IEEE SLT 2026

CARD audio captioning

Cross-component Audio Representation Distillation for Encoder-Free Audio Captioning. Listen to a clip, reveal the model's caption, and compare it with what human annotators wrote.

How this page works: the captions below were generated beforehand by the released model CARD-Qwen3-4B-AudioCaps with the paper's decoding settings, and are identical to its live output for these clips. To caption your own audio, run the model live in Colab.
1. Choose a clip
No clip selected 0:00 / 0:00
2. Compare with the human captions
How CARD works
Training only: a frozen CLAP-HTSAT teacher is distilled by component, its early stages into the projector and its later stages into the language model. The teacher is removed, so there is no audio encoder at inference.
55.4
CIDEr-D on AudioCaps
13.2 M
audio interface parameters, 34.4 M with the encoder
0.4 ms
audio interface latency, 11.1 ms with the encoder
0
audio encoders at inference

Numbers from the paper (Table II and Section IV-B).

Cite this work 
@inproceedings{kolluri2026card,
  title     = {{CARD}: Cross-component Audio Representation Distillation for Encoder-Free Audio Captioning},
  author    = {Kolluri, Ganesh Pavan Kartikeya Bharadwaj and Zhang, Yuchen and Kampouridis, Michael and Shekhar, Ravi},
  booktitle = {Proceedings of the IEEE Spoken Language Technology Workshop (SLT)},
  year      = {2026}
}