A study on kindergarten teacher–child interaction quality based on a multimodal emotion recognition model integrating speech and text
Year: 2026
Type: article
Abstract: Emotion recognition offers a route to quantifying interaction quality in early childhood education, yet existing methods suffer from unimodal input, poor real-time performance and weak generalization. This study proposes a speech-text multimodal model built on a Mamba encoder with a GatedxLSTM fusion mechanism, in which frame pooling and temporal masking resolve the temporal misalignment between modalities and a learnable gate adjusts modality contribution. Experiments on IEMOCAP and FAU-AIBO, together with an in-house KTC-Pilot corpus of authentic kindergarten interactions used for ecological validation, show that the approach outperforms mainstream baselines in F1, WA and UA. Three child-speech-aware adaptations, namely pitch and formant augmentation, silence-aware frame pooling and disfluency-tolerant text encoding, close the domain gap with authentic kindergarten dialogue, and modality-dropout training, cross-modal reconstruction of missing text and gating regularisation maintain robustness under incomplete inputs. A contamination audit against the CLAP pretraining corpora confirms that the gains do not arise from data leakage. A downstream evaluation using an Interaction Quality Index calibrated against expert observer ratings achieves a Pearson correlation of 0.692 with expert scores. An on-device deployment study on two embedded platforms further shows that the quantised framework sustains an end-to-end latency of 377 ms per utterance at a real-time factor of 0.075 and a board power of 9.6 W while retaining an F1 score of 0.706 on authentic kindergarten audio, and a stage-wise data-handling design keeps raw audio, spectral features and transcripts inside the classroom node so that only a scalar interaction quality value is written to persistent storage. Ablation and visualization analyses confirm the contribution of each structural component to recognition performance and interpretability.
Source: Scientific Reports
Authors: Jinjin Han, Xuejun Yuchi, Xing Qi
Institution: Weifang Engineering Vocational College
Language: English
Topic: Emotion and Mood Recognition
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
SDG: Quality education
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7215032987