Publications

2026

1. EMNLP

   

   VIBE: Video Instruction-aligned Background Music Generation

   Aryan Vijay Bhosale, Vaibhavi Lokegaonkar, Vishnu Raj, and 5 more authors

   In EMNLP (Findings) 2026, 2026

   Abs arXiv Website

   A text-and-video-to-music generation model that brings semantic control and instruction adherence to video-to-music. VIBE introduces a depth-wise cross-layer conditioning mechanism to overcome the representational bottleneck of static cross-modal conditioning in diffusion-autoregressive architectures, together with a reward-modeling taxonomy that optimizes both verifiable constraints (tempo, key) and subjective qualities (musicality, multimodal alignment) through a five-stage training curriculum.

   A text-and-video-to-music generation model that brings semantic control and instruction adherence to video-to-music. VIBE introduces a depth-wise cross-layer conditioning mechanism to overcome the representational bottleneck of static cross-modal conditioning in diffusion-autoregressive architectures, together with a reward-modeling taxonomy that optimizes both verifiable constraints (tempo, key) and subjective qualities (musicality, multimodal alignment) through a five-stage training curriculum.

2. NeurIPS

   

   DuplexWorld: Can Voice Agents Help You Get Through the Day?

   Aryan Vijay Bhosale, Harshit Rajgarhia, Akhil Pothanapalli, and 3 more authors

   In NeurIPS (RTCA) ’26, 2026

   Abs arXiv Website

   A benchmark for holistically evaluating speech-to-speech voice agents beyond database tool-calling. DuplexWorld spans six domains where voice agents are especially useful, with agents evaluated on eleven conversation types across 156 scenarios (350+ hours of dialogue), each probing conversational and analytical ability.

   A benchmark for holistically evaluating speech-to-speech voice agents beyond database tool-calling. DuplexWorld spans six domains where voice agents are especially useful, with agents evaluated on eleven conversation types across 156 scenarios (350+ hours of dialogue), each probing conversational and analytical ability.

   
3. EMNLP

   

   TORUS: A Test of Rendering-Understanding Self-Coherence for Unified Audio Models

   Aryan Vijay Bhosale, Harshit Rajgarhia, Abhishek Mukherji, and 1 more author

   In EMNLP 2026 Workshop on Speech and Audio Language Models (SALMA), 2026

   Abs arXiv Website

   The first self-coherence test for audio-native unified models, asking whether a model’s generation and understanding heads agree about the same audio rather than evaluating each capability in isolation. TORUS comprises 48 three-stage self-coherence tests with 432 six-option questions spanning speech, sound, and music across five task families.

   The first self-coherence test for audio-native unified models, asking whether a model’s generation and understanding heads agree about the same audio rather than evaluating each capability in isolation. TORUS comprises 48 three-stage self-coherence tests with 432 six-option questions spanning speech, sound, and music across five task families.

4. NeurIPS

   

   Exploring Textual Guidance for Video-to-Music Generation Models

   Vaibhavi Lokegaonkar, Aryan Vijay Bhosale, Vishnu Raj, and 5 more authors

   In NeurIPS 2026 Creative AI Track, 2026

   Abs arXiv Website

   A text-conditioned video-to-music generation model for fast, high-quality, semantically and stylistically controllable background music creation. An autoregressive module aligns visual and textual inputs for high-level musical planning, and Diffusion Transformers refine the plan into high-quality audio.

   A text-conditioned video-to-music generation model for fast, high-quality, semantically and stylistically controllable background music creation. An autoregressive module aligns visual and textual inputs for high-level musical planning, and Diffusion Transformers refine the plan into high-quality audio.

2024

1. Preprint

   

   Anomaly detection using Diffusion-based methods

   Aryan Bhosale, Samrat Mukherjee, Biplab Banerjee, and 1 more author

   2024

   arXiv
2. Preprint

   

   Deep evolving semi-supervised anomaly detection

   Jack Belham, Aryan Bhosale, Samrat Mukherjee, and 2 more authors

   2024

   arXiv