We gratefully acknowledge support from
the Simons Foundation and member institutions.
Full-text links:

Download:

Current browse context:

cs.SD

Change to browse by:

References & Citations

DBLP - CS Bibliography

Bookmark

(what is this?)
CiteULike logo BibSonomy logo Mendeley logo del.icio.us logo Digg logo Reddit logo ScienceWISE logo

Computer Science > Sound

Title: AST: Audio Spectrogram Transformer

Abstract: In the past decade, convolutional neural networks (CNNs) have been widely adopted as the main building block for end-to-end audio classification models, which aim to learn a direct mapping from audio spectrograms to corresponding labels. To better capture long-range global context, a recent trend is to add a self-attention mechanism on top of the CNN, forming a CNN-attention hybrid model. However, it is unclear whether the reliance on a CNN is necessary, and if neural networks purely based on attention are sufficient to obtain good performance in audio classification. In this paper, we answer the question by introducing the Audio Spectrogram Transformer (AST), the first convolution-free, purely attention-based model for audio classification. We evaluate AST on various audio classification benchmarks, where it achieves new state-of-the-art results of 0.485 mAP on AudioSet, 95.6% accuracy on ESC-50, and 98.1% accuracy on Speech Commands V2.
Subjects: Sound (cs.SD); Artificial Intelligence (cs.AI)
Cite as: arXiv:2104.01778 [cs.SD]
  (or arXiv:2104.01778v1 [cs.SD] for this version)

Submission history

From: Yuan Gong [view email]
[v1] Mon, 5 Apr 2021 05:26:29 GMT (402kb,D)
[v2] Tue, 6 Apr 2021 20:29:37 GMT (405kb,D)
[v3] Thu, 8 Jul 2021 20:16:28 GMT (798kb,D)

Link back to: arXiv, form interface, contact.