OpenTSLab/MachineHearing

Machine Hearing Benchmark (MHB)

This repository provides the Machine Hearing Benchmark (MHB) introduced in Machine Hearing: Can Audio LLMs Generalize Beyond Human Auditory Perception?

Multimodal LLMs have made substantial progress in perceiving the physical world, but their capabilities remain largely centered on signals that humans can directly see or hear. The physical world contains waves, vibrations, and oscillations far beyond human auditory perception. MHB investigates whether Audio Large Language Models (ALLMs) can generalize to this broader form of machine hearing.

MHB comprises 6,128 instances across 13 classification tasks and seven domains: Astronomy, Earth Science, Healthcare, Electromagnetic, Underwater Acoustics, Bioacoustics, and Manufacturing. It spans highly heterogeneous input regimes, with sampling rates from approximately 10⁻⁵ to 10⁶ Hz, sequence lengths from 10¹ to 10⁶ samples, and both single- and multi-channel inputs. Each instance consists of an input, a task-specific textual prompt, and a predefined label set.

Dataset Licenses and Distribution

MHB integrates data from multiple sources. The following table lists the source or processed-release licenses and whether the corresponding input data are included in this repository.

MHB Task	Data Source	Source / Release License	Input Files	Reference
GWOSC	GWOSC_GW_Event	CC BY-NC-SA 4.0 (SciTS)	Included	SciTS
LEAVES	LEAVES	CC BY 4.0	Included	NADC
Iquique	Iquique	Not specified	Reconstruction required	SeisBench
Llaima	Llaima	CC BY-NC 3.0	Included	Mendeley Data
Yaseen	Yaseen	Not specified	Reconstruction required	Official GitHub
CirCor_loc	CirCor	ODC-BY 1.0	Included	PhysioNet
RadSeg	RadSeg	MIT	Included	RadSeg
RadioML	RadioML2016.10A	CC BY-NC-SA 4.0	Included	DeepSig
ShipsEar	ShipsEar	Not specified	Reconstruction required	Official database
DeepShip	DeepShip	Not specified	Reconstruction required	Official GitHub
MarmAudio	MarmAudio	CC BY 4.0	Included	Zenodo
SDUST_bear	sdust_bearing via RMIS	MIT (RMIS processed release)	Included	RMIS
PU_cur	pu_cur via RMIS	MIT (RMIS processed release)	Included	RMIS

License notes. MHB's original benchmark contributions, to the extent the maintainers hold the relevant rights, are released under CC BY-NC-SA 4.0. Original data sources and processed releases retain their own applicable conditions and attribution requirements. No waveform files from Iquique, Yaseen, ShipsEar, or DeepShip are redistributed here; their metadata and reconstruction scripts are provided instead.

Dataset Structure

The repository contains a unified meta_data.jsonl file, task-specific data folders, and processing scripts. Each task folder follows the naming convention Domain-Task-Task_Name, with input data stored under raw_input/.

├── README.md
├── meta_data.jsonl
├── process/
│   ├── MHB.png
│   ├── map_deepship.py
│   ├── map_iquique.py
│   ├── map_shipsear.py
│   ├── map_yaseen.py
│   └── process_raw2wav.py
├── Astronomy-GWOSC-Gravitational_Wave_Event_Detection/
│   ├── raw_input/
│   └── wav_input/                  # generated on demand
├── Astronomy-LEAVES-Light_Curve_Classification/
│   └── raw_input/
├── ...
└── Manufacturing-PU_cur-Motor_Current_Fault_Diagnosis/
    └── raw_input/

# wav_data.jsonl is generated at the repository root.

• meta_data.jsonl: The complete set of 6,128 instances and annotations, including entries for the four datasets that must be obtained externally.
• <Dataset_Folder>/raw_input/: Input files in their original formats (csv, npy, wav, or flac). Directly hosted tasks contain their files; the four externally sourced tasks are reconstructed locally.
• process/map_*.py: Scripts for selecting and mapping externally obtained data to the exact names and relative paths in meta_data.jsonl.
• process/process_raw2wav.py: Optional conversion of six non-WAV tasks to wav_input/, also generating wav_data.jsonl.

The four externally hosted datasets can be reconstructed locally using the corresponding mapping scripts. The wav_input/ directories and wav_data.jsonl are generated on demand by process/process_raw2wav.py.

