The Evaluation Method of Speech Interference Effect Based on Deep Learning
Hengfeng Fu, Sen Wang · 2020
Evaluating the effect of interference is critical for many communication systems, especially when faced with increasing threat of unintentional and intentional interference. In this work, we proposed two assessment models based on deep learning to illustrate the complex nonlinear relationship between utterance and impairment level. We evaluate their performance on a realistic dataset. A convolution neural network (CNN) based model is proposed to evaluate interference effect with the spectrogram. To avoid overfitting in a specific interference scenario, several transfer learning strategies and domain adaption are used to enhance the CNN model. The experimental result shows that the domain adaption based model provides the best accuracy on both the source domain (0.89) and the target domain (0.85).