Your language filter can't see a sister language: data lessons from a low-resource build
In the first post of this series I wrote that Crimean Tatar had no dataset good enough for serious work, and that I started by fixing an existing speech set. In the second a recogniser learned to hear our speech,…