Singular Spectrum Analysis for Source Separation in Drone-Based Audio Recording
[EN]The usage of drones is increasingly spreading into new fields of application, ranging from agriculture to security. One of these new applications is sound recording in areas of difficult access. The challenge that arises when using drones for this purpose is that the sound of the recorded source...
| Autores: | , , , , , |
|---|---|
| Tipo de recurso: | artículo |
| Estado: | Versión publicada |
| Fecha de publicación: | 2021 |
| País: | España |
| Institución: | Universidad de Salamanca (USAL) |
| Repositorio: | GREDOS. Repositorio Institucional de la Universidad de Salamanca |
| OAI Identifier: | oai:gredos.usal.es:10366/169817 |
| Acceso en línea: | http://hdl.handle.net/10366/169817 |
| Access Level: | acceso abierto |
| Palabra clave: | Drone Audio recording Source separation Egonoise cancellation Singular spectrum analysis |
| Sumario: | [EN]The usage of drones is increasingly spreading into new fields of application, ranging from agriculture to security. One of these new applications is sound recording in areas of difficult access. The challenge that arises when using drones for this purpose is that the sound of the recorded sources must be separated from the noise produced by the drone. The intensity of the noise emitted by the drone depends on several factors such as engine power, propeller rotation speed, or propeller type. Noise reduction is thus one of the greatest challenges for the next generations of unmanned aerial vehicles (UAVs) and unmanned aerial systems (UAS). Even though some advances have been made on that matter, drones still produce a considerable noise. In this article, we approach the problem of removing drone noise from single-channel audio recordings using blind source separation (BSS) techniques, and in particular, the singular spectrum analysis algorithm (SSA). Furthermore, we propose an optimization of this algorithm with a spatial complexity of O(nt), which is significantly lower than the naive implementation which has a spatial complexity of O(tk2) (where n is the number of sounds to be recovered, t is the signal length and k is the window size). The best value for each parameter (window length and number of components used to reconstruct the source) is selected by testing a wide range of values on different noise-sound ratios. Our system can greatly reduce the noise produced by the drone on said recordings. On average, after the recording has been processed by our method, the noise is reduced by 1.41 decibels |
|---|