IssueTracker

Issue: Google is asking their STT clients to migrate to new conformer models There currently is an "opt-out" option but in June, all clients will be forced to migrate to the new models. However, the accuracy of the new conformer models is, for some combinations of model and data types, “not better” than the accuracy of the legacy models.

In any case, once the migration opt-out flag is set for the models customers would like to use (i.e. phone_call) in the classic models.

What you expected to happen:

This issue is not about specific recognition problems but the overall model performance degradation and the forced migration.

• To remove the forced migration would allow them to use the classical models which they know work well in some contexts.
• To use the classic model, customers used to go to the Google Speech Console, in the "Preview Features" tab. There was a list of toggle that allowed them decide which classic model names would point to the new conformer models. The list is not there anymore so customers can't make any test.
• Customer want to know improvements made in the conformer models.

Please refer to the following link [1].

[1] https://cloud.google.com/speech-to-text/docs/conformer-migration