Responsibilities include:

• Pioneered a dual-microphone enhancement framework integrating internal and external ear microphone signals to improve intelligibility and mitigate speech drop in low-SNR conditions.
• Designed a low-parameter, streamable U-Net-based architecture with a cross-attention mechanism to fuse two-channel feature representations for real-time deployment.
• Collected and curated a real-world evaluation dataset using a dummy-head HAT recording system for robust, ecologically valid testing.
• Introduced SpeechDROPBERT, a novel SpeechBERT-based evaluation metric for quantifying speech-drop effects.
• Achieved DNSMOS OVRL 2.40 on the curated test set vs. 2.25 in a beamformer + speech enhancement baseline, and SpeechDROPBERT 0.88 vs. 0.60 baseline, demonstrating substantial perceptual and robustness gains

Responsibilities include:

• Developed TENNs, a state-space model optimized for the Akida SNN chip, enabling efficient multimodal processing.
• Designed aTENNuate, a real-time speech enhancement model (Interspeech 2025) and explored LoRA-based adaptation.
• Optimized LLM training and implemented a Triton-based GPU kernel for FFT convolution in signal processing.
• Developed secure inference & speaker verification, leveraging state-space models for enterprise applications.
• Developing TENNs TTS model

Responsibilities include:

• Developed diarization systems for multilingual, multi-speaker environments with code-mixing, automating 40 hours of conversational data annotation.
• Refined speaker detection by preprocessing audio and fine-tuning speaker activity detection using x-vectors.
• Enhanced speaker boundaries with VB-HMM clustering for improved diarization accuracy.
• Achieved DER of 28.04 for speaker diarization and 37.72 for language diarization on the DISPLACE dataset.