.png){kind=icon}
Return
Learning audio classification models on generated data using a sequential encoder enhanced self-supervised audio spectrogram transformer
.png){kind=icon}
PRE
AI
DOI：10.1007/s00530-026-02755-3.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Audio-classification tasks usually confront a lack of high-quality training samples especially for specific cases with rare categories of sounds. Hence, to ensure sufficient available data for learning audio-classification models, one may resort to audio generative models for obtaining training audio samples with supervision and variance. Nevertheless, current Transformer-based models for audio processing usually fail to optimise sequential relationships between patches in a spectrogram. In this regard, we propose a Sequential Encoder enhanced Self-Supervised Audio Spectrogram Transformer (SE-SSAST) for learning audio classification models on generated data. The proposed approach consists of a spectrogram Transformer module and a sequential encoder module, focusing on learning and patches’ representations and performing inter-patch encoding. Then, the experimental results for the models evaluated on real-world audio classification datasets indicate effectiveness for the proposed approach, through employing models learnt on generated audio data.
Keywords:
Audio classification
Generated data
Sequential encoder
Spectrogram transformer
Self-supervision
Journal
IF:
3.1
Papers:
2.8K
Citations:
2.7K
Organization
Cited Papers
ShareSaveShareSaveData augmentation on convolutional neural networks to classify mechanical noise.png)
APPLIED ACOUSTICS
APPLIED ACOUSTICSIF3.6
PREAI
ShareSaveShareSaveShareSaveGenerative emotional AI for speech emotion recognition: The case for synthetic emotional speech augmentationAPPLIED ACOUSTICS
OAAI
APPLIED ACOUSTICSIF3.6
OAAIShareSaveShareSaveShareSave
View more