FoleyBench: A Benchmark for Video-to-Audio Models
The first large-scale benchmark for Foley-style video-to-audio generation: non-speech, non-music sound that must align causally with the visuals.
I'm a PhD student at Carnegie Mellon University, advised by Chris Donahue and Bhiksha Raj.
My research is on multimodal intelligence, particularly audio, and on reliable ways to evaluate it.
I've worked on:
- Efficient audio-language models (Mellow)
- Benchmarks for audio understanding and generation (MMAU-Pro, FoleyBench)
- Evaluation metrics that align with human judgment (AURA Score, MACE)
Previously, I was a founding engineer at Cekura (YC F24), where I built the evaluation layer for voice agents, which now runs on 60K+ calls a day. Before that, I completed an MS at CMU, research internships at MIT and EPFL, and a BTech in electrical engineering from IIT Delhi.
Research
MMAU-Pro: A Challenging and Comprehensive Benchmark for Holistic Evaluation of Audio General Intelligence
A benchmark for audio general intelligence across speech, sound and music, including spatial, multi-audio, and long-form questions that need several reasoning steps at once.
AURA Score: A Metric for Holistic Audio Question Answering Evaluation
A metric for open-ended audio question answering that pairs LLM reasoning with an audio-entailment check. State-of-the-art correlation with human judgments on AQEval, a new 10k-response benchmark.
Mellow: A Small Audio Language Model for Reasoning
A small audio-language model that competes with much larger models on audio reasoning tasks, together with ReasonAQA, a large synthetic dataset for audio reasoning.
Vision Language Models Are Few-Shot Audio Spectrogram Classifiers
Evaluates whether vision-language models such as GPT-4o and Gemini can classify sounds from spectrogram images alone. Few-shot, they reach expert-level accuracy.
News
- Aug 2026Started my PhD at Carnegie Mellon University.
- Jun 2026AURA Score accepted to Interspeech 2026.
- Nov 2025FoleyBench accepted to ICASSP 2026.
- Nov 2025MMAU-Pro accepted to AAAI 2026.
- Sep 2025Mellow accepted to NeurIPS 2025.
- Jul 2025Envelope morphing accepted to WASPAA 2025.