Vocal Cord Vibration Signal Recognition Model Based on Feature Engineering Preprocessing

The vocal cord vibration signal carries as much rich phonetic information as the speech signal. However, conventional automatic speech recognition (ASR) technology cannot effectively deal with vocal cord vibration signal. In this article, a wearable throat speech detection sensor (TSDS) based on graphene sheets is designed. Subsequently, a large number of Chinese and English words are collected, and a dataset of vocal cord vibration for recognition tasks is established. Applying the knowledge of transfer learning, we use XGBoost applied to bearing fault detection as a source model, combined with feature engineering technology principal component analysis (PCA), and propose an extreme gradient boosting model fused with PCA for efficient identification of vocal fold vibration signals. The model reduces the data dimension, saves the operation time, and can automatically extract the characteristic parameters with a large contribution to the vocal cord vibration signal at the same time.