Few-Shot, Language-Agnostic Keyword Spotting (Smart India Hackathon 2024)

Recognition: Winner, Smart India Hackathon 2024 Grand Finale, problem statement 1680, set by the National Technical Research Organisation (NTRO), Government of India.

Problem area

The problem statement asked for a system that detects user-defined spoken keywords in audio streams. The streams may be in any language, including low-resource ones, and only a handful of recorded examples are available per keyword. This combines two hard settings in speech processing: few-shot learning (new keywords without large labelled datasets) and language-agnostic recognition (no dependence on a particular language’s acoustic or lexical model).

Outcome

Our team’s solution placed first at the Grand Finale. The problem statement was set by a government technical agency, so technical details of the solution are not disclosed here.