Music Audio Sentiment Classification Based on CNN-BiLSTM and Attention Model
Zhen Chen, Changhui Liu · 2021
In order to improve the accuracy rate of sentiment classification of music audio, a neural network model combining CNN and BiLSTM and attention mechanism is proposed. First, Librosa is used to extract the preliminary audio features of songs. Those are suitable for audio sentiment classification are selected through comparative experiments. Second, the obtained preliminary feature matrix is taken as the model input, and CNN is used to further extract its local features. Then BiLSTM is used to obtain global features and attention model is used to weight different feature vectors to highlight important feature vectors. Finally, softmax is used for classification. This paper applies the combination model which has good effect in text to audio field, the experimental results show that the classification accuracy rate reaches 85.4%, which is significantly higher than that of the single model and has a better classification effect