Speech Recognition Systems for Code Switched Speech (in collaboration with Microsoft Research)
Guide: Prof. Preethi Jyothi
Code-switching is a practice of using two or more languages in a single utterance. Due to lack of significant amounts of clean and labelled data, a robust speech recognition system can't be built directly. We explore techniques of using monolingual language models of the individual languages and combine them in a way to build a combined language model which can then be used to build a robust speech recognition system. Future work includes exploring usage of linguistic rules and encoding them to be able to generalize better.
Opponent Modelling in Scrabble
Guide: Prof. Shivaram Kalyanakrishnan
We look to design an inference agent to model the opponent's rack. Using information about the opponent's rack, the agent can play aggressively or defensively. We use simple Bayesian Inference based model to devise a probability distribution over possible racks. We impose further constraints by introducing information about previous opponent's racks. We use this information in forward Monte Carlo simulations to find better moves.
Defect Detection from Product Reviews
Guide: Prof. Pushpak Bhattacharyya (in collaboration with Accenture)
We aim to find the specific sentences from product reviews which directly delineate the defect of the product from the review. We first build a product ontology which captures the defect-related phrases and words. For any candidate test sentence, we use a pattern matching approach to find the similarity with the ontology. We use POS tags and word embeddings for better generalization.
Interactive Virtual Chat Bot
Philips Research, India
The aim was to develop a chat bot for medical applications. I worked on several parts in the same, one of the important ones included the problem of lip-syncing. This included studying various co-articulation models to obtain the best mapping of visemes with an additional constraint of ensuring smooth transitions. Apart, I also worked on building the framework for scene recording and generation dynamically.
Improving Language Models using Cross-Scripted Text
Microsoft Research, India
Cross-scripted text is the transliterated version of the native language text in some other language. These can act as secondary sources of data, especially in contexts where native language text is limited. We aim to exploit this data source using a robust transliterator and using the transliterated text to improve the language models for the native language.