Temporal Audio Narratives: Event Reasoning Through Large Language Models
IEEE Access · 2026 · 14 · 74566-74575
DOI: 10.1109/access.2026.3693145 ↗Kurz gefasst
Fachartikel in der Zeitschrift IEEE Access, 2026. Verfasst von Nikolai A. Kravchenko und 2 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0).
Abstract (Originalsprache)
Large audio-language models enable scene understanding from raw audio but require substantial computational resources, limiting edge deployment. This paper proposed and evaluated a hybrid two-stage pipeline that combines sound event detection with compact chain-of-thought reasoning for efficient audio scene captioning. Using a curated subset of 204 CompA-R-Instructions samples (70 train, 134 test) with ground-truth event timestamps, a LoRA-adapted Gemma-3N model achieved superior LLM-as-judge scores (0.673) versus state-of-the-art audio LLMs (0.274–0.397), despite operating only on symbolic event inputs. Therefore, the reported performance should be interpreted as an upper-bound estimate under idealized event annotations, since the current study uses oracle event timestamps rather than outputs from a real sound event detection system. The event-driven approach demonstrated greater robustness (0.11 vs. 0.18–0.21), concise outputs, and predictable hardware-bound latency, highlighting structured event conditioning and reasoning distillation as an effective architectural bias for resource-constrained audio understanding.
Zitieren
APA
Kravchenko, N. A., Gorodnichev, M., & Mkrtchian, G. (2026). Temporal Audio Narratives: Event Reasoning Through Large Language Models. IEEE Access, 14, 74566-74575. https://doi.org/10.1109/access.2026.3693145
Vancouver
Kravchenko NA, Gorodnichev M, Mkrtchian G. Temporal Audio Narratives: Event Reasoning Through Large Language Models. IEEE Access. 2026;14:74566-74575. doi:10.1109/access.2026.3693145
BibTeX
@article{Kravchenko2026,
author = {Kravchenko, Nikolai A. and Gorodnichev, Mikhail and Mkrtchian, Grach},
title = {Temporal Audio Narratives: Event Reasoning Through Large Language Models},
journal = {IEEE Access},
year = {2026},
volume = {14},
pages = {74566--74575},
doi = {10.1109/access.2026.3693145}
}
EUVIMED