Transfer learning for automatic ABRSM grade evaluation

his thesis explores the use of transfer learning for Automatic Performance Assess-ment, a research area that aims to develop software applications that can analyse and grade musical performances. Such applications can assist students in their learning process, especially in their daily practice. However, current Automatic Performance Assessment applications are not reliable and effective enough for professional music education.
A dataset of karaoke performances was created, consisting of Amazing Grace record-ings from Smule’s “Digital Archive of Mobile Performances”. The recordings were graded across the criteria of Pitch, Time, Tone, Shape and Performance, follow-ing the ABRSM grading guidelines. Hand-crafted features were extracted using Essentia, pYIN, DTW and vector distance measures as a baseline for performance evaluation. Features were also extracted from a music-tagging CNN and two audio compression VAE models: one trained on the same dataset and one trained on a different dataset. Linear kernel SVM classifiers were trained using these features and their test ROC AUC metrics were compared.
It was found that hand-crafted features performed better than NN-based features for Pitch, Time and Performance criteria, while NN-based features performed better than hand-crafted features for Tone and Shape criteria. The results suggest that transfer learning may be beneficial for some aspects of performance assessment.