Actually, the technical report is designed to try and give a preliminary answer those very questions! I’ll send you a link to it in your email and perhaps you could give me some feedback?
Practically for Turkish:
- The error rate is too high to make pretranscription faster than transcribing from scratch for a native speaker.
- The model could be used for forced alignment, with a fairly high error rate, but given enough data and a bootstrapping process you should be able to improve this fairly rapidly
- For keyword spotting (e.g. doing voice search in a large audio archive, the Turkish model will give you reasonable precision, but lower recall, e.g. you will find examples of what you are looking for, but perhaps not all of them)
- For closed vocabulary (digits 0-9 + evet, hayır) the error rate is around 10%, so for every 10 words you speak into it, one will be wrong.
It’s worth noting that for Turkish there was very little final training data, only 1:52:31. You could get better results by not using the default splits of training-development-test.
Another key observation is that it is not good to have the training set unbalanced towards one speaker. The more speakers in the training set the better. For Basque for example there was 10:51:34 hours of training data and 53 speakers. For Tatar there was 11:49:15 but only 2 speakers.