Francis Tyers
Francis Tyers
> PMC should give clear criteria and maybe a calendar I think having a calendar for changeover is an excellent idea.
Are there already recordings of these sentences? Have you cross-checked with the data in the latest release of Common Voice?
I may be wrong (we need to check with @mozgzh), but I think that if you remove sentences that already have recordings then we will lose access to those recordings....
Regarding WER: Common Voice does not train models, it only releases data, so we have no way of knowing -- aside from reports from the community -- if the WER...
Ok, so the majority can be removed. Those 135.537 should be left.
@ajay2110 if you still need this, please feel free to try out the PR #88.
@Olga-Yakovleva It's actually grapheme based, we found that for Chuvash a grapheme->phoneme model is not necessary, at least for training Ossian and Merlin. The input format is a `.txt` file...
You're welcome! (Тархасшӑн!) At least regarding Ossian/Merlin it works pretty well with Russian loan words. If you try something like "Манӑн мотоцикл пур, санӑн та мотоцикл пур-и ?" you will...