Tasks

All 13 tasks in MHB are formulated as classification tasks, with detailed task definitions provided in Appendix A.1 of the paper.

Domain	Task	Instances	Classification Objective	Dataset Folder
Astronomy	GWOSC	512	Gravitational-wave event detection	Astronomy-GWOSC-Gravitational_Wave_Event_Detection
Astronomy	LEAVES	1,024	Variable-star classification from light curves	Astronomy-LEAVES-Light_Curve_Classification
Earth Science	Iquique	502	Earthquake focal-depth classification	Earth_Science-Iquique-Earthquake_Focal_Depth_Classification
Earth Science	Llaima	358	Volcanic seismic-event classification	Earth_Science-Llaima-Volcanic_Seismic_Event_Classification
Healthcare	Yaseen	200	Heart-sound classification	Healthcare-Yaseen-Heart_Sound_Classification
Healthcare	CirCor_loc	697	Auscultation-location identification from PCG recordings	Healthcare-CirCor_loc-Auscultation_Location_Identification
Electromagnetic	RadSeg	482	Radar waveform/coding-scheme classification	Electromagnetic-RadSeg-Coding_Scheme-Classification
Electromagnetic	RadioML	550	Radio modulation classification	Electromagnetic-RadioML-Radio_Modulation_Classification
Underwater Acoustics	ShipsEar	84	Vessel and sound-source classification	Underwater_Acoustics-ShipsEar-Vessel_And_Sound_Source_Classification
Underwater Acoustics	DeepShip	527	Vessel-type classification	Underwater_Acoustics-DeepShip-Vessel_Type_Classification
Bioacoustics	MarmAudio	300	Marmoset vocalization-type classification	Bioacoustics-MarmAudio-Marmoset_Vocalisation-Classification
Manufacturing	SDUST_bear	572	Bearing fault diagnosis from vibration signals	Manufacturing-SDUST_bear-Vibration_Fault_Diagnosis
Manufacturing	PU_cur	320	Bearing fault diagnosis from motor-current signals	Manufacturing-PU_cur-Motor_Current_Fault_Diagnosis
Total	13 tasks	6,128

meta_data.jsonl Format

Each line is a JSON object representing one instance. The following is a shortened example; the text fields are illustrative:

{
  "task_name": "GWOSC",
  "dataset_name": "GWOSC_GW_Event",
  "id": "GW_3413075",
  "data_type": "csv",
  "input_ts": {
    "num_channel": 1,
    "channel_detail": [],
    "path": "Astronomy-GWOSC-Gravitational_Wave_Event_Detection/raw_input/0003747_gn_m1_14_87_m2_3_63_mtotal_18_q_4_1.csv",
    "length": 64000,
    "timestamps": [],
    "fs": 16000
  },
  "input_text": "Task-specific classification instruction",
  "gt_text": "The classification result is: normal",
  "gt_result": {
    "class_list": ["normal", "anomaly"],
    "gt_class": ["normal"]
  },
  "meta_data": {}
}

Field	Description
task_name	MHB task identifier (e.g., GWOSC, CirCor_loc, SDUST_bear).
dataset_name	Original or processed-source dataset identifier.
id	Unique identifier of the instance.
data_type	Input file format: csv, npy, wav, or flac.
input_ts.path	Path to the input file, relative to the MHB repository root. For externally sourced datasets, the corresponding mapping script restores this file locally.
input_ts.length	Input sequence length in time steps.
input_ts.num_channel	Number of input channels.
input_ts.channel_detail	Optional channel descriptions; empty when not provided.
input_ts.fs	Original sampling rate in Hz.
input_ts.timestamps	Optional timestamp information, usually an empty list.
input_text	Task-specific textual instruction given to the model.
gt_text	Reference answer in text form.
gt_result.class_list	List of valid classification labels.
gt_result.gt_class	Ground-truth class label as a list.
meta_data	Source-specific metadata (e.g., trace_name for Iquique or ori_file_path and seg_index for ShipsEar); fields vary by task.

Paths in meta_data.jsonl are relative to the repository root. For example, input data can be loaded using record["input_ts"]["path"] relative to the repository root.

For models that generate additional explanations or text, users may append an output-format constraint to the end of input_text during inference, such as:

"Output ONLY the classification result. Do not output any other text, explanation, or punctuation."

