Speech Recognition using EfficientNet
Qidong Lu, Yingying Li, Zhiliang Qin, Xiaowei Liu, Yun Xie
Abstract
Aiming at the problem of speech recognition, this paper trains the latest convolutional network model EfficientNet with better performance to improve its recognition accuracy. To improve the generalization ability of the model, data augmentation operations such as noise addition, speed adjustment, pitch change are performed on the original speech waveforms. Because timedomain waveform can hardly describe the essential characteristics of an audio clip, Mel power spectrogram and its incremental characteristics are utilized as two-dimensional network inputs to accurately display the internal information of the signal. The experimental results show that method proposed in this paper can achieve high word classification accuracy according to the confusion matrix and accuracy variation curve of verification set on Google speech commands dataset.