Hulme, Edward, Marshall, David ORCID: https://orcid.org/0000-0003-2789-1395, Sidorov, Kirill ORCID: https://orcid.org/0000-0001-7935-4132 and Jones, Andrew 2024. Acoustic classification of guitar tunings with deep learning. Presented at: DLfM 2024, Stellenbosh, South Africa, 27 June 2024. Proceedings of the 11th International Conference on Digital Libraries for Musicology. ACM, pp. 6-14. 10.1145/3660570.3660574 |
Preview |
PDF
- Accepted Post-Print Version
Download (1MB) | Preview |
Abstract
A guitar tuning is the allocation of pitches to the open strings of the guitar. A wide variety of guitar tunings are featured in genres such as blues, classical, folk, and rock. Standard tuning provides a convenient placing of intervals and a manageable selection of fingerings. However, numerous other tunings are frequently used as they offer different harmonic possibilities and playing methods. A robust method for the acoustic classification of guitar tunings would provide the following benefits for digital libraries for musicology: (i) guitar tuning tags could be assigned to music recordings; these tags could be used to better organise, retrieve, and analyse music in digital libraries, (ii) tuning classification could be integrated into an automatic music transcription system, thus facilitating the production of more accurate and fine-grained symbolic representations of guitar recordings, (iii) insights acquired through guitar tunings research, would be helpful when designing systems for indexing, analysing, and transcribing other string instruments. Neural networks offer a promising approach for the automated identification of guitar tunings as they can learn useful features for complex discriminative tasks. Furthermore, they can learn directly from unstructured data, thereby reducing the need for elaborate feature extraction techniques. Thus, we evaluate the potential of neural networks for the acoustic classification of guitar tunings. A dataset of authentic song recordings, which featured polyphonic acoustic guitar performances in various tunings, was compiled and annotated. Additionally, a dataset of synthetic polyphonic guitar audio in 5 different tunings was generated with sample-based audio software and tablatures. Using audio converted into log mel spectrograms and chromagrams as input, convolutional neural networks were trained to classify guitar tunings. The resulting models were tested using unseen data from disparate recording conditions. The best performing systems attained a classification accuracy of 97.5% (2 tuning classes) and 73.9% (5 tuning classes). This research provides evidence that neural networks can classify guitar tunings from music audio recordings; produces novel annotated datasets that contain authentic and synthetic guitar audio, which can serve as a benchmark for future guitar tuning research; proposes new methods for the collection, annotation, processing, and synthetic generation of guitar data.
Item Type: | Conference or Workshop Item (Paper) |
---|---|
Date Type: | Published Online |
Status: | Published |
Schools: | Computer Science & Informatics |
Publisher: | ACM |
ISBN: | 9798400717208 |
Date of First Compliant Deposit: | 2 July 2024 |
Last Modified: | 23 Jul 2024 21:39 |
URI: | https://orca.cardiff.ac.uk/id/eprint/170217 |
Actions (repository staff only)
Edit Item |