Experiences

Led and mentored a team of three engineers in the design, development, and delivery of end-to-end speech solutions.
Trained multilingual STT model, leveraging a massive dataset of 100,000+ hours to achieve high-accuracy across diverse Indian languages and dialects.
Developed code-mixed STT models that handle seamless English-Indian language switching without compromising on quality, delivering accuracy comparable to monolingual benchmarks”
Trained Proof-of-Concept (POC) models for Speech-to-Speech (S2S) and Speech-to-Text Translation (S2TT) systems.
Optimized distributed training pipelines via DeepSpeed and PyTorch, orchestrating efficient GPU utilization for faster iteration on massive-scale datasets.

Designed and implemented scalable data preparation pipelines for large-scale speech and text corpora, including data cleaning, normalization, deduplication, tokenization, and efficient sharding for distributed training.
Led the transition from Hybrid to Transformer-based STT by benchmarking toolkits like NeMo, K2, and PyTorch, resulting in a 10% average relative improvement in accuracy across multiple languages.
Improved domain-specific accuracy by 15% through Parameter-Efficient Fine-Tuning (PEFT), enabling rapid model adaptation for specialized client requirements without the overhead of full fine-tuning.
Adapted pre-trained LLMs for Indic languages, enhancing linguistic coverage and model performance in multilingual and low-resource environments.
Worked extensively on tokenization strategies and language modeling workflows, to improve decoding and overall STT performance.

Trending Tags