Baseline and dataset generation are in ESPNet2 chime7_dasr recipe
This repository contains forced alignment segmentation for the CHiME-6 dataset, produced as in https://github.com/nateanl/chime6_rttm (details are in the CHiME-6 challenge description paper [1]).
Important
Here we extend the forced alignment annotation also for the training set, and provide it also in the form of
JSON files.
JSON forced alignment segmentation has a dummy value for the words entry (due to licensing we can't release the transcriptions here). The words field is not removed as it makes these more convenient to use in the baseline scripts e.g. in https://github.com/espnet/espnet/blob/master/egs2/chime7_task1/asr1/local/get_lhotse_manifests.py if one wants to use this segmentation in place of the manual one for GSS and ASR.
[1] Watanabe, S., Mandel, M., Barker, J., Vincent, E., Arora, A., Chang, X., et al. CHiME-6 challenge: Tackling multispeaker speech recognition for unsegmented recordings. https://arxiv.org/abs/2004.09249