Transkripcija kitarskih tablatur z globokimi transformerskimi modeli

Automatic guitar tablature transcription remains an under-explored problem in automatic music transcription due to limited real-world annotated data with fretboard annotations and the instrument-specific complexity of assigning tones to a specific fretting position. In this thesis, a transformerbased sequence-to-sequence model for direct audio-to-tablature transcription is presented, inspired by MT3. The model predicts tokens from a vocabulary tailored for guitar tablatures without relying on intermediate sheet music representations. A dedicated tablature tokenization schema is introduced, and a systematic ablation study is conducted to analyse the impact of input audio representation (CQT versus Mel spectrogram), input window length, and vocabulary design. The results indicate that transcription accuracy is strongly influenced by input window length and vocabulary complexity, while the choice of audio representation has minimal effect. The proposed approach achieves performance comparable to existing guitar transcription systems and surpasses them on several evaluation metrics, while highlighting remaining challenges related to dataset limitations, training methods, and generalization.