This system implements a recognizer-based verification system.
To train a model using this, first make sure SpeechBrain is installed via pip. This will change when the toolkit is released soon, but for now it can be installed from test PyPI using the following command:
pip install speechbrain
Next, ensure you have access to the OGI kids' speech repository, say in /path/to/cslu_kids
You can run an experiment with:
cd verification/train
python train.py hparams/train.yaml --data_folder /path/to/cslu_kids
The different loss terms listed in hparams/train.yaml can be used to train a model for different tasks.
...
ctc_weight: 0.0
align_weight: 0.0
verify_weight: 1.0
...The ctc and align weight (together) can be used to train a system with reasonable alignments between evidence and posterior, as outlined in our paper (see citation below).
The verify weight can be used to train a model to make predictions about whether each utterance has an anomaly, based on the recording quality labels provided in the CSLU data.
@inproceedings{plantinga2019towards,
title={Towards Real-Time Mispronunciation Detection in Kids' Speech},
author={Plantinga, Peter and Fosler-Lussier, Eric},
booktitle={2019 IEEE Automatic Speech Recognition and Understanding Workshop (ASRU)},
pages={690--696},
year={2019},
organization={IEEE}
}