Advanced Methods for Speaker Verification and Low-Compute Speech Recognition

Abstract: Speaker verification requires better speaker representations and scoring mechanisms. The first doctoral seminar (Seminar I) addressed the limitations of conventional speaker verification pipelines through Transformer-based modeling. Novel speaker embeddings, termed s-vectors, was proposed using a Transformer encoder architecture to replace traditional TDNN-based embeddings. In addition, to overcome the shortcomings of statistical scoring backends such as PLDA, Transformer Encoder-based Speaker Authenticator (TESA), a neural scoring mechanism that directly models speaker similarity was proposed. The second doctoral seminar (Seminar II) extends this work along multiple complementary directions. Firstly, a new Indian multilingual speaker classification dataset, INX SpeakerHub, covering 10 major Indian languages, was created. ECAPA TDNN was trained with INX-SpeakerHub and speaker verification performance was analysed. Then, to address limitations of margin-based training using additive angular margin (AAM) loss in speaker verification under weak competitor conditions, a novel architecture-agnostic discriminative objective, Phantom AAM loss, was proposed. Next, s-vectors 2.0, an enhanced version of the previously proposed s-vectors, was introduced to further strengthen the speaker embeddings. Finally, the scope of the work was broadened to low-compute automatic speech recognition (ASR), where a parameter-efficient Transformer-based ASR architecture was proposed to reduce computational requirements while maintaining word error rates. Overall, the contributions of Seminar II strengthen the thesis by extending speaker verification research toward data resources, system-level evaluation, robust training objectives, enhanced speaker embeddings and resource-aware speech recognition.

Event Details
Title: Advanced Methods for Speaker Verification and Low-Compute Speech Recognition
Date: December 19, 2025 at 11:30 AM
Venue: ESB 244 / Google meet
Speaker: Ms. Metilda Sagaya Mary N J (EE18D013)
Guide: Dr. S Umesh
Type: PHD seminar