Optimization of the Music Appreciation Teaching Model under the Effect of Fine-Grained Emotion Analysis

This study introduces the extraction of fine-grained multimodal emotional features, optimizes the music appreciation teaching model in colleges, and enhances students’ recognition and understanding of musical emotion. Fine-grained in this study refers to the local analysis at the input representation level, not the expansion of the number of sentiment labels. The lyrics channel extracts WordPiece tokens, lyric sentences, and global semantic information in turn; the audio channel parses Mel spectrum, short-time energy, zero-crossing rate and spectrum flux according to time frame. After being hierarchically aggregated, local clues are mapped to four quadrants of happiness, sadness, anger and relaxation in the valence-arousal space. The four types of labels undertake the functions of classification supervision and teaching display; they do not constitute a fine-grained label system. The model adopts a dual-channel structure. Pre-trained Bidirectional Encoder Representations from Transformers (BERT) is introduced into the text end to complete contextual semantic coding and emotional guidance weighting. At the audio end, the Convolutional Neural Network (CNN) is used to extract the local time-frequency structure; then, Long Short-Term Memory (LSTM) is employed to model the inter-frame evolution relationship. Sequence-level bidirectional cross-attention is used for soft alignment between tokens and audio frames. There is no need to preset word-by-word timestamps, and the related lyric units and acoustic intervals can be connected through the global weight matrix. The public Million Song Dataset (MSD) is used for training and verification. In terms of teaching application, 215 first-year students in a university are selected to carry out a four-week music appreciation teaching intervention. The experimental group receives the assistant teaching of emotion recognition; the control group adopts the traditional teaching. The observable classroom participation performance, interaction situation and musical emotion understanding score are included in the teaching effect evaluation. The results show that the F1-score of the proposed method on MSD is 91.2%, higher than 82.7% of Bidirectional LSTM (BiLSTM) and 78.3% of multimodal support vector machine. After the teaching intervention, the average score of emotional understanding ability in the experimental group increases by 38.6% compared with that in the control group, and the frequencies of speech and gaze duration increase simultaneously. The findings indicate that the local emotion clues at tokens, sentences and audio frames can support the four-quadrant emotion judgment and provide structured feedback for the music appreciation class.