Speaker Change Detection Metrics Using Interval-Overlap Evaluation

The patent introduces an evaluation metric system that acts as an intermediary between the model predictions and the ground truth annotations. Instead of directly comparing predicted transition points with subjective annotations, the system uses precision and recall metrics that evaluate predictions based on their temporal proximity to annotated intervals. This intermediary evaluation framework reduces the impact of annotation subjectivity by allowing predictions within a tolerance window to be considered correct, thereby improving model performance consistency while maintaining the ability to identify speaker transitions.

The patent changes the evaluation parameter from exact time point matching to temporal interval overlap measurement. Instead of requiring predictions to match annotated transition points exactly, the system evaluates predictions based on whether they fall within or overlap with annotated time intervals. This parameter change allows the model to achieve both high measurement precision (by evaluating temporal accuracy) and high reliability (by tolerating reasonable variations in annotation timing), directly resolving the contradiction between precision and reliability.