Maitreyi M

Led the development of a SOTA multilingual Text-to-Speech (TTS) system, using CUDA on H100 GPUs, spanning 9 languages (8 Indic) and powering key projects including real-time voice dubbing and speech LLM voice bots. Leveraged an emotion classifier to map text and speech embeddings, enabling automated selection of reference speakers in TTS generation. Contributed to cutting-edge speech systems including Language Identification (LID) Model, Automatic Speech Recognition (ASR) Model, and speech-based large language models (Speech LLMs). Worked on WebRAG and DocRAG; Continuous Pretraining (CPT) of Krutrim’s in-house large language model (LLM). Implemented end-to-end DevOps workflows for deploying trained TTS models in production, including API development, and automated CI/CD pipelines. Built and maintained large-scale data cleaning and processing pipelines for over 8 TB of speech and text data.

Analysed and trained on time series data to predict possible failures in a machine with 93% accuracy [predictive maintenance]. Worked on weather data to predict the solar irradiance the following day with 87% accuracy [irradiation prediction].

Trained deep neural network (DNN) to derive relationships between a person’s health and their traits. Handled firsthand the complexities of fairness and bias in data.