Year: 2025
Type: article
Abstract: The study adapts several machine-learning and deep-learning architectures to recognize 63 traditional instruments in weakly labelled, polyphonic audio synthesized from the proprietary Sound Infusion collection. Ten thousand 5s clips were algorithmically generated, features such as Mel-spectrograms, MFCCs, and VGGish embeddings were extracted, and six models were evaluated. The re-implemented Han et al. Convolutional Neural Network (CNN) attained the best result (micro F1 = 0.55; macro F1 = 0.50), approaching published performance on mainstream Instrument Recognition in Musical Audio Signals (IRMAS) data. Results highlight data scarcity and class imbalance as key obstacles for culturally diverse MIR.
Source: Open Journal of Social Sciences
Authors: Sovathanak Meas, Rezza Moieni
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: diamond
OpenAlex ID: W4414397451