Ashima Sood

Balancing token distribution can dramatically improve performance in underrepresented domains, challenging the assumption that more data is always better.

Fragile gains in low-resource ASR performance are exposed, with standard methods outperforming complex alternatives in Garhwali speech recognition.

Sarcasm detection in audio is revolutionized by a framework that quantifies prosodic incongruity, achieving state-of-the-art performance across multiple datasets.