Deep Learning Model for Chinese Dialect Speech Recognition Based on Key Vector Optimized Temporal Features

We propose a deep learning model for Chinese dialect speech recognition based on key vector de-optimization of temporal features. Unlike traditional Hidden Markov Models (HMMs) and their variants, the model uses stacked convolutional and long-short-term memory networks to extract features from the processed speech data, and is trained using an end-to-end approach, while the key vector module is used to dynamically extract key vectors to optimize the model features. Our proposed model achieves a word error rate (CER) of 2.802\% on the XXX dataset, which is 47.213\% lower than the GLU-based speech recognition model, and 0.632\% lower than the CNN- and RNN-based models.