MDT-AI090 Spontaneous Speech Datasets of Expressive Paralinguistics

Compared to emotional expressions, paralanguage demonstrates continuous and context-dependent variations during conversations, making it inherently more challenging to describe. To enhance the ability of large-scale models to effectively learn human paralanguage features, Magic Data has developed a novel “High-Quality Spontaneous Speech Datasets of Expressive Paralinguistics” This dataset is the result of a collaboration between the company's experienced product specialists and senior voice synthesis consultants, involving meticulous design, iterative refinement of the labeling framework, and production through a professional data pipeline. It is characterized by a high sampling rate of 48 kHz, recordings collected in controlled quiet environments, and contributions from a demographically diverse pool of speakers varying in region, age, and gender to ensure speech diversity. The dataset encompasses a wide array of topics across 20 domains, including daily life, leisure and entertainment, education and training, and healthcare.