Work

LaSO-Enhanced Prototypical Networks for Few-Shot Classroom Sound Classification and Event Detection

Year: 2026
Type: article
Abstract: Automated analysis of classroom audio using machine learning can classify and detect sounds tied to classroom activity, helping education researchers understand instructional patterns and student engagement, offering a practical alternative to in-person observation or manual review of multi-class recordings. Conventional deep-learning approaches require large training sets obtained by manually combing through long recordings to find sufficient examples of each sound event, which is impractical and may not generalize across schools and teaching contexts. This paper proposes a few-shot learning framework integrating Label Set Operations (LaSO) with Prototypical Networks for classroom sound classification and real-time event detection. Our approach extracts scalogram-based time-frequency representations from audio segments and employs three complementary LaSO transformation networks (intersection, union, subtraction) to synthesize enriched prototypes capturing complex inter-sample relationships. We extend prior work from a 3-class to a 5-class taxonomy and introduce a complete sound event detection (SED) pipeline for classroom recordings of arbitrary length. Experiments on a multi-school dataset with 5-fold cross-validation show that LaSO-ProtoNet achieved 96.7% accuracy on 5-way 5-shot classification. Cross-domain evaluation via the leave-one-school-out (LOSO) protocol demonstrated robust generalization, with 85.2% average accuracy on unseen schools. For classroom activity detection on untrimmed recordings, our classroom activity detection (CAD) pipeline achieved 95.6% segment-based F1, 91.3% intersection-based F1 (DTC=GTC= 0.5), and 87.3% event-based F1 at ±0.25s primary collar. Ablation studies confirm each LaSO component’s contribution, with the complete model showing consistent gains over baseline Prototypical Networks across all configurations.
Source: IEEE Access
Language: English
Open Access status: gold
Estimated APC paid: $2,075
OpenAlex ID: W7204965985
FWCI:
Cites: 43
Cited by:
Related to: 10