Clotho-AQA: A Crowdsourced Dataset for Audio Question Answering (arXiv 2204.09634)

Forum topic · 小凯 · 2026-07-05

Summary

This forum post indexes the arXiv paper Clotho-AQA (arXiv:2204.09634, April 2022) by Samuel Lipping, Parthasaarathy Sudarsanam, Konstantinos Drossos, and Tuomas Virtanen. Clotho-AQA is a crowdsourced dataset designed for audio question answering, extending the Clotho audio captioning corpus with question-answer pairs about environmental sounds. The paper targets the intersection of audio understanding, multi-modal learning, and question answering, providing a benchmark resource for training and evaluating models that answer natural-language questions about audio content. The surrounding commentary situates the work within broader trends in information retrieval and retrieval-augmented generation, noting how audio QA complements text and vision QA datasets. Readers should consult the original PDF at https://arxiv.org/abs/2204.09634 for exact dataset statistics, collection methodology, and baseline results, as this post primarily provides metadata, categorization, and cross-references to related multi-modal retrieval and audio QA resources rather than reproducing the paper's quantitative findings.

Clotho-AQA: A Crowdsourced Dataset for Audio Question Answering

Overview

This post catalogs the paper Clotho-AQA: A Crowdsourced Dataset for Audio Question Answering (arXiv, April 2022).

| Field | Detail | |-------|--------| | Authors | Samuel Lipping, Parthasaarathy Sudarsanam, Konstantinos Drossos, Tuomas Virtanen | | Link | https://arxiv.org/abs/2204.09634 | | Type | Academic paper / dataset resource | | Categories | Multi-modal, Question Answering |

Key points

Clotho-AQA is a crowdsourced dataset for audio question answering (AQA), built as an extension of the Clotho audio captioning dataset.
It enables models to answer natural-language questions about environmental audio recordings, combining audio understanding with language comprehension.
The work addresses a gap in multi-modal QA: while text and image QA are well studied, question answering over non-speech audio had comparatively fewer dedicated resources.
The authors are affiliated with audio signal processing research (Tampere University group), continuing the Clotho line of audio captioning resources.

Context in the field

The post situates Clotho-AQA within broader discussions of information retrieval and multi-modal learning, alongside related entries such as:

Listen, Think, and Understand / OpenAQA dataset (May 2023) — a later audio QA dataset.
Cross-modal retrieval surveys and vision-language retrieval works (ColPali, EA-VTR), which share the goal of aligning language queries with non-textual content.

Notes and limitations of this post

This entry is largely a metadata summary and cross-reference index; the detailed methodology, dataset statistics, annotation protocol, and baseline model results are not reproduced here. Readers should verify quantitative claims against the original PDF at arXiv:2204.09634.
The generic commentary sections in the source post (on RAG, ranking, and engineering checklists) are template content and do not reflect specifics of the Clotho-AQA paper itself.

Citation

Original paper: Clotho-AQA: A Crowdsourced Dataset for Audio Question Answering. arXiv:2204.09634, April 2022.