Publication

Abstract : Automatic identification of guitar-playing technique is important for music analysis, identification of performers, and interactive learning. This paper presents a hybrid Transformer-classifier model that utilizes sophisticated spectrogram-based feature extraction techniques to improve classification accuracy. Short-Time Fourier Transform (STFT), Mel-Frequency Cepstral Coefficients (MFCC), Constant-Q Transform (CQT), Continuous Wavelet Transform (CWT), and Gammatone spectrograms are used in order to extract the fine-grained frequency and time-domain features of the guitar sound. These spectrograms are processed through Vision Transformers (ViTs) and Swin Transformers, which extract spatial and temporal features in an efficient manner. These extracted features are subsequently classified through ML classifiers. The experimental demonstrates that the MFCC delivers the highest classification accuracy when coupled with Vision Transformer and Swin Transformer. More precisely, with SVM and MLP classifiers, both models record an accuracy of 97 %. In the case of Linear SVM, Vision Transformer outperforms marginally with a 98 % accuracy, and Swin Transformer records 97 %. These findings highlight the high performance of hybrid Transformer-classifier models in guitar technique recognition without compromising computational efficiency.