You don't need _no_ errors, you just need low errors. Aside from Common Voice, there are also a lot of resources at openslr. Also, the amount of data you need

You don't need _no_ errors, you just need low errors. Aside from Common Voice, there are also a lot of resources at openslr. Also, the amount of data you need is often vastly overestimated, with advances in pretraining and transfer learning and the fact that most languages don't have as terrible an orthography as English.