Undersökning av ljud-till-MIDI-transkriptionsgeneralisering : Utvärdering av en pianotranskriptionsmodell på gitarr

Automatic Music Transcription (AMT) systems are typically trained on specific instruments, raising questions about their ability to generalize across instruments with differing acoustic properties and playing techniques. This study investigates how well the MT3 model—originally trained for piano transcription—performs on acoustic guitar recordings. Using the GuitarSet dataset, we evaluate transcription accuracy through Onset F1 and Onset & Offset F1 scores. The results reveal a substantial performance gap: while MT3 achieves high accuracy on piano (Onset F1: 95.95%; Onset & Offset F1: 83.46%), its performance drops considerably for guitar (Onset F1: 57.00%; Onset & Offset F1: 16.94%). This decline does not correlate strongly with polyphony or note duration, suggesting that the underlying cause lies in differences in timbre and articulation between piano and guitar. Further analysis of monophonic excerpts reveals the model’s tendency to produce spurious notes, likely due to its sensitivity to transient, guitar-specific sounds such as string noise, which differ from the more percussive nature of piano onsets. These findings highlight the limitations of cross-instrument generalization in AMT and point to the need for instrument-specific adaptation—whether through model fine-tuning, architectural changes, or more diverse training data. Future work could explore adaptation strategies, incorporate note velocity information, analyze the effect of spectral differences and extend evaluations to additional instruments such as electric guitar to improve the generalizability of AMT systems.