Sequential Modeling by Leveraging Non-Uniform Distribution of Speech Emotion

The expression and perception of human emotions are not uniformly distributed over time. Therefore, tracking local changes emotion within a segment can lead to better models for speech recognition (SER), even when the task is provide sentence-level prediction emotional content. A challenge exploring sentence that most existing corpora only annotations (i.e., one label per sentence). This labeling approach appropriate leveraging dynamic trends sentence. We propose framework splits into fixed number chunks, generating chunk-level patterns. relies on rankers unveil pattern sentence, creating continuous curves. Our trains SER model with xmlns:xlink="http://www.w3.org/1999/xlink">sequence-to-sequence formulation by retrieved proposed method achieves best xmlns:xlink="http://www.w3.org/1999/xlink">concordance correlation coefficient (CCC) performance arousal (0.7120), valence (0.3125), dominance (0.6324) MSP-Podcast corpus. In addition, we validate experiments IEMOCAP MSP-IMPROV databases. further compare curves time-continuous traces. evaluation demonstrates these chunk-label effectively capture displaying time-consistency property similar traces annotated listeners. learns meaningful, complementary, information contributes improvement predictions attributes.