Audio Transfer learning

Fine-tuning image models to tackle domain shift and class imbalance with PyTorch and torchaudio in audio data

Over the past few years, the study of environmental sound classification (ESC) has become very popular due to the intricate nature of environmental sounds. This paper reports our study on employing various acoustic features aggregation and data enhancement approaches for the effective classification of environmental sounds. The proposed data augmentation techniques are mixtures of the reinforcement, aggregation, and combination of distinct acoustics features. These features are known as spectrogram image features (SIFs) and retrieved by different audio feature extraction techniques. All audio features used in this manuscript are categorized into two groups: one with general features and the other with Mel filter bank-based acoustic features. Two novel and innovative features based on the logarithmic scale of the Mel spectrogram (Mel), Log (Log-Mel) and Log (Log (Log-Mel)) denoted as L2M and L3M are introduced in this paper. In our study, three prevailing ESC benchmark datasets, ESC-10, ESC-50, and Urbansound8k (Us8k) are used. Most of the audio clips in these datasets are not fully acquired with sound and include silence parts. Therefore, silence trimming is implemented as one of the pre-processing techniques. The training is conducted by using the transfer learning model DenseNet-161, which is further fine-tuned with individual optimal learning rates based on the discriminative learning technique. The proposed methodologies attain state-of-the-art outcomes for all used ESC datasets, i.e., 99.22% for ESC-10, 98.52% for ESC-50, and 97.98% for Us8k. This work also considers real-time audio data to evaluate the performance and efficiency of the proposed techniques. The implemented approaches also have competitive results on real-time audio data.

We’re on a journey to advance and democratize artificial intelligence through open source and open science.