All learning paths Learn See Try Make Notice who is missing from it What this lesson is about Count the speakers by age, place and gender. The gap you do not count is the gap the model will inherit.
A hundred Kannada clips, taken over three weekends at one college. Your partner wants to send them tomorrow. What first?
A Ask the college for its class list, and copy the ages from that. B Count how many are men and how many are women, and send that along. C Count your speakers by age, place, gender and how they earn.
Check my recall Why this matters The survey behind this lesson names it: bias that comes from how you sampled a population. Older speakers, women and people without a smartphone are the ones most often missing.
Count who is in a collection and name, in writing, who is missing from it.
The survey behind this lesson names it: bias that comes from how you sampled a population. Older speakers, women and people without a smartphone are the ones most often missing.
Count your speakers by age, place, gender and how they earn. A gap you never counted becomes the gap the finished model carries. Report the gap with the collection instead of quietly fixing it later. Task A hundred Kannada recordings, collected over three weekends at one college.
Weak approach Reporting one hundred speakers and twelve hours, with no other count.
Stronger approach The same hundred, counted four ways. All are aged eighteen to twenty-five. Eighty are men. Every one is a student, and all live in one city.
Why the stronger approach works Both teams collected the same audio. Only the second one can tell you who the model will fail.
Try a changed situation The work is no longer speech. You are gathering Hindi posts from one app, used mostly by young men in cities. What do you do before sending it?
Make something yourself Build a coverage table for one real collection. Count the speakers by age, place, gender and work. Then write a paragraph naming who is missing and why.
Lesson 3 of 12 on Bharat-language data craft next Write a guideline others can follow Part 1 ends in a work sample you can send The rules this work sample is read against
Read the primary or official source: Mehrabi et al., Bias and Fairness survey (arXiv)