Advancing the frontiers of AI, audio processing, and natural language understanding through rigorous research and innovation.
We present Moûsai, a family of efficient text-to-music diffusion models that generate high-quality music from text descriptions. Our models achieve state-of-the-art results while being significantly more computationally efficient than previous approaches.
We introduce a novel approach for confidence estimation in speech recognition using OpenAI's Whisper model, enabling more reliable automatic speech recognition systems with calibrated confidence scores.
We present methods to significantly improve Whisper's ability to recognize rare and domain-specific words in zero-shot settings, addressing a key limitation of large speech recognition models.
We introduce CLadder, a comprehensive benchmark for evaluating causal reasoning capabilities in large language models, providing insights into their understanding of cause-and-effect relationships.
We investigate how language models handle moral exceptions and edge cases, exploring their alignment with human moral judgment and the nuances of ethical reasoning.
We present an approach for detecting hostile content in Hindi text using pre-trained language models, contributing to safer online spaces for regional language communities.
We introduce DARS, a novel approach for enhancing coding agent performance through dynamic action re-sampling and adaptive tree traversal strategies.
We develop an adversarial fine-tuning approach for classifying self-reported breast cancer posts on Twitter, improving healthcare-related information extraction from social media.