. Evaluated on the standardized UCLASS speech dataset under a Windows 11 system using Python 3.11 with TensorFlow and PyTorch, our proposed framework achieves an outstanding accuracy of 96.5%, precision of 95.8%, recall of 96.2%, and F1-score of 96.0%. These findings show that the integrated Transformer-LSTM pipeline can effectively support early diagnosis and real-time intelligent clinical screening.