Automatic Speech-Based Smoking Status Identification

Identifying the smoking status of a speaker from speech has a range of applications including smoking status validation, smoking cessation tracking, and speaker profiling. Previous research on smoking status identification mainly focuses on employing the speaker's low-level acoustic features such as fundamental frequency (F-0), jitter, and shimmer. However, the use of high-level acoustic features, such as Mel Frequency Cepstral Coefficients (MFCC) and filter bank (Fbank) for smoking status identification, has rarely been explored. In this study, we utilise both high-level acoustic features (i.e., MFCC, Fbank) and low-level acoustic features (i.e., F-0, jitter, shimmer) for smoking status identification. Furthermore, we propose a deep neural network approach for smoking status identification by employing ResNet along with these acoustic features. We also explore a data augmentation technique for smoking status identification to further improve the performance. Finally, we present a comparison of identification accuracy results for each feature settings, and obtain the best accuracy of 82.3%, a relative improvement of 12.7% and 29.8% on the initial audio classification approach and rule-based approach, respectively.