ART (Audio Reasoning Tasks)

Dataset Description

ART (Audio Reasoning Tasks) is a benchmark for evaluating the audio reasoning capabilities of multimodal large language models (MLLMs). Unlike benchmarks that evaluate individual audio capabilities in isolation, ART focuses on tasks that require models to combine multiple skills for understanding and reasoning over audio signals.

The benchmark consists of 9 000 audio samples spanning nine audio reasoning tasks, with 1 000 samples per task and over 30 hours of audio in total. Each sample contains a spoken question together with the audio information required to answer it. All questions are formulated so that the expected answer can be reduced to Yes or No.

ART is fully balanced with respect to both tasks and target labels - each task contains 500 samples with the expected answer Yes and 500 samples with the expected answer No.

The benchmark was designed around two main principles:

1. A task should require reasoning over multiple properties of the audio signal rather than being solvable using the output of a single specialized audio-processing module.
2. A task should be approachable by a person without professional training in audio, speech, or music.

For more details, see the ART paper.

Example Usage

The dataset can be loaded using the datasets library:

from datasets import load_dataset

dataset = load_dataset("amu-cai/ART", "all", split="test")
task = load_dataset("amu-cai/ART", "audio_arithmetics", split="test")

Supported Tasks

ART contains nine tasks covering different forms of reasoning over speech, environmental sounds, and other acoustic information.

Task	Abbreviation	Description
Audio Arithmetics	AA	Performing simple arithmetic reasoning based on sounds heard in the recording.
Audio Transformation Detection	ATD	Determining whether one recording is a transformed version of another.
Cross-Recording Language Identification	CRLI	Comparing languages spoken across recordings and reasoning about the identified languages.
Cross-Recording Speaker Identification	CRSI	Comparing speaker identities across recordings.
Selective Text Inference	STI	Reasoning about spoken content selected according to properties of the speakers.
Sound Reasoning	SR	Answering questions that require recognizing sounds and reasoning about the objects or events producing them.
Speech Features Comparison	SFC	Comparing recordings with respect to speech characteristics such as accent.
Text and Sound Reasoning	TSR	Combining spoken-language understanding with recognition of non-speech sounds.
Text and Temporal Localization Reasoning	TTLR	Combining spoken-language understanding with temporal or acoustic-scene information.

For example, tasks may require determining whether two speakers are the same person, comparing the number of sound events, identifying relationships between languages, or combining the semantic content of speech with environmental sounds.

Languages

Most synthesized speech in ART is in English.

The Cross-Recording Language Identification (CRLI) task additionally contains speech in four languages - Estonian, Finnish, Hungarian, and Polish.

Dataset Structure

Data Instances

An example instance from the Audio Arithmetics task:

{
  'instance_id': 'aa_4160014',
  'task': 'audio_arithmetics',
  'audio': {
    'path': None,
    'array': array([ 0.00054932,  0.00054932,  0.00045776, ...,
      -0.00027466, -0.00024414, -0.00045776]),
    'sampling_rate': 16000
  },
  'answer': 'yes',
  'question': 'Are there more bell rings than cat meows?',
  'template_id': 'aa_4',
  'question_id': 'aa_416',
  'slots': '{"n_sound_A": 4, "n_sound_B": 3, "sound_A": "bell", "sound_B": "cat"}'
}

Data Fields

The benchmark provides the information necessary to associate each audio recording with:

• instance_id (string): Unique identifier of the instance.

• task (string): Name of the audio reasoning task the instance belongs to.

• audio (audio): Audio recording containing the spoken question and the audio information required to answer it.

  • path (string or None): Path associated with the audio file. May be None when the audio is decoded directly from the dataset.

  • array (array): Decoded audio waveform.

  • sampling_rate (int): Sampling rate of the audio waveform, equal to 16 kHz.

• answer (string): Ground-truth answer to the question, eiter yes or no.

• question (string): Textual representation of the question spoken in the audio.

• template_id (string): Identifier of the template used to generate the question.

• question_id (string): Identifier of the generated question.

• slots (string): JSON-encoded dictionary containing the values used to fill the question template. Its contents depend on the task and template.

Data Distribution

Task	Samples	Templates	Speakers	Utterances	Sounds	Total duration
AA	1 000	6	-	-	5	3h 46m 10s
ATD	1 000	4	-	-	4	3h 53m 30s
CRLI	1 000	6	12	12	-	3h 32m 01s
CRSI	1 000	4	4	8	-	3h 46m 17s
STI	1 000	4	4	36	-	3h 09m 47s
SR	1 000	15	-	-	20	3h 08m 15s
SFC	1 000	4	10	3	-	2h 37m 07s
TSR	1 000	8	4	16	17	3h 25m 24s
TTLR	1 000	4	4	15	8	3h 00m 47s
Total	9 000	55	22	86	25	30h 19m 18s

Each task contains exactly 1 000 instances, consisting of 500 Yes answers and 500 No answers.

