Skip to main content

Lund University Publications

LUND UNIVERSITY LIBRARIES

DMEL : THE DIFFERENTIABLE LOG-MEL SPECTROGRAM AS A TRAINABLE LAYER IN NEURAL NETWORKS

Martinsson, John LU and Sandsten, Maria LU (2024) 49th IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2024 In ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings p.5005-5009
Abstract

In this paper we present the differentiable log-Mel spectrogram (DMEL) for audio classification. DMEL uses a Gaussian window, with a window length that can be jointly optimized with the neural network. DMEL is used as the input layer in different neural networks and evaluated on standard audio datasets. We show that DMEL achieves a higher average test accuracy for sub-optimal initial choices of the window length when compared to a baseline with a fixed window length. In addition, we analyse the computational cost of DMEL and compare to a standard hyperparameter search over different window lengths, showing favorable results for DMEL. Finally, an empirical evaluation on a carefully designed dataset is performed to investigate if the... (More)

In this paper we present the differentiable log-Mel spectrogram (DMEL) for audio classification. DMEL uses a Gaussian window, with a window length that can be jointly optimized with the neural network. DMEL is used as the input layer in different neural networks and evaluated on standard audio datasets. We show that DMEL achieves a higher average test accuracy for sub-optimal initial choices of the window length when compared to a baseline with a fixed window length. In addition, we analyse the computational cost of DMEL and compare to a standard hyperparameter search over different window lengths, showing favorable results for DMEL. Finally, an empirical evaluation on a carefully designed dataset is performed to investigate if the differentiable spectrogram actually learns the optimal window length. The design of the dataset relies on the theory of spectrogram resolution. We also empirically evaluate the convergence rate to the optimal window length.

(Less)
Please use this url to cite or link to this publication:
author
and
organization
publishing date
type
Chapter in Book/Report/Conference proceeding
publication status
published
subject
keywords
adaptive transforms, audio classification, Deep learning, learnable Mel spectrogram, STFT
host publication
2024 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2024 - Proceedings
series title
ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings
pages
5 pages
publisher
IEEE - Institute of Electrical and Electronics Engineers Inc.
conference name
49th IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2024
conference location
Seoul, Korea, Republic of
conference dates
2024-04-14 - 2024-04-19
external identifiers
  • scopus:85195408870
ISSN
1520-6149
ISBN
9798350344851
DOI
10.1109/ICASSP48485.2024.10446816
language
English
LU publication?
yes
id
32c15a29-82d8-4558-9537-0ece70f446fc
date added to LUP
2024-09-12 14:27:35
date last changed
2024-09-12 16:20:13
@inproceedings{32c15a29-82d8-4558-9537-0ece70f446fc,
  abstract     = {{<p>In this paper we present the differentiable log-Mel spectrogram (DMEL) for audio classification. DMEL uses a Gaussian window, with a window length that can be jointly optimized with the neural network. DMEL is used as the input layer in different neural networks and evaluated on standard audio datasets. We show that DMEL achieves a higher average test accuracy for sub-optimal initial choices of the window length when compared to a baseline with a fixed window length. In addition, we analyse the computational cost of DMEL and compare to a standard hyperparameter search over different window lengths, showing favorable results for DMEL. Finally, an empirical evaluation on a carefully designed dataset is performed to investigate if the differentiable spectrogram actually learns the optimal window length. The design of the dataset relies on the theory of spectrogram resolution. We also empirically evaluate the convergence rate to the optimal window length.</p>}},
  author       = {{Martinsson, John and Sandsten, Maria}},
  booktitle    = {{2024 IEEE International Conference on Acoustics, Speech, and Signal Processing, ICASSP 2024 - Proceedings}},
  isbn         = {{9798350344851}},
  issn         = {{1520-6149}},
  keywords     = {{adaptive transforms; audio classification; Deep learning; learnable Mel spectrogram; STFT}},
  language     = {{eng}},
  pages        = {{5005--5009}},
  publisher    = {{IEEE - Institute of Electrical and Electronics Engineers Inc.}},
  series       = {{ICASSP, IEEE International Conference on Acoustics, Speech and Signal Processing - Proceedings}},
  title        = {{DMEL : THE DIFFERENTIABLE LOG-MEL SPECTROGRAM AS A TRAINABLE LAYER IN NEURAL NETWORKS}},
  url          = {{http://dx.doi.org/10.1109/ICASSP48485.2024.10446816}},
  doi          = {{10.1109/ICASSP48485.2024.10446816}},
  year         = {{2024}},
}