Pengenalan akor pada lagu utuh menjadi tantangan karena adanya campuran suara gitar, vokal, drum, dan instrumen lain yang membuat spektrum frekuensi tumpang tindih dan sulit dipisahkan. Untuk mengatasi masalah tersebut, penelitian ini mengembangkan sistem pengenalan akor gitar secara otomatis menggunakan pemisahan sumber suara dan klasifikasi citra spektrogram berbasis Deep learning. Dataset terdiri dari 2000 file audio akor gitar dalam 10 kelas (A, Am, Bm, C, D, Dm, E, Em, F, G) yang diambil dari dataset publik berdasarkan jurnal Osmalskyj et al. (2012), dan diaugmentasi menjadi 4000 file untuk meningkatkan keberagaman. Data dikonversi ke citra Mel Spectrogram berukuran 224×224 piksel dan dibagi menjadi 70% pelatihan, 20% validasi, dan 10% pengujian. Dua arsitektur diuji, yaitu CNN konvensional dan MobileNetV2 berbasis transfer learning. Selama pelatihan, MobileNetV2 mencapai akurasi 0.853 (train) dan 0.789 (val), dengan F1-score 0.79. Model terbaik digunakan untuk menguji data testing dan menghasilkan akurasi 0.81 serta macro F1-score 0.80, mengungguli CNN konvensional. Arsitektur akhir mencakup GlobalAveragePooling2D, BatchNormalization, Dense, dan Dropout, dilatih dengan Adam dan callback pelatihan. Pada pengujian nyata, sistem memisahkan trek gitar menggunakan Demucs, lalu Melakukan segmentasi per 3 detik dan klasifikasi akor. Antarmuka Gradio menampilkan hasil prediksi beserta confidence score. Hasil ini membuktikan kombinasi Mel Spectrogram, augmentasi data, Source Separation, dan arsitektur MobileNetV2 dalam pengenalan akor gitar
Chord recognition in complete songs is challenging due to the mixture of guitar, vocals, drums, and other instruments, which creates overlapping frequency spectra that are difficult to separate. To address this issue, this study develops an automatic guitar chord recognition system using Source Separation and Deep learning–based Spectrogram image classification. The dataset consists of 2,000 audio files of guitar chords in 10 classes (A, Am, Bm, C, D, Dm, E, Em, F, G) obtained from a public dataset based on Osmalskyj et al. (2012), to 4,000 files to increase diversity. The data were converted into 224×224 pixel Mel Spectrogram images and split into 70% training, 20% validation, and 10% testing. Two architectures were evaluated, namely conventional CNN and transfer learning–based MobileNetV2. During training, MobileNetV2 achieved an accuracy of 0.853 (train) and 0.789 (val), with an F1-score of 0.79. The best model was used for testing, resulting in an accuracy of 0.81 and a macro F1-score of 0.80, outperforming conventional CNN. The final architecture consisted of GlobalAveragePooling2D, BatchNormalization, Dense, and Dropout Layers, trained using Adam optimizer with training callbacks. In real-world testing, the system separated the guitar track using Demucs, then performed 3-second segmentation and chord classification. A Gradio interface was developed to display prediction results along with confidence scores. These findings demonstrate the effectiveness of combining Mel Spectrograms, data augmentation, Source Separation, and the MobileNetV2 architecture for automatic guitar chord recognition
