{Christophe Cerisara Romain Serizel F ́elix Gontier}

Abstract

utomated audio captioning is the multimodal task of describingenvironmental audio recordings with fluent natural language. Mostcurrent methods utilize pre-trained analysis models to extract rele-vant semantic content from the audio input. However, prior infor-mation on language modeling is rarely introduced, and correspond-ing architectures are limited in capacity due to data scarcity. Inthis paper, we present a method leveraging the linguistic informa-tion contained in BART, a large-scale conditional language modelwith general purpose pre-training. The caption generation is condi-tioned on sequences of textual AudioSet tags. This input is enrichedwith temporally aligned audio embeddings that allows the model toimprove the sound event recognition. The full BART architectureis fine-tuned with few additional parameters. Experimental resultsdemonstrate that, beyond the scaling properties of the architecture,language-only pre-training improves the text quality in the multi-modal setting of audio captioning. The best model achieves state-of-the-art performance on AudioCaps with 46.5 SPIDEr.