Interests: Machine Learning, Natural Language Processing, Speech Recognition

Papers & Preprints

  1. Stem-driven Language Models for Morphologically Rich Languages
    paper

    Yash Shah†, Ishan Tarunesh†, Harsh Deshpande, Preethi Jyothi
    (under review at ICASSP 2020)

  2. Seg-LM: A Factored-Output Language Model for Morphologically Rich Languages
    paper

    Yash Shah†, Ishan Tarunesh†, Preethi Jyothi
    (preprint; was submitted to EMNLP 2019)

† joint first authors

Research Projects

  • [WIP] Active Adversarial Accent Adaptation
    code

    Undergraduate Thesis with Prof. Preethi Jyothi, IIT Bombay, July 2019 - present

  • Self-supervised Representation Learning for Raw Audio
    report
    code

    Course Project (CS753) with Prof. Preethi Jyothi, IIT Bombay, July 2019 - November 2019

  • Exploring Online Algorithms for Causal Bandits
    report
    code

    Course Project (CS747) with Prof. Shivaram K., IIT Bombay, July 2019 - November 2019
    (one of the few teams that were awarded AA grade right away based on the project)

  • Recurrent Generative Models for Human Motion Synthesis
    report
    code

    R&D Project II (CS485) with Prof. Arjun Jain, IIT Bombay, Jan-May 2019

  • Exploring Hybrid Language Models for Morphologically Rich Languages
    report
    code

    Seminar (CS396) with Prof. Preethi Jyothi, IIT Bombay, July-Nov 2018

  • 3D Shape Recognition using Neuromorphic Tactile Sensing
    code

    Summer research internship with Prof. Alcimar Soares, SINAPSE Lab, NUS, May-July 2018

  • Alternate Loss Functions for Neural Language Modeling
    report
    code

    R&D Project I (CS490) with Prof. Preethi Jyothi, IIT Bombay, Jan-May 2018

Research Implementations

  • OpenASR-py: A minimal toolkit for end-to-end ASR and related tasks
    code

    OpenASR-py is a minimal, PyTorch based open source toolkit for end-to-end automatic speech recognition (ASR) related tasks, which borrows many elements from OpenNMT-py and at the same time provides simpler, task-specific reimplementations of several others. Due to the highly modular and transparent codebase, it can be used as a starting point for research projects in ASR as well as other less explored topics such as domain adaptation, adversarial training, active learning etc.

  • Discovering Latent Dependency Structure in VAEs
    report
    code

    Course Project (CS726) with Prof. Sunita Sarawagi, IIT Bombay, Jan-April 2019
    PyTorch implementation of Variational Autoencoders with Jointly Optimized Latent Dependency Structure by Jiawei He et al. that appeared in ICLR 2019. We also made two extensions to the approach proposed in the paper: (i) replacing the top-down inference modules by a LSTM network to model the dependency between latent variables, and (ii) extending the idea for sequential data along with deriving the corresponding ELBO term.

  • Language Modeling for Morphologically Rich Languages: Character-Aware Modeling for Word-Level Prediction
    code

    TensorFlow implementation of the paper by Daniela Gerz et al. that appeared in TACL 2018. My code is more focused towards three of the set of 50 languages experimented on by the authors, namely Hindi, Tamil and Kannada.

  • Generating Sequences With Recurrent Neural Networks
    report
    code

    Course Project (EE769) with Prof. Amit Sethi, IIT Bombay, Jan-April 2018
    TensorFlow implementation of the classic paper by Alex Graves on handwriting synthesis (sequence generation in general). We coupled the base model with a NLM so that sentences could be generated and written on the fly, given a prior context.