Aug 2022 – Jan 2024 · 1 yr 5 mos · Bengaluru, Karnataka, India

• Worked on building in-house NLP applications and expanding the scope of NLP for our use cases.
• 1. Contributed to a project titled 'Joint Multilingual PostASR Correction & ITN', aimed at developing a single NLP model capable of handling both PostASR correction and ITN. This approach streamlines the process, eliminating the requirement for separate models for different tasks and languages.
• 2. Worked on a project named 'INTENDD: A Unified Contrastive Learning Approach for Intent Detection and Discovery'. INTENDD introduces a comprehensive framework for identifying and uncovering intents in dialogue utterances from Task-oriented Dialogue (ToD) systems. This research has been recognized and accepted at EMNLP 2023. For more details, the paper can be accessed at: https://aclanthology.org/2023.findings-emnlp.947/.
• 3. Our research initiative titled 'Scaling Neural ITN for Numbers and Temporal Expressions in Tamil: Findings for an Agglutinative Low-resource Language'. This project delves into various fascinating facets of ITN. It gained acceptance in the EMNLP Industry track for 2023. Additional information can be found at: https://aclanthology.org/2023.emnlp-industry.42/.
• 4. Engaged in multiple fundamental NLP challenges, including zero-shot intent detection and named entity recognition (NER), constructing retrieval-augmented generation (RAG) pipelines, online learning, etc.

Jan 2022 – Jul 2022 · 6 mos

• 1. Worked on building a speech recognition model which can transcribe millions of calls (which are multilingual) received per day between customers and delivery partners, between delivery partners and customer care, etc.
• 2. The ASR model that we built was based on Wav2Vec2.0 which is tuned on in-house and AI4Bharat data. We have also published a research paper which described a series of experiments (on pre/post processing, LMs, inference tricks, etc.) where our WER improves from 80%+ to ~30%, which is highly serviceable for the use-cases.
• 3. Use cases can be to analyze and understand these calls are fairly evident such as catching fraud, building product features to reduce unneeded calls, delivery partners training, understanding the sentiments of the delivery partner as well as the customer for assigning automatic ratings.
• 4. This research has been accepted in 24th International Conference on Speech and Computer, 2022.
• Paper link - https://link.springer.com/chapter/10.1007/978-3-031-20980-2_41

Apr 2021 – Nov 2021 · 7 mos · Singapore

• 1. Worked on creating an annotated dataset for anomaly detection in the energy consumption of buildings.
• 2. The dataset contains meter readings for one year from 1413 buildings at hourly timestamps.
• 3. Build a web-based tool for detecting anomalies in energy time series.
• 4. We have benchmarked the dataset on eight machine learning-based models.
• 5. The dataset along with the benchmarking results has been submitted to ARDUOUS workshop 2022 (https://text2hbm.org/arduous/).
• 6. Currently, we are working on extending the dataset for more than a year and building possible deep learning models for anomaly detection problems.

Nov 2020 – Oct 2021 · 11 mos · Bengaluru, Karnataka, India

• 1. Created multi-device corpus in neutral and whispered speech called "wSPIRE".
• 2. wSPIRE is so far the largest multidevice corpus for whispered speech.
• 3. We published the wSPIRE corpus as a database paper in the O-COCOSDA 2021
• (https://www.colips.org/conferences/cocosda2021/wp/)
• 4. Worked on building a deep learning model for the recording device classification using wSPIRE
• corpus.
• 5. Build a Dual Attention Pooling Network for recording device classification.
• 6. This work has been accepted in ICASSP 2022 (https://2022.ieeeicassp.org/)
• wSPIRE paper - https://ieeexplore.ieee.org/document/9660449
• wSPIRE corpus - https://spire.ee.iisc.ac.in/spire/wspire.php
• wSPIRE presentation - https://www.youtube.com/watch?v=qI_3eK9NrdE&t=104s
• ICASSP paper - https://ieeexplore.ieee.org/document/9747700
• Advisor: Dr. Prasanta Kumar Ghosh, Associate Professor, Department of Electrical Engineering.

May 2020 – Sep 2020 · 4 mos · Kathmandu, Bāgmatī, Nepal

• 1. Worked on Unsupervised Multilingual speech recognition and synthesis on code switched script for the low-resource language.
• 2. Project focused on three languages (Hindi, Nepali, and English).
• 3. Also responsible for the collection and creation of the database.
• Advisor: Suresh Manandhar, Director, Naamii AI Research Institute

Jan 2020 – Mar 2020 · 2 mos · Karnataka, India

• 1. Worked on Kannada Text-to-Speech Model.
• 2. Build a web-interface for our kannada Text-to-Speech.
• 3. Volunteered and presented a project titled, ‘Pitch-follower’ based on a time-frequency graph in a science exhibition organized at the IISc.
• TTS interface demo: https://www.youtube.com/watch?v=Y4KA4br3isE
• Advisor: Dr. Prasanta Kumar Ghosh, Associate Professor, Department of Electrical Engineering.