The complete benchmark therefore contains 4 500 samples for each answer class.

Dataset Creation

Task Selection

ART was created using a multi-stage procedure.

Candidate tasks were first proposed by speech and natural language processing experts. Tasks were required to combine multiple audio-processing capabilities and to remain approachable without professional training.

Tasks involving subjective judgments, such as emotion classification or subjective audio-quality assessment, were excluded. Tasks that could not be expressed as Yes/No questions were also excluded to allow evaluation without relying on another language model as a judge.

This process resulted in the nine tasks included in ART.

Task Templates

Instances were generated using task-specific templates. Each template contains slots populated with values such as sentences, speakers, sounds, or other task-specific attributes.

The use of templates enables systematic generation of examples while maintaining control over the distribution of tasks and answers.

In total, ART uses 55 templates across its nine tasks.

Speech

The spoken questions were synthesized using a common voice to provide consistent prompts across the benchmark.

Additional utterances required by individual tasks were generated using speech samples selected from existing speech datasets. English speaker material was derived from GLOBE, while the multilingual speech used for Cross-Recording Language Identification was selected from VoxPopuli.

The multilingual portion contains speakers of Estonian, Finnish, Hungarian, and Polish.

Speech synthesis was performed using a pipeline based on Voicebox.

Sounds

Non-speech sounds were manually selected from recordings available under the Creative Commons CC0 license in Freesound.

The selected material included 13 short sounds, four short tunes, and eight background sounds.

Short sound recordings were manually trimmed where necessary to contain a single sound event.

Audio Processing

After generating the necessary questions, utterances, and sounds, recordings were automatically composed according to their corresponding task templates.

All recordings were normalized to −20 dBFS. Silence between components was manually adjusted to make the prompts natural and ensure that individual recordings could be distinguished.

Short sounds and tunes concatenated with questions were limited to a maximum duration of five seconds and faded out where necessary. For tasks involving background audio, the background signal was attenuated by 20 dB to preserve speech intelligibility.

Evaluation

ART supports a straightforward Yes/No evaluation protocol:

Answer the question from the audio.
Answer only "Yes" or "No".

Because every instance has a deterministic binary target, models can be evaluated without using an LLM as a judge.

The paper additionally investigates a descriptive evaluation setting in which models are allowed to provide unrestricted answers:

Answer the question from the audio.

Descriptive answers can subsequently be evaluated manually or using an LLM-based judge.

Limitations

ART is intended to measure whether multimodal models can perform reasoning over audio signals by combining multiple audio-related capabilities. However, the nine tasks do not represent every possible form of audio reasoning.

Consequently, strong performance on ART should not by itself be interpreted as evidence that a model has human-level audio reasoning capabilities.

The benchmark also relies substantially on synthesized speech and deliberately uses relatively clean audio. This design reduces ambiguity and helps isolate reasoning failures from failures caused by poor acoustic quality, but it does not represent the full range of challenging conditions encountered in real-world audio.

Citation Information

The ART paper is available in ACL Anthology.

If you use ART in your research, please cite:

@inproceedings{christop-etal-2026-benchmark,
    title = "A Benchmark for Audio Reasoning Capabilities of Multimodal Large Language Models",
    author = "Christop, Iwona  and
      Czy{\.z}nikiewicz, Mateusz  and
      Sk{\'o}rzewski, Pawe{\l}  and
      Bondaruk, {\L}ukasz  and
      Kubiak, Jakub  and
      Lewandowski, Marcin  and
      Kubis, Marek",
    editor = "Demberg, Vera  and
      Inui, Kentaro  and
      Marquez, Llu{\'i}s",
    booktitle = "Proceedings of the 19th Conference of the {E}uropean Chapter of the {A}ssociation for {C}omputational {L}inguistics (Volume 1: Long Papers)",
    month = mar,
    year = "2026",
    address = "Rabat, Morocco",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.eacl-long.42/",
    doi = "10.18653/v1/2026.eacl-long.42",
    pages = "953--983",
    ISBN = "979-8-89176-380-7",
    abstract = "The present benchmarks for testing the audio modality of multimodal large language models concentrate on testing various audio tasks such as speaker diarization or gender identification in isolation. Whether a multimodal model can answer the questions that require reasoning skills to combine audio tasks of different categories cannot be verified with their use. To address this issue, we propose Audio Reasoning Tasks (ART), a new benchmark for assessing the ability of multimodal models to solve problems that require reasoning over audio signal."
}