Abstract
Quantifying auditory perception of blending between sound sources is a relevant topic in music perception, but remains poorly explored due to its complex and multidimensional nature. Previous studies were able to explain the source-level blending in musically constrained sound samples, but comprehensive modelling of blending perception that involves musically realistic samples was beyond their scope. Combining the methods of Music Information Retrieval (MIR) and Machine Learning (ML), this investigation attempts to classify sound samples from real musical scenarios having different musical excerpts according to their overall source-level blending impression.Monophonically rendered samples of 2 violins in unison, extracted from in-situ close-mic recordings of ensemble performance, were perceptually evaluated and labeled into blended and non-blended classes by a group of expert listeners. Mel Frequency Cepstral Coefficients (MFCCs) were extracted, and a classification model was developed using linear and non-linear feature transformation techniques adapted from the dimensionality reduction strategies such as Principal Component Analysis (PCA), Linear Discriminant Analysis (LDA), and t-Stochastic Neighbourhood Embedding (t-SNE), paired with Euclidean distance measure as a metric to evaluate the similarity of transformed feature clusters. Results showed that LDA transformed raw MFCCs trained and validated using a separate train-test data set and Leave-One-Out Cross-Validation (LOOCV) resulted in an accuracy of 87.5%, and 87.1% respectively in correctly classifying the samples into blended and non-blended classes. In this regard, the proposed classification model which incorporates "ecological" score-independent sound samples without requiring access to individual source recordings advances the holistic modeling of blending.