When we set out to add AI interviews to RecruitKar, the "easy" version was obvious: transcribe English audio, score it, done. The hard, and far more useful, version was doing it in the languages Indian candidates actually speak. Here is what surprised us.
1. Code-switching is the norm, not the exception. Real Indian interview answers are rarely pure Hindi or pure English. They are Hinglish, Tanglish, a sentence that starts in Marathi and lands on an English technical term. A model that expects one language per utterance falls apart, so handling mixed-language input turned out to be the single biggest quality lever.
2. Accent and dialect variance is huge within one language. Tamil in Chennai versus Madurai, Hindi across states: the acoustic spread inside a single language is wider than the gap between some entirely different ones. Broad, real-world speech coverage mattered far more than clean studio data.
3. Scoring has to be language-fair. A candidate answering brilliantly in Telugu should not score lower than a mediocre English answer just because the pipeline understands English better. Separating what was said from how well the model heard it was essential to keep the evaluation fair.
4. Proctoring plus video changes the trust equation. Voice alone is convenient, but recorded, proctored video is what makes a hiring team actually trust a first round enough to skip it themselves.
The payoff: a recruiter can screen 50 candidates across five languages overnight and wake up to scored transcripts, without a single scheduling call. For small Indian teams hiring without a recruiter, that is the difference between a two-week and a six-week hire.
If you are working on multilingual speech or hiring tech, I would love to compare notes. And if you are hiring in India, that is exactly who we built RecruitKar for.
Top comments (0)