Mar 2023 – Apr 2024 · 1 yr 1 mo · Bengaluru, Karnataka, India · Remote

Oct 2022 – Feb 2023 · 4 mos · Bengaluru, Karnataka

Feb 2022 – Sep 2022 · 7 mos · Remote

Mar 2020 – Feb 2022 · 1 yr 11 mos · Remote

Oct 2017 – Sep 2019 · 1 yr 11 mos · Stuttgart Area, Germany · On-site

• 1. Style Modeling in End-to-End Speech Synthesis(2019): Design and implementation of novel
• style modeling approach based on adversarial autoencoder for learning latent representations for style control and transfer in end-to-end speech synthesis.
• Foundation for the End-to-End Speech Synthesis project at the Institut für Maschinelle Sprachverarbeitung (IMS).
• Current state-of-the-art implementations e.g. Global Style Token Tacotron and Variational Autoencoder Tacotron.
• Novel style modeling algorithm based on Adversarial Autoencoders.
• Built male and female voice models.
• Experiments on style control in terms of disentanglement, interpolation, and combination were
• performed and showed great promise for future research in speech augmentation.
• Experiments on style transfers were performed and the proposed model outperformed the existing
• state-of-the-art models in AB preferential tests.
• 2. Implicit Emotion Word Recognition Using Deep Learning(2018): Developed an implicit
• emotion word recognition model using convolutional neural networks(CNNs) for recognizing the emotional word absent in the sentence.
• The project was implemented for Implicit Emotions Shared Task’ (IEST)
• To predict the emotion of a tweet based solely on the context surrounding an emotion word that
• is absent from the sentence.
• Qualitative and quantitative analysis was performed on the model’s output.
• 3. Sentiment Analysis of Amazon Product Reviews Using Domain Adversarial Training(2017): Developed a sentiment classifier for Amazon product reviews while addressing the problem of Domain Adaptation.
• Various modeling techniques were implemented ie single domain-based and cross-domain adversarial training approach.
• Two different sets of training regimes were tested for the cross-domain adversarial training approach.
• Experiments on the effects of different types of activation functions and optimizers in the deep neural network models were performed and analyzed.

Jul 2016 – Sep 2017 · 1 yr 2 mos · On-site

• 1. Content and news platform (Indus Daily)
• a. Complete frontend and backend development for the platform.
• b. Managing the configuration of android client using firebase. This feature proved very handy for A/B testing.
• 2. Analytics SDK
• a. Developed data management android sdk for gathering analytics data from the device and sending it to backend server in controlled and optimised manner.
• b. New design led to the 85% reduction in cost as compare to previous version which was deployed in production.
• 3. Pincode and geotagging Module for location services
• a. Developed Rest api and implemented geospatial queries on mongo db to tag lat-long locations to their corresponding pincodes

Jul 2015 – Jun 2016 · 11 mos · On-site

• Text to Speech
• a. Implemented voice models for 6 Indic languages (Hindi, Marathi, Tamil, Telugu, Malayalam and Bengali) using HMM based training.
• b. Developed Android client with both offline and online voice support. Product shipped on more than 5 million devices across 20+ device models.
• c. Also developed an SDK for other Android applications to enable them to use our Text to Speech service within their own app.

Jan 2015 – May 2015 · 4 mos · Noida Area, India · On-site

• Project : L OS upgrades for US based devices
• a. Note 3, S4 &Tab S.
• b. Interacted and worked with multiple teams to perform end to end
• Development.
• c. Get familiarised and used Samsung software for creation of binary.
• d. Successfully performed SCAT, BVT, CSC checker and uploaded required
• updated binary to their proprietary server.
• e. Constantly communicated with the US team to understand requirements of their
• market.
• f. Complete documentation of the project.
• g. Technology Stack: - Samsung proprietary software, Android and its tools.