Multilingual Query-by-Example KWS for Indian Languages using Transliteration

Query-by-Example Keyword Spotting (QbE KWS) detects query audio within target audio. A common approach for multilingual QbE KWS uses phoneme posteriors as representations, with a shared phoneme dictionary across languages. We propose a novel method that replaces phoneme-based representations with transliteration, unifying transcripts from multiple Indian languages into the Devanagari script, a text script used for Hindi and Marathi. We train a Multilingual ASR model to predict transliterated Devanagari text from audio across 10 Indian languages. The character logits from this ASR serve as both query and target audio features. Using the Kathbath dataset for training and the IndicSUPERB QbE evaluation set, our approach achieves significant improvements. The average MTWV increased from 0.015 (IndicSUPERB) to 0.504, and performance rose from 0.387 to 0.504, surpassing the best-performing Marathi ASR baseline. This demonstrates the effectiveness of transliteration for multilingual KWS.