Training data does not improve initial Custom Speech model compared to the baseline model

[KatieProchilo]

Using data provided by the Custom Speech team, the first Custom Speech model that users train will not improve compared to the baseline model when we train it using data in the training folder.

This results in the pipeline rightfully failing when the word error rate does not improve, but this means no releases will be created, which is a huge loss for users who want to learn about that.

Solution

The training data at that link should improve recognition against the evaluation test data (audio + human-labeled transcripts) in the testing folder at that link.

[KatieProchilo]

This has been filed as CSE feedback.