CoEDL / elpis

🙊 software for creating speech recognition models.
https://elpis.readthedocs.io/en/latest/
Apache License 2.0
151 stars 33 forks source link

Training fails when dataset contains empty annotations #345

Open benfoley opened 9 months ago

benfoley commented 9 months ago

During training, evaluation stage can fail when training dataset contains empty annotations.

ValueError: one or more ground truths are empty strings.

Quick fix is to set training settings min_duration to 1 rather than default 0.

Probably better to exclude empty annotations from dataset.