Loading the testimony.
Here is the problem
A speaker says a short sentence in Awadhi. The general-purpose model does not fail by getting a word wrong. It gives up and repeats one syllable until it runs out of room.
Before: a general speech model
After: retrained on Awadhi
The testimony
Recorded answers from Awadhi speakers describing the customs around birth, marriage and mourning. Each quotation is the corpus transcription, unaltered, and plays in the speaker's own voice. The English beneath it is machine translation, offered so the testimony can be followed; where the translator was not confident, none is shown.