Prateeth Nayak

I am currently a Senior ML-Engineer at Apple Inc., where I am involved in improving, scaling and optimizing of Voice Trigger System for Siri across all Apple Devices.

Previously, I was a researcher/computer-scientist at Stanford Research Institute (SRI) International, where I worked on cross-disciplinary machine learning & deep learning applications. At SRI Intl. I've worked on Model Quantization/Pruning (SRI spin-off LatentAI), AMD and DASL.

I completed my Masters at Univ. of Minnesota, Twin Cities and was a part of Prof. Mingyi H. research group.

Email  /  Bio  /  Google Scholar

• [June 2024] Work on new learning technique & architecture for keyword-spotting "RepCNN: Micro-sized, Mighty Models for Wakeword Detection" accepted at InterSpeech'24.paper
• [June 2023] Work on Voice-Trigger Mitigation model "Audio-to-Intent Using Acoustic-Textual Subword Representations from End-to-End ASR" accepted at ICASSP'23.paper
• [Sept 2022] Work on personalized Voice-Trigger "Improving Voice Trigger Detection via Metric Learning" accepted at InterSpeech'22.paper
• [Apr 2021] Joined Apple Inc. as ML Engineer at Siri !
• [Dec 2020] Molecule-Encoding paper accepted at Machine Learning for Molecules @ NeurIPS 2020
• [Nov 2020] Preprint of work "Zero-Shot Learning with Knowledge Enhanced Visual Semantic Embeddings" available. preprint
• [Oct 2020] Preprint of work "Transformer based Molecule Encoding for Property Prediction" available. preprint
• [Feb 2020] Joined full-time at SRI Intl. as Computer Scientist, AIC Group, San Diego !
• [Dec 2019] Graduated Masters in ECE from Univ. of Minnesota, Twin Cities
• [Dec 2019] Paper accepted at Efficient Machine Learning & Cognitive Computing @ NeurIPs 2019.

I'm interested in applied research on Model Quantization/Pruning, Foundation Model optimizations, and on-device ML. I delve much into cross-disciplinary projects.

We introduce a new architecture and approach to training a model for keyword detection for very low compute and memory hardware. This involves the trick of over-parameterizing the model during training and re-parameterizing during inference resulting in 2x lesser peak memory usage 10x faster runtime on-device.

We improve the voice trigger detection performance of smartphone assistant using small number of utterances from target speaker in a multi-task learning framework to output a personalized embedding for each utterance.

We introduce a modified Transformer based architecture to take in atom-level feature vectors as inputs and aim to encode a a better molecule representation space which helps effectively include the structure-property relations of a molecule.

We present the comparison of data-free post-training quantization methods for state-of-art models trained on datasets like ImageNet. To better analyze quantization results, we describe the overall range and local sparsity of valuesafforded through various quantization schemes.

• Stanford Research Institute (SRI) Intl., San Diego CA [Feb'20 - Mar'21]
  Computer Scientist
  Advanced Analytics/Aritifical Intelligence Center (AIC) group; Research focus on Molecular-Property prediction, DASL, and Zero-Shot Learning.


• Stanford Research Institute (SRI) Intl., Princeton NJ [Feb'19 - Aug'19]
  Deep Learning & Optimization Intern
  Computer Vision Group; Research focus on Post-Training and Training-Aware DNN Model Quantization techniques. Developed framework for post-training quantization for weights-only 8-bit or lower precision for previously trained in Tensorflow, Darknet(Yolo)


• Univ. of Minnesota, Twin Cities [Aug'18 - Dec'18]
  Graduate Teaching Assistant
  Responsible for grading, preparing and teaching course materials for Linear Control Systems EE-4231


• Boston Scientific, Arden Hills MN [May'18 - Aug'18]
  Machine Learning Intern
  Corporate Research Team; Developed end-to-end machine learning pipelines for analysing time-series data obtained from wearable devices of Chronic-Pain patients to find correlation between the physiological signals and reported pain scores.