Built “Manushya-7B” and Designed “VoxClash”

I’m Indrayudh Mandal

ML Researcher at Deccan AI and Speech NLP Researcher at Lingo Labs, IIT Gandhinagar — working on speech models, evaluation systems, and Voice AI.

Latest news Projects Experience 2011–14 Era Gallery Get in touch

About Me

I am Indrayudh Mandal, an ML researcher at Deccan AI and a speech NLP researcher at Lingo Labs, IIT Gandhinagar, where I work under the guidance of Prof. Mayank Singh.

My work spans audio models, Automatic Speech Recognition for Indic languages, text-to-speech evaluation, and evaluation research for language and vision-language systems, with an emphasis on building robust pipelines that turn research ideas into usable tools.

News

  • [November 2026] Benchmarked Indic multi-party conversational diarization — three systems scored on 12 languages and 273 hours. Post.
  • [October 2026] Showcased a demo of Manushya-7B at Tech Expo 2026 during Amalthea, IIT Gandhinagar. Photos.
  • [October 2026] Built Manushya-7B, an open-weight full-duplex speech-to-speech model, at Lingo Labs, IIT Gandhinagar.
  • [October 2026] Wrote about VoxClash TTS 2.0 — the leaderboard now covers nine languages and ranks models by use case.
  • [September 2026] VoxClash by Deccan AI went live — a blind, human-preference arena for evaluating text-to-speech models.
  • [April 2026] Our instruction-following benchmark at Deccan AI is out — we evaluated multi-constraint prompts on frontier models, and constraint type predicted failure rate better than model choice.
  • [February 2026] Attended the India AI Impact Summit 2026 in Delhi on 18 February. Met peers from IIT Gandhinagar and Lingo Labs, and teams at Soket AI Labs and BharatGen.
  • [January 2026] Joined Deccan AI as an ML Researcher.
  • [December 2025] Attended IndoML 2025 at BITS Pilani Hyderabad Campus as part of the Graduate Forum, with colleagues from Lingo Labs.
  • [December 2025] Attended the Regional AI Impact Conference 2025 at Mahatma Mandir, Gandhinagar — an IndiaAI pre-summit for the India–AI Impact Summit 2026, hosted with IIT Gandhinagar.
  • [June 2025] Attended the Microsoft Research India Academic Summit 2025 virtually.
  • [February 2025] Part of Lingo Labs at CoLab 2025, IIT Gandhinagar’s Industry Open House, where the lab showcased work in responsible AI, AI for education, and NLP for India.
  • [January 2025] Part of Team Lingo Labs at the Curiosity Carnival at IIT Gandhinagar.
Earlier news

Experience

  • Deccan AI (January 2026 — present)

    ML Researcher

    Designed the evaluation plan for VoxClash, a blind human-preference TTS leaderboard across Indic and Indian-English voices.

    Worked on the instruction-following benchmark evaluating multi-constraint prompts on frontier models.

    Worked on trajectory-based workflow evaluation using a rule-based plus LLM-as-a-judge approach.

    Developed sanity and contamination checkers for Terminal-Bench tasks to improve evaluation reliability.

    Worked on a spatial reasoning evaluation for VLMs.

    Deccan AI

  • Lingo Labs, IIT Gandhinagar (2024 — present)

    Graduate Researcher / Speech NLP Researcher

    Built Manushya-7B, an open-weight full-duplex speech-to-speech model.

    Built an ASR system for Indic languages from scratch, along with TTS.

    Developed a cross-lingual STS model that processes audio from an Asian tonal language to English through a multi-stage pipeline.

    Lingo Labs

  • Wipro Technologies, Bengaluru (2022 — 2023)

    Software Developer

    Transformed legacy COBOL modules into efficient JCL scripts, reducing processing time by about 40%.

    Reduced defect leakage by around 30%, improving compliance and software quality.

    Wipro
What’s not in CV 2011–2014 game modding. Opens on its own page.

Research Interests

Speech and language
Automatic Speech Recognition Text-to-Speech Voice AI Audio Models Cross-Lingual Semantics Indic Languages
Evaluation
Agent Evaluation VLM Benchmarks Model Evaluation

Projects

VoxClash A blind, human-preference TTS leaderboard. Listeners compare model-versus-model clips; ranks come from Elo scores with issue tags for where a voice breaks down. TTS Evaluation Deccan AI Leaderboard · Writeup

Manushya-7B Built an open-weight full-duplex speech-to-speech model at Lingo Labs. Speech Lingo Labs Demo

Self Supervised ASR An end-to-end ASR pipeline built from scratch in PyTorch, with self-supervised pre-training using a Transformer encoder and CTC fine-tuning, focused on Indic languages. ASR Self-supervised PyTorch Lingo Labs GitHub

CoT-Llama3.3-70B Implemented a Chain of Thought reasoning framework using Meta's Llama 3.3 70B model through the Groq API. LLM Python PyTorch Prompt Engineering View repository

Education

  • Indian Institute of Technology Gandhinagar (2024 — present)

    M.Tech in Artificial Intelligence

    Research focus: NLP, ASR · Advisor: Prof. Mayank Singh

    IIT Gandhinagar

  • Chandigarh University (CU-IDOL) (2022 — 2024)

    Masters in Computer Applications

    Chandigarh University

  • Maulana Abul Kalam Azad University of Technology (WBUT) (2019 — 2022)

    Bachelors in Computer Applications

    MAKAUT

Teaching

Indian Institute of Technology Gandhinagar

  • CS613: Natural Language Processing, Fall 2024
  • ES114: Probability, Statistics and Data Visualization, Spring 2025
  • ES112: Computing
IIT Gandhinagar

Skills

Natural Language Processing Automatic Speech Recognition Machine Learning Deep Learning Python PyTorch TensorFlow Transformers Data Preprocessing Model Evaluation Research and Development Statistical Analysis Git and GitHub

Contact

I'm always open to discussing research opportunities, collaborative work, or connecting with people building thoughtful AI systems.

Email LinkedIn GitHub X