This study investigates the use of the parametric trajectory model to perform unsupervised speech detection and segmentation in a noisy air traffic control audio. The process of detecting and segmenting speech is subdivided into two primary tasks: the binary distinction of speech and noise, and the ability to identify the beginning and end of speech segments. For each of these two tasks, the parametric trajectory model algorithm is applied in both a model-based (prior training) and a blind (no training) approach. The model is also trained created completely unsupervised. The results show that the parametric trajectory model can be applied to detect and segment speech in noisy audio with a high degree of success using either approach. While the model approach provided significantly fewer false alarms, the addition of a simple heuristic to the blind approach effectively produces the same level of performance when measured using the F-measure.
展开▼
