Speech Emotion Recognition Using Transfer Learning and Adaptive Timeframe Segmentation

Emotion recognition from speech is a key component of affective computing, enabling machines to interpret human emotional states through vocal cues. However, achieving robust performance in real-world conditions remains a significant challenge due to variability in speech duration, background noise, and emotional expression. This paper presents an efficient method for adapting the pretrained YAMNet model—originally designed for short audio event classification—to the task of sentence-level speech emotion recognition. Traditional fixed-frame approaches often struggle to capture the uneven distribution of emotional cues across an utterance and typically require extensive overlapping, which increases computational cost. To address this, we propose a non-uniform adaptive timeframe segmentation strategy that divides each sentence into a fixed number of windows, regardless of its duration. This method achieves better sentence-level error rates without compromising computational efficiency. We evaluate our method on the Berlin Emotional Speech Database under clean and noisy conditions, including real-world and reverberated noise scenarios. Compared to the best-performing fixed-overlap baseline (85% overlap), our method achieves a 2.76% relative error rate reduction with 29% lower computational cost. When matched for complexity, it yields up to a 9.5% error reduction. These improvements are statistically significant and consistent across different signal-to-noise ratios, demonstrating the robustness and efficiency of the proposed adaptation strategy for real-world emotion recognition applications.