Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition

Name: Vividh-ASR: A Complexity-Tiered Benchmark and Optimization Dynamics for Robust Indic Speech Recognition
Uploaded: 2026-10-23T10:21:09.066127+00:00
Duration: 425 s
Description: Reverse Multi-Stage Fine-Tuning (R-MFT) uses a parameter-efficient 244M Whisper model to outperform larger models on spontaneous Indic speech. The method employs a high initial learning rate and a hard-to-easy curriculum, combined with internal analysis showing adaptation focuses only on the decoder while preserving the encoder's robust acoustic features.

Reverse Multi-Stage Fine-Tuning (R-MFT) uses a parameter-efficient 244M Whisper model to outperform larger models on spontaneous Indic speech. The method employs a high initial learning rate and a hard-to-easy curriculum, combined with internal analysis showing adaptation focuses only on the decoder while preserving the encoder's robust acoustic features.

Open the full episode