Diptanu Sarkar
hello [at] diptanu [dot] com

I am a computer science graduate student at the Rochester Institute of Technology. Currently, I am a thesis student affiliated to Language Technology Group under Dr. Marcos Zampieri. I am also a member of the Center on Access Technology research group under Dr. Micheal Stinson.

Earlier, I worked as a software engineering intern at Wayfair's platform engineering team. I've earned my bachelor's degree in electronics engineering from NIT Agartala, India in 2015. I also worked with the Infosys Engineering team in Bengaluru, India.

I will join Microsoft's Cloud and AI team as a Software Engineer post my graduation.

I am passionate about large-scale distributed systems and natural language processing.

In this study, we explore various triggers and targets of hate speech on social media. We propose state-of-the-art models for document and token-level aggression detection using transformers.

A model based on word importance in utterances using machine learning for Automatic Speech Recognition (ASR) systems to provide better phone captioning and accessibility to the deaf or hard of hearing (DHH) community.

Developed an automatic language identification model using the Bi-gram, Naive Bayes, Artificial Neural Network to detect ten different natural languages. The model is trained using the WiLI-2018 benchmark dataset, and the highest accuracy achieved on the test dataset is 99.7% with paragraph text.

Designed and developed an online easy text sharing tool. Started as a hobby project, now deployed for free.

Implemented a part-of-speech tagger in the English language using the Hidden Markov Model, Bayesian Net, and Naive Bayes. Then, compared the performance of the Forward-Backward Algorithm and the Viterbi Algorithm. The model resulted in over 91.2% word accuracy with 63.6% sentence accuracy.

Built image classification deep learning architectures - AlexNet, VGG16, and ResNet using transfer learning and fine-tuning in PyTorch. Final model accuracies achieved are AlexNet-81.2%, VGGNet-85.6%, ResNet-84.7% on 10K test images.

In this article, the importance of asymptotic analysis is explained, followed by the introduction to asymptotic notations. The worst, average, and best case time complexity analysis are also briefly discussed.

Music stimulates strong human emotions and feelings. Music platforms provide highly customized playlists to every user along with playlists based on moods. Emotions are subjective, and the subjective nature of emotions makes emotion detection a very challenging task when applied to music. Previously, music emotion detection solely relied on acoustic features. In recent studies, it’s observed that using music lyrics features along with acoustic features significantly improves the classification result.

Language Identification in Natural Language Processing is the process of identifying the spoken language in speech utterances. This blog examines three different models to recognize languages automatically - Dynamic Hidden Markov Networks model, Deep Neural Network model, and Long Short-Term Memory Recurrent Neural Network model.

Appreciate the aesthetics? Credit: Jon Barron.
Last updated: 01 Nov 2020