Report on Current Developments in Bioacoustic Signal Processing and Music Emotion Recognition

General Direction of the Field

The field of bioacoustic signal processing and music emotion recognition is witnessing a significant shift towards leveraging advanced deep learning techniques, particularly in the context of transfer learning and self-supervised models. Researchers are increasingly focusing on cross-species transfer learning, where models pre-trained on human-generated sounds are being adapted to recognize and classify vocalizations of other species, such as bats and birds. This approach not only enhances the applicability of existing models but also contributes to a deeper understanding of out-of-distribution signal processing.

In the realm of underwater acoustic target recognition (UATR), the use of pre-trained models, both from the audio domain and general image recognition (ImageNet), is being explored to address the scarcity of labeled data. The findings suggest that while ImageNet models slightly outperform audio-specific models in this domain, the choice of pre-training and fine-tuning strategies significantly impacts model performance. This underscores the importance of tailored approaches for different data modalities.

Feature engineering and time-frequency representation remain critical for improving model performance, especially in complex acoustic environments. Recent studies highlight the impact of combining various time-frequency features, demonstrating that specific combinations can outperform single features. This approach is particularly relevant for bioacoustic signals, where the transformation of raw data into meaningful representations is essential for accurate classification.

The generalization capabilities of bioacoustic classifiers are also being extensively studied, with a particular focus on transfer learning methods and dataset characteristics. Fine-tuning and knowledge distillation are emerging as effective strategies, with cross-distillation showing promise in improving in-domain performance. However, shallow fine-tuning is found to be more robust for generalizing to complex soundscapes, emphasizing the need for balanced and comprehensive labeling practices in bioacoustic datasets.

In music emotion recognition (MER) and emotional music generation (EMG), the field is moving towards more objective evaluation metrics and diverse audio encoders to mitigate the inherent biases introduced by subjective assessments. The use of Frechet Audio Distance (FAD) and multiple encoders is proposed to provide a more robust measure of music emotion, enhancing both recognition and generation tasks.

Lastly, few-shot learning approaches are being adapted for bioacoustic event detection, with novel strategies for constructing negative prototypes and adaptive learning losses to improve model performance across varying task durations. These advancements are crucial for addressing the challenges posed by limited annotated data and varying vocalization durations in bioacoustic applications.

Noteworthy Developments

• Cross-species transfer learning in bat bioacoustics: Initial findings suggest that models pre-trained on human speech generate the most distinctive representations of bat song syllables, paving the way for improved out-of-distribution signal processing.
• Transfer learning in underwater acoustic target recognition: ImageNet pre-trained models slightly outperform audio-specific models in passive sonar classification, highlighting the potential of pre-trained models to address data scarcity in UATR.
• Generalization in bird sound classification: Cross-distillation and shallow fine-tuning emerge as effective strategies for improving in-domain and out-of-domain performance, respectively, in large-scale bird sound classification.
• Objective evaluation in music emotion recognition: The use of Frechet Audio Distance and diverse audio encoders demonstrates potential for mitigating emotion bias in both recognition and generation tasks.
• Few-shot learning for bioacoustic event detection: An adaptive learning framework with a negative selection strategy improves performance by 12.84%, addressing challenges in constructing negative prototypes and varying vocalization durations.