Combining Symbolic Music Representation and Empirical Performance Data for Analysis of Solo Vocal Performances
Lead Research Organisation:
University of Glasgow
Department Name: School of Engineering
Abstract
The aim of this project is to perform automatic note segmentation of monophonic music by using both symbolic representation of music (score) and empirical data from audio recordings. The results of this will be stored in a database which can then be queried to gain technical information about how a performer plays or sings.
Organisations
People |
ORCID iD |
| Nicholas Bailey (Primary Supervisor) | |
| Keziah Milligan (Student) |
Studentship Projects
| Project Reference | Relationship | Related To | Start | End | Student Name |
|---|---|---|---|---|---|
| EP/N509668/1 | 01/10/2016 | 30/09/2021 | |||
| 1652899 | Studentship | EP/N509668/1 | 01/10/2015 | 18/04/2019 | Keziah Milligan |
| Description | If we wish to analyse musical performances - to investigate the aspects of the performance itself or an audience's perception of it - we first need a method for measuring performances. This method must be capable of yielding data about the smallest details: the precise time a note begins, the pitch of the note and how the pitch changes over the course of the note. From this information, a performer's pitch and timing accuracy can be found, as well as the pitch and time features which result in certain perceptual effects. This project seeks to create a reliable and accurate automatic technique for collecting these data on a large scale. By using a biologically-inspired model, I have successfully created nonlinear filters in software, which allow time-frequency analysis of a signal at a greater resolution than the uncertainty principle would suggest possible. When driven by an input - in this case, musical audio - a filter (or 'detector' as we refer to them) will respond when its centre frequency is close to a frequency present in the input. This means we can know both the frequencies in the audio and the times at which they appear because we are not trying to measure time and frequency directly; they are part of the state of the system, which we can access directly. The output of this 'DetectorBank' software can then be analysed to identify note events, a category which includes both the onset times of notes and changes which occur over the course of a note (for example, pitch changes due to vibrato). An automatic method for classifying these events into note onsets and intra-note events would be the next step in this software development. |
| Exploitation Route | Further work would involve extending the software to also track frequency continuously. This could be achieved using the same data (DetectorBank output) as the onset detection software. Additionally, if the software can be optimised to work in real time, this may be of use in other fields of engineering which also come up against the general problem of detecting the precise time a frequency component appears in a signal. For example, fault detection in mechanical systems often uses spectral analysis. |
| Sectors | Digital/Communication/Information Technologies (including Software) |