Applying Deep Learning Algorithms for Speech Recognition in Speech-Impaired Children

Abstract - Speech impairment affects millions of children worldwide, creating significant barriers to communication, education, and social development. This paper investigates the application of deep learning algorithms for automatic speech recognition (ASR) specifically adapted to the speech patterns of children with speech-language disorders. We evaluate and compare four deep learning architectures Convolutional Neural Networks (CNNs), Long Short-Term Memory networks (LSTMs), Transformer-based models, and hybrid CNN-LSTM frameworks trained and validated on augmented speech corpora drawn from children with articulation disorders, dysarthria, and stuttering. Mel-frequency cepstral coefficients (MFCCs) and spectrogram features serve as primary input representations. Experimental results demonstrate that the hybrid CNN-LSTM model achieves the highest word error rate (WER) reduction, reaching 78.4% recognition accuracy on the test set, outperforming conventional Hidden Markov Model (HMM) baselines by over 31 percentage points. Transfer learning from adult speech corpora, combined with child-specific data augmentation, further improves robustness to irregular phoneme production. The findings confirm that deep learning-based ASR holds substantial promise as an assistive technology for speech-impaired children, with practical implications for therapeutic tools and inclusive educational platforms. Key Words: speech recognition, deep learning, speech impairment, children, LSTM, CNN, transformer, assistive technology, MFCC, dysarthria.