Exploring Prosodic Features Modelling for Secondary Emotions Needed for Empathetic Speech Synthesis.

Area of Science:

• Speech synthesis
• Computational linguistics
• Affective computing

Background:

• Empathetic speech synthesis requires modeling secondary emotions, which are subtle and challenging to capture.
• Existing methods often rely on large datasets and deep learning, making them resource-intensive.
• Secondary emotions in speech synthesis remain under-explored.

Purpose of the Study:

• To develop a low-resource system for synthesizing secondary emotions in speech.
• To model subtle emotional prosody features using machine learning.
• To create an emotional text-to-speech (TTS) system capable of generating five distinct secondary emotions.

Main Methods:

• Handcrafted feature extraction and low-resource machine learning for prosody modeling.
• Quantitative model-based transformation for fundamental frequency contour shaping.
• Rule-based approaches for modeling speech rate and mean intensity.
• Development of an emotional TTS system synthesizing anxious, apologetic, confident, enthusiastic, and worried emotions.

Main Results:

• A proof-of-concept system for synthesizing secondary emotions in speech was successfully developed.
• The system utilizes a low-resource-intensive approach, avoiding the need for extensive emotion-specific databases.
• A perception test demonstrated that participants could identify synthesized emotions with a hit rate exceeding 65%.

Conclusions:

• The proposed low-resource approach is effective for synthesizing secondary emotions in speech.
• This method offers a viable alternative to resource-intensive deep learning models for emotional TTS.
• The system demonstrates potential for creating more nuanced and empathetic synthetic voices.