Hello! I'm Pavan.
I do research in speech technology. I'm working towards a PhD with a focus on automatic speech assessment and recognition.
S Pavankumar Dubagunta
RESEARCH ASSISTANT
Email:
​
Office:
+41 277 217 714
Address:
Idiap Research Institute,
303-2 Centre du Parc, Rue Marconi 19,
Martigny CH-1920.
​
​
​
EXPERIENCE
2017-Present
Research Assistant
IDIAP RESEARCH INSTITUTE
Automatic speech assessment and recognition.
2015-2017
Senior Speech Engineer
INTERACTIVE INTELLIGENCE (now GENESYS)
I worked on building ASR models for six languages that went into the commercial product for small vocabulary ASR systems. I worked on deep learning techniques for improved efficacy and analysed ASR hypotheses to improve lexicons, LMs and phone definitions.
2015-2017
Lead Engineer
Senior Software Engineer
SAMSUNG R&D INSTITUTE INDIA
I worked on robust feature extraction techniques, implemented data selection techniques for ASR training, built and tested acoustic models using large data for multiple languages.
EDUCATION
2017-2021 (in progress)
Docteur ès Sciences (PhD)
ÉCOLE POLYTECHNIQUE FÉDÉRALE DE LAUSANNE (EPFL)
Automatic Speech Assessment and Recognition
2011-2013
Master of Science by Research
INDIAN INSTITUTE OF TECHNOLOGY MADRAS
Electrical Engineering.
Thesis: Feature normalisation for robust speech recognition.
2006-2010
Bachelor of Engineering
ANDHRA UNIVERSITY
Electronics and Communication Engineering.
LANGUAGES & SKILLS
Python - Proficient
Bash - Proficient
C++/C - Advanced
Kaldi - Proficient
Keras-Tensorflow / PyTorch - Advanced
MATLAB/Octave - Advanced
SELECT PUBLICATIONS
ASR: SEGMENTAL TRAINING
Hybrid neural network based automatic speech recognition (ASR) systems can be trained better using linguistic segment level confidence measure based objectives.
DIALECT IDENTIFICATION
Dialects can be identified better using knowledge from speech production and articulation, using raw signal based convolutional neural networks (CNNs).
DEPRESSION DETECTION
Voice source carries information indicative of a person's depression, which can be captured using raw signal based CNNs.