Dual-level segmentation method for feature extraction enhancement strategy in speech emotion recognition

Abstract

The speech segmentation approach could be one of the significant factors contributing to a Speech Emotion Recognition (SER) system's overall performance. An utterance may contain more than one perceived emotion, the boundaries between the changes of emotion in an utterance are challenging to determine. Speech segmented through the conventional fixed window did not correspond to the signal changes, due to the random segment point, an arbitrary segmented frame is produced, the segment boundary might be within the sentence or in-between emotional changes. This study introduced an improvement of segment-based segmentation on a fixed-window Relative Time Interval (RTI) by using Signal Change (SC) segmentation approach to discover the signal boundary concerning the signal transition. A segment-based feature extraction enhancement strategy using a dual-level segmentation method was proposed: RTI-SC segmentation utilizing the conventional approach. Instead of segmenting the whole utterance at the relative time interval, this study implements peak analysis to obtain segment boundaries defined by the maximum peak value within each temporary RTI segment. In peak selection, over-segmentation might occur due to connections with the input signal, impacting the boundary selection decision. Two approaches in finding the maximum peaks were implemented, firstly; peak selection by distance allocation, and secondly; peak selection by Maximum function. The substitution of the temporary RTI segment with the segment concerning signal change was intended to capture better high-level statistical-based features within the signal transition. The signal's prosodic, spectral, and wavelet properties were integrated to structure a fine feature set based on the proposed method. 36 low-level descriptors and 12 statistical features and their derivative were extracted on each segment resulted in a fixed vector dimension. Correlation-based Feature Subset Selection (CFS) with the Best First search method was applied for dimensionality reduction before Support Vector Machine (SVM) with Sequential Minimal Optimization (SMO) was implemented for classification. The performance of the feature fusion constructed from the proposed method was evaluated through speaker-dependent and speaker-independent tests on EMO-DB and RAVDESS databases. The result indicated that the prosodic and spectral feature derived from the dual-level segmentation method offered a higher recognition rate for most speaker-independent tasks with a significant improvement of the overall accuracy of 82.2% (150 features), the highest accuracy among other segmentation approaches used in this study. The proposed method outperformed the baseline approach in a single emotion assessment in both full dimensions and an optimized set. The highest accuracy for every emotion was mostly contributed by the proposed method. Using the EMO-DB database, accuracy was enhanced, specifically, happy (67.6%), anger (89%), fear (85.5%), disgust (79.3%), while neutral and sadness emotion obtained a similar accuracy with the baseline method (91%) and (93.5%) respectively. A 100% accuracy for boredom emotion (female speaker) was observed in the speaker-dependent test, the highest single emotion classified, reported in this study.