The limitations in dialect speech recognition include, among others, the complicated distribution of speech features, limited and diverse corpora, which influences the recognition of intelligent speech systems. The influence of nonlinear variation of dialect speech characteristics and the problem of cross-regional speech transfer is considered in this paper, by developing an algorithm of dialect speech feature extraction and smart recognition classification based on big data. The initial approach is to build a speech feature library through a large-scale speech dataset and to obtain a multi-dimensional feature representation with the help of a joint feature fusion mechanism comprising of Mel-Frequency Cepstral Coefficient (MFCC) and spectrogram. Second, a dual-channel Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) network are proposed to perform dual roles of extracting speech features and modeling temporal features respectively. Third, a mechanism of attention is used to improve feature weight distribution, thus making the model more sensitive to the minute differences in dialect speech. Lastly, a multi-class Support Vector Machine (SVM) is integrated to have intelligent recognition and classification. A self-constructed test set of samples of dialect speech were used in experiments. Findings indicated that the designed model had a recognition rate of 97.3% and an average feature extraction rate of 21.56%, which means that it has excellent generalization and real-time utilization.
Yaohui Yuan (Thu,) studied this question.