Breath-based augmentative and alternative communication (AAC) offers a vital communication method for people with severe motor impairments when other input methods are unsuitable. Within the context of AAC, current machine learning techniques for classifying breath patterns are often slow or difficult to retrain on devices with limited computational resources. This study presents a deep learning-based convolutional neural network breathing feature classification (CNN-BFC) methodology for breath-based AAC using a continuous wavelet transform of breath signals, with time–frequency features extracted using a CNN. Various classification methods were explored in this study for comparison, such as utilizing transfer learning from a pretrained network, retraining a CNN network from scratch, and existing dynamic time warping (DTW) classification methods. Tested with data from 25 healthy subjects at Loughborough University, it was found that using a pretrained CNN network as a feature extractor for a support vector machine (SVM) gave the lowest training time while maintaining high accuracy compared to transfer learning and DTW. The SVM classifier using CNN feature extraction achieved a low training and prediction time of only 8.95 and 0.29 s per subject respectively, significantly outperforming the training time of the transfer learning method at 195 s, and the prediction time of DTW at 34.9 s. • Research of a real-time CNN-based classifier for breath pattern recognition in AAC. • Prediction time of 0.29 s outperforms DTW on diverse breath patterns. • Rapid retraining (9 s) of CNN-TL-SVM at 85% accuracy for per-user AAC adaptation. • Investigation of data augmentation and CNN architectures suitable for multimodal AAC.
Simatwo et al. (2026) studied this question.