Answers are grounded in this paper and cite their sources.
Machine learning algorithms, when trained on audio recordingsfrom a limited set of devices, may not generalize well to sam-ples recorded using other devices with different frequency re-sponses. In this work, a relatively straightforward method is in-troduced to address this problem. Two variants of the approachare presented. First requires aligned examples from multipledevices, the second approach alleviates this requirement. Thismethod works for both time and frequency domain represen-tations of audio recordings. Further, a relation to standardiza-tion and Cepstral Mean Subtraction is analysed. The proposedapproach becomes effective even when very few examples areprovided. This method was developed during theDetectionand Classification of Acoustic Scenes and Events(DCASE)2019 challenge and won the 1st place in the scenario with mis-matched recording devices with the accuracy of 75%. Sourcecode for the experiments can be found online
No takes yet. Share an insight, caveat, or question.
Michał Kosmider (2021) studied this question.