The Test of Auditory-Vocal Affect (TAVA) dataset.

Area of Science:

• Linguistics
• Psychology
• Computer Science

Background:

• Speech Emotion Recognition (SER) research often struggles to disentangle linguistic content from affective vocal cues.
• Existing datasets may not adequately isolate paralinguistic information for focused analysis.
• Vocal affect plays a crucial role in communication, particularly for individuals with language impairments.

Purpose of the Study:

• To introduce the TAVA dataset, a novel resource for Speech Emotion Recognition (SER).
• To facilitate research by disentangling paralinguistic (affective) and linguistic information in speech.
• To enable studies on affect perception in clinical and non-clinical populations.

Main Methods:

• Developed the TAVA dataset comprising 352 audio recordings of emotionally expressive English speech.
• Created transformed electroglottographic (tEGG) versions of speech signals to suppress phonetic content while preserving affective cues.
• Collected over 120,000 crowd-sourced ratings for valence, arousal, and dominance on both original and tEGG signals.

Main Results:

• The TAVA dataset provides paired audio and tEGG signals, allowing for direct comparison of affect perception.
• Crowd-sourced ratings offer detailed insights into how affective cues are perceived across different signal types.
• tEGG signals effectively isolate vocal affect, demonstrating their utility in SER research.

Conclusions:

• The TAVA dataset is a valuable resource for advancing SER by offering phoneme-reduced, affect-rich speech representations.
• This dataset can support research into vocal affect sensitivity in populations with communication difficulties.
• It enables dissociation of linguistic and affective processing, broadening SER applications.