Leveraging LSTM Models for Overlap Detection in Multi-Party Meetings

Detecting and Counting Overlapping Speakers in Distant Speech Scenarios

“…Conversely, using for training FA-based labels, obtained on per-speaker headset devices, improves the performance of the algorithm even when testing it on manually labeled data. This differs from the approach in [17] where FA-based labels were used both for training and testing, a procedure which could possibly be subject to bias in evaluation if not double-checked with manual annotation.…”

“…We use our synthetic dataset to determine, in a controlled environment, whether FA can be considered as a reliable labeling procedure for the purpose of OSDC. In the past, FA-based labeling has been adopted in the CHiME-6 challenge for training and evaluation of diarization systems as well as in [17], with mixed results according to [19].…”

“…Supervised neural network based OSD systems have been shown to outperform more classical approaches [15][16][17]. Notably, Sajjan et al [17] have shown that a neural network based approach for joint VAD+OSD can outperform a Gaussian Mixture Model (GMM) based system on real-world distant speech data such as the AMI meeting dataset [18]. Recent work focusing on OSD only [19,20] has shown even more impressive results in near-field conditions.…”

“…Building on these previous works, we study supervised joint VAD+OSD and speaker counting in distant speech scenarios. We propose a Temporal Convolutional Network (TCN) architecture for these tasks, and evaluate it against previous works on joint VAD+OSD [17] and speaker counting [22] on AMI and CHiME-6 [14]. Because, to the best of our knowledge, supervised speaker counting has never been studied on real-world data, we also explore how the class imbalance problem can be mitigated using data augmentation.…”

“…Because, to the best of our knowledge, supervised speaker counting has never been studied on real-world data, we also explore how the class imbalance problem can be mitigated using data augmentation. Finally, we investigate the use of forced alignment (FA) as a labeling procedure for frontend speech segmentation applications, as used in [17] on AMI and in CHiME-6 [14]. This paper is structured as follows.…”