The feature extraction based on texture image information for emotion sensing in speech.

Area of Science:

• Speech processing
• Computer vision
• Affective computing

Background:

• Emotion recognition in speech is crucial for human-computer interaction.
• Traditional methods often struggle with subtle emotional cues.
• Spectrograms offer a visual representation of speech, potentially containing emotion-related texture information.

Purpose of the Study:

• To propose a novel texture image feature for Emotion Sensing in Speech (ESS).
• To evaluate the effectiveness of this feature across different languages and corpora.
• To demonstrate the superiority of the proposed feature over existing methods.

Main Methods:

• Spectrograms of speech signals were transformed into recognizable images.
• Image contrast was enhanced using a cubic curve.
• Texture Image Information (TII) was extracted using Laws' masks.
• Support Vector Machine (SVM) was employed as a classifier.
• Cross-corpus evaluation was performed using EMO-DB, eNTERFACE, and KHUSC-EmoDB.

Main Results:

• The proposed TII-based feature significantly improved classification accuracy in ESS systems.
• The 2-D TII feature effectively discriminated between different emotions, complementing pitch and formant tracks.
• De-noising 2-D spectrogram images proved more manageable than de-noising 1-D speech signals.

Conclusions:

• The novel texture image feature (TII) shows significant promise for enhancing Emotion Sensing in Speech.
• Visual perception-inspired feature extraction offers a robust approach for analyzing emotional states in speech.
• The TII feature provides a valuable addition to existing speech emotion recognition techniques.