A Supervised Factorial Acoustic Model for Simultaneous Multiparticipant Vocal Activity Detection in Close-talk Microphone Recordings of Meetings

Using automatic speech recognition (ASR) word error rates (WERs) as a metric, the systems in (1) and (3) appear to have yield similar performance, in spite of significant additional architectural differences. Systems of type (2) have not been fielded for segmentation for ASR, and therefore cannot be directly compared. Although approaches of type (3) offer a significant advantage, namely the opportunity to directly constrain the number of simultaneously vocalizing participants, they come with the caveat of a variable acoustic vector size, since conversations/meetings can have variable numbers of participants. To overcome this difficulty, unsupervised acoustic models have been deployed [4], which do not require acoustic model training data (or training time). Our previous work has shown that this severly limits the number of features, as well as the minimum frame size. The aim of the current work is to develop a supervised acoustic model, capable of producing accurate density estimates for large feature vectors extracted from short frames, for scenario (3).