Siddharth Divi

Applied Scientist working on Speech Recognition. Hiker, backpacker, and explorer.

Hi, I’m Sid.

I’m an Applied Scientist working on Speech Recognition at Amazon, based in the Bay Area. My work sits at the intersection of machine learning and human-computer interaction — building the systems that let Alexa understand how people speak, across languages, accents, and acoustic conditions.

Before that, I was at Amazon Web Services, and before that, a grad student at Purdue University where I researched Federated Learning. I grew up in Chennai, India, did my undergrad at SSN College of Engineering, and made my way west through Indiana and eventually to California.

What I Work On

At Amazon’s Ambient Perceptual Technologies team, I work on keyword spotting, voice activity detection, and automatic speech recognition — the core components that power Alexa’s ability to hear and respond. A lot of my work involves making these systems more efficient, more fair, and more robust across the world’s many voices and languages.

Some things I’ve worked on:

  • On-device models that decide whether you’re talking to Alexa (DVAD) — shipped as the gating component for Alexa+
  • Keyword spotting for automotive environments with challenging acoustics
  • Fairness improvements across gender-based cohorts, reducing performance gaps for millions of users
  • Model compression via Quantization-Aware Training and Knowledge Distillation

Outside of Work

When I’m not building systems that listen, I’m out in places worth talking about — hiking, backpacking, and traveling. The Bay Area is a great base for it: Yosemite, the Sierra Nevada, and the Pacific Coast are all a few hours away.

Some favorites so far: the Emigrant Wilderness, Desolation Wilderness, Clouds Rest, Glacier National Park, Norway, Iceland, and Alaska.

Research

My academic work focused on Federated Learning — specifically, how to make FL systems more personalized and more fair. I introduced new metrics for evaluating FL fairness and proposed a framework that unifies knowledge distillation with personalization. This work was presented at the ICML 2021 Federated Learning Workshop.

Earlier work spans time series forecasting and species classification, published at IWANN 2019, CVC 2019, and CLEF 2018.

Let’s Connect