Machine Learning, Speech Processing, Data Science

Noor Fathima Khanum M.G.

Email: ************@****.*** k Mobile: +1-608-***-**** k www.linkedin.com/in/noor-fathima Skills

Languages: Python, C, C++, Bash scripting, MATLAB, Mathematica

Technologies/Frameworks: PyTorch, Tensor

ow, Keras, Kaldi, Festival, HTS-HTK, OpenCv, Docker, Flask Education

University of Wisconsin-Madison Madison, WI

Master of Science in Machine Learning and Signal Processing; GPA: 3.86 Sep. 2018 { May. 2020 Relevant Courses:Mathematical Foundations of Machine Learning, Matrix Methods in Machine Learning, Learning Based Methods in Computer Vision, Probability and Stochastic Processes, Security and Privacy in Data Science

Visvesvaraya Technological University Bangalore, India Bachelor of Engineering in Electronics and Communications Engineering Sep. 2011 { Jun. 2015 Relevant Courses: Random Processes, Speech Processing, Digital Signal Processing Research and Employment Experience

Qualcomm Corporate Research and Development San Diego, CA Systems Engineering Intern - Machine Learning (Mentor: Christopher Lott, Sr.Dir Engineering) May 2019 - Aug 2019 Transformer Neural Network k Compression k Quantization-aware training k Tensor Decomposition

Developed highly optimized neural network architecture and computation kernels for on-device execution.

Trained and optimized Transformer Neural Network for machine translation task in PyTorch framework.

Set up compression and quantization-aware training pipelines for Transformer model.

Explored tensor decomposition based compression techniques resulting in 10x compression with less than 10% performance drop.

AVIATR Lab, Waisman Center k University of Wisconsin Madison Madison, WI Project Assistant - Machine Learning (Mentor: Dr. Ender Tekin, Associate Scientist) Oct 2018 - May 2019 Transfer Learning k Image Classi cation k Generate image descriptors aiding disabled students

Used Docker containers to set up training pipeline for deep neural network based image classi cation.

Used Transfer Learning for image classi cation and achieved 12% increase in accuracy.

Used techniques such as Bagging and Boosting to overcome data imbalance problem.

Cogknit Semantics Bangalore, India

Research Engineer - Speech Systems Feb 2017 - Jul 2018 Automatic Speech Recognition (ASR) k Keyword Search System (KWS) k Video Tagging and Closed Captioning

Trained a variety of neural network based acoustic models (GMM-HMM, DNN-HMM, LSTM-CTC, TDNN) for speech recognition of low-resource Indian languages.

Implemented a lattice-based KWS system.

Wrote production-level code to implement inference engines for the speech algorithms.

Built web-services for speech algorithms using Flask and containerized them using Docker.

Text-to-Speech (TTS) Lab k PES Institute of Technology Bangalore, India Research Assistant - Speech Systems (Mentor: Dr. V Ramasubramanian) Jul 2015 - Dec 2016 Text to Speech System k Prosody Transplantation k Speech Data Collection and Annotation

Worked on segmentation techniques for data annotation in Festival platform.

Trained a TTS engine using HTS-HTK framework for Kannada - an Indian language.

Incorporated prosody for improving naturalness of synthesized speech using Unit-Selection framework. Publications

Noor Fathima, Tanvina Patel, Mahima C, Anuroop Iyengar \TDNN based Multilingual Speech Recognition for Low-Resource Indian Languages", InterSpeech 2018. 2nd place in the Microsoft Low Resource Speech Recognition Challenge

Tanvina Patel, Krishna D.N, Noor Fathima, Mahima C, Nisar Shah, Anuroop Iyengar, \Development of Large Vocabulary Speech Recognition System with Keyword Search for Manipuri" InterSpeech 2018

M. G. Khanum Noor Fathima, Mythri Thippareddy, M. Arunakumari, H.C Mamatha, H. N Supriya, A. Sricharan, V. Ramasubramanian, \Phonetically conditioned prosody transplantation for TTS: Unit granularity, context, prosody styles", in Oriental COCOSDA 2016; pp. 99-104. doi: 10.1109/ICSDA.2016.7918992

Mythri Thippareddy, Noor Fathima, D. N Krishna, Sricharan, V. Ramasubramanian, \Phonetically conditioned prosody transplantation for TTS: 2-stage phone-level unit-selection framework", in Speech Prosody 2016; Speech Prosody 2016 (2016): 781-785

D.N Krishna, M. G. Khanum Noor Fathima, Mythri Thippareddy, Sricharan, V. Ramasubramanian, \Phonetically conditioned prosody transplantation for TTS: A segmental unit-selection framework", in 13th IEEE Conference INDICON-2016

Projects

Adversarial Robustness: Implemented metric learning based defense mechanism to improve neural network’s resistance to adversarial examples generated using popular attacks such as FGSM and Carlini-Wagner.

Semi-Supervised Phrase Localization in Images (Arxiv: link): Developed a novel system utilizing CNN and LSTM neural networks. The network links visual regions in images to corresponding textual segments including phrases and words without explicitly providing supervision during training.

Texture Classi cation: Classi ed textures which appear very similar within data containing a highly imbalanced class distribution. Using statistical and segmentation based features, achieved 95% accuracy consistently.

Date Master Paintings: Developed an Image Processing technique to validate authenticity of paintings claimed to be works of renowned painters like Van Gogh and Vermeer. This was achieved by using weave maps of the paintings which in turn display the variation in thread width of the canvas.

Movie Genre Classi cation: Captured multi-modal (audio and image) information from movie trailers using pre-trained VGG19 for image features and CNN-based network trained on Google Audioset database for audio features.

Contact this candidate