Vishal Keswani

        

×
Vishal Keswani ☰
  ABOUT PROJECTS PUBLICATIONS AWARDS        PORTFOLIO  

FULL TIME
[Check Portfolio-Resume for description]

INTERNSHIP
[Click Portfolio-Resume for description]

ACADEMIC
[Click Portfolio-CV for description]



×

Automatic Speech Recognition for interview transcription

@VMock India Pvt Lmt

  • Streamlined the 3 step ASR pipeline, removing third party dependency and saving upto $2k per month variable costs
  • Voice Activity Detection: compared low-latency vad models, used silero-vad to find no speech regions for audio splitting
  • Audio Transcription: developed in-house speech-to-text capability with Whisper, testing across sizes and frameworks
  • Phoneme Alignment: used wav2vec2 for generating time-stamps of text over audio for subtitling and targeted feedback
  • Reduced word error rate by 51% on technical transcripts and 29% overall, reduced latency by 20% using onnx format

ABOUT ME

Hey there! I am Vishal, Data Scientist at VMock Inc. I have around 3 years of experience in the field. Currently, my work is focused on Language Processing in the Career Acceleration space. I use the latest advancements in NLP and Speech Processing to help people in their professional journey. My days are spend working with deep learning models to automate solutions and perfecting the pipelines that ensure their functioning as robust technical products.

Since my JEE days I have had an inclination towards Mathematics and numbers in general. During my Bachelor’s at IIT Kanpur, I formed a solid foundation in Statistics and the use of Data in tackling real-world problems. Building onto the foundation, I did several courses and internships in Machine Learning and decided to join the data revolution. During my Master’s, I focused on working with unstructured textual data as the advent of Transformers called for my attention (pun intended).



PUBLICATIONS

Formulating Neural Sentence Ordering as the Asymmetric Traveling Salesman Problem (Long,Oral)
14th International Conference on Natural Language Generation (INLG'21)
[PDF] [PPT] [Talk] [Code]

Hypernym Detection in Financial Domain via Context-Free and Contextualized Word Embeddings
2nd Workshop on Financial Technology and Natural Language Processing (IJCAI'20)
[PDF] [PPT] [Talk] [Code]

Unimodal and Bimodal Sentiment Analysis of Internet Memes
14th Workshop on Semantic Evaluation (COLING'20)
[PDF] [Code]



TECHNICAL

VMock'23 Awarded STAR Employee for exceptional performance
FinNLP'20 Ranked 1st in the FinSim Shared Task on hypernym detection
SemEval'20 Stood 1st in SemEval Task on Sentiment Analysis of Memes
SURGE'18 Received Commendation Certificate for excellent performance
TA201A Bagged the Best Staff’s Pick Project Award in a batch of 380
JEE Mains Ranked in top 0.05% nationally, stood 1st in city (Ajmer)
JEE Advanced Ranked in top 0.15% nationally, stood 2nd in city (Ajmer)
CBSE Received Certificate of Merit from K.V.S. for standing in top 1.5%

EXTRA CURRICULARS

VMock'21 Won 2nd prize for best Coordinated Attire in Diwali'21
Galaxy'17 Secured the 1st position in British Parliamentary Debate
12th Became School Captain of K.V.1 Ajmer, also won Best Student Award
KVS Best Regional Speaker in Youth Parliament Competition'13-14 at Zonal level
10th Captained School Cricket Team of E.P.S. Ajmer, won Fair-Play Award
EPS Stood in top 150 in Tata Building India Essay Competition'11-12 in 2 million