Most speech bots hear an Indian accent and completely lose their digital minds. I RESCUE them, turning raw, messy street audio into bulletproof training data.
From crisp custom pipelines to battle-tested Indic TTS and ASR, I micromanage every single syllable. The result?
Voice AI that actually understands the mother tongue without sounding like a broken microwave.
Data Scientist
Data Scientist working across the full speech pipeline: dataset construction, model fine-tuning, and evaluation, with a focus on Indic languages.
If you’ve ever tried getting an off-the-shelf speech model to understand rapid Hindi-English code-switching, you already know the pain: the algorithms are brilliant, but the data is a ghost town. That realization is basically how I ended up down the rabbit hole of building speech tech for Indic languages from scratch.
Early on, I realized you can tweak model architectures all day, but in this domain, clean corpora are so scarce that the dataset is the model. So instead of staying in the clean sandbox of modeling, I went end-to-end. I started out dealing with the unglamorous, ground-level stuff; collecting raw, messy audio in the wild, cleaning out ambient street noise, and designing the pipelines that turn scattered audio files into structured, high-fidelity training data. From there, I moved into fine-tuning TTS systems that actually sound human, and setting up rigorous human-annotated ground truths to stress-test ASR models so we aren’t just relying on vanity metrics.
These days, I’m putting that obsession to work as a Data Scientist at Datasetsmith.com, where my day job is building production-grade speech pipelines and dataset workflows designed to survive the real world. When I log off, I usually end up right back at the terminal and tinkering with voice agents, messing around with transliteration quirks, and deliberately pushing speech systems until they snap just to see what broke first.
End-to-end ASR pipelines with WER benchmarking across six platforms including AssemblyAI, Sarvam, Azure, and Meta Omnilingual. Custom XTTS V2 fine-tuning and OmniVoice voice cloning deployed for Indic languages and enterprise BFSI use cases.
Production pipelines covering segmentation, format conversion, denoising, speaker diarization, and DNSMOS-based quality scoring at scale. Handles both single-speaker TTS corpora and two-person conversational audio including podcasts and phone call recordings.
Studio-grade recording setups with HITL (Human-in-the-Loop) annotation: automated pipelines handle first-pass processing while human reviewers validate uncertain outputs before data enters training. Covers single-speaker TTS corpora, conversational datasets, and automated quality frameworks.
Beyond core speech work, domains I've built and shipped in.
Automotive-grade wake word systems for enterprise clients covering the full pipeline from data collection through signal processing to production deployment. Optimised for low-latency, always-on detection in resource-constrained edge environments.
Object detection and tracking with YOLO-based models for real-time and batch inference workloads. Includes frame extraction pipelines, annotation workflow management, and dataset curation for consistent model training.
Linux-native development with Docker containerisation, Azure GPU VMs for model training, and RunPod for speech processing pipelines including OmniVoice voice cloning. AWS and Cloudflare R2 for scalable audio storage across production workloads.
A working toolkit for speech data collection, model training, evaluation, and deployment.
A selection of speech and language projects covering synthesis, transliteration, and large-scale synthetic corpora.
End-to-end fine-tuning of XTTS V2 on a curated Hindi speech corpus — covering data collection, preprocessing, speaker normalisation, and GPU training on RunPod infrastructure. Produces natural-sounding Hindi synthesis with accurate phoneme rendering, built to handle the phonetic complexity of Indic language inputs.
A rule-based NLP engine that converts English text to Devanagari script at 86.6% accuracy, using a 7-tier lookup cascade from a curated 7,300+ word dictionary to a neural phoneme model. Handles edge cases like silent letters, suffix morphology, and multi-word phrases. Deployed as a Flask REST API with a live web interface.
A 15-language parallel synthetic speech corpus built with zero-shot voice cloning on OmniVoice — roughly 100 hours per language. Every language carries the same matched set of emotions and delivery styles: happy, sad, angry, storytelling and narration, news read, and commentary, so downstream models get consistent expressive coverage across the whole set rather than uneven per-language data.
My path into speech technology — from a diploma in computer science to building production speech systems full time.
Datasetsmith.com
Building production speech data pipelines and dataset workflows for TTS and ASR, with a focus on Indic languages. Work spans raw audio collection, automated processing, human-in-the-loop annotation, and evaluation against ground-truth transcripts.
Gradscaler Technologies Pvt. Ltd.
Worked on Python projects supporting speech model training, including data quality tooling and audio capture applications for TTS corpus collection.
ITM SLS Baroda University
Lateral entry admission following the diploma. Currently in the third year, studying alongside full-time work in speech technology.
ITM SLS Baroda University
Graduated with a CGPA of 8.26.
I'm open to roles, collaborations, and consulting in speech AI, audio data engineering, and production ML systems. If you're working on something in voice tech — especially for Indic languages — let's talk.
Whether you have a project in mind, want to discuss speech technology, or explore collaboration opportunities, I'd love to hear from you.