Reconstructing Externally Hosted Datasets

The input data for Iquique, Yaseen, ShipsEar, and DeepShip are not bundled in this release. Download the source data from the linked official providers, then run the relevant script from the MHB repository root. Each script uses meta_data.jsonl to select the exact MHB subset, creates the required task folder and raw_input/ directory, and preserves the relative paths already defined in the metadata. No changes to meta_data.jsonl are needed.

For the MHB mapping scripts, install the necessary Python packages:

pip install numpy h5py soundfile

Iquique (502 instances)

Source: SeisBench — Iquique.

Obtain the SeisBench Iquique dataset and locate the folder containing waveforms.hdf5 and metadata.csv. Run:

python process/map_iquique.py /path/to/Iquique

The script uses meta_data.trace_name to extract the specified three-component seismic traces from the HDF5 file and saves 502 NPY files with the names and paths defined in meta_data.jsonl.

Yaseen (200 instances)

Source: Official Yaseen GitHub repository.

Download or clone the repository and extract the five category archives (AS, MR, MS, MVP, and N). The extracted directories are named AS_New_3주기/, MR_New_3주기/, MS_New_3주기/, MVP_New_3주기/, and N_New_3주기/. For example, from the downloaded source directory:

for archive in *_New_3주기.rar; do unrar x "$archive"; done

Use unrar or a recent compatible RAR extractor. Then, from the MHB root:

python process/map_yaseen.py /path/to/Classification-of-Heart-Sound-Signal-Using-Multiple-Features-

The script selects and copies the 200 specified WAV files from the five categories.

ShipsEar (84 instances)

Original data: ShipsEar official website (access to the complete dataset may require contacting its authors).

Train/test split: ShipsEar — An Unofficial Train-Test Split, using its 12-class split.

Download ShipsEar and apply the linked repository's 12-class train-test-split.py to produce a folder such as:

shipsear_with_split-12class/
├── train/
└── test/

Then run:

python process/map_shipsear.py /path/to/shipsear_with_split-12class/test

Following the segmentation protocol associated with the referenced 12-class split, the mapping script extracts 30-second windows at 15-second intervals, including the designated Trawler test excerpt, and writes the 84 MHB WAV files to their specified raw_input/ paths.

DeepShip (527 instances)

Original data: Official DeepShip GitHub.

Train/test split: Method for Splitting the DeepShip Dataset.

Download DeepShip and apply the linked repository's train.txt, test.txt, and code.ipynb to obtain the segmented test data. The split code generates a directory named DeepShip_test, with the following structure:

DeepShip_test/
├── Cargo/
├── Passengership/
├── Tanker/
│   └── 20170822-133/
│       └── 172736.wav/
│           └── 00027.wav
└── Tug/

Pass the path to the DeepShip_test directory (the one containing the four category folders):

python process/map_deepship.py /path/to/DeepShip_test

The script selects the 527 designated three-second test segments, renames each output WAV using the sample id, and copies the files to the corresponding raw_input/ location.

RAW-to-WAV Conversion

For Audio LLM APIs that only accept WAV files, MHB provides process/process_raw2wav.py. First reconstruct Iquique, because its NPY inputs are not included in the repository. From the MHB root, run:

pip install numpy pandas
python process/process_raw2wav.py

The script converts GWOSC, LEAVES, Iquique, Llaima, RadSeg, and RadioML (3,428 instances) into each task's wav_input/ directory.

The script also generates wav_data.jsonl in the MHB root, containing all 6,128 instances. For the six converted tasks, it updates data_type and input_ts.path, length, num_channel, and fs; it additionally sets input_ts.channel_detail to [] for Iquique, RadSeg, and RadioML. The other seven tasks remain unchanged and retain their original WAV/FLAC input references where applicable.

Citation

If you use MHB, please cite:

@misc{liu2026machinehearing,
  title         = {Machine Hearing: Can Audio LLMs Generalize Beyond Human Auditory Perception?},
  author        = {Liwei Liu and Jun Tao and Xingyuan Li and Xuenan Xu and Chen Zhang and Junyi Wang and Boya Zhang and Xiaoyu Yang and Zhen-Hua Ling and Wen Wu and Chao Zhang},
  year          = {2026},
  eprint        = {2611.10693},
  archivePrefix = {arXiv},
  primaryClass  = {cs.SD},
  url           = {https://arxiv.org/abs/2611.10693}
}

Please also cite the original dataset publications and the relevant processing or split releases, including SciTS, RMIS/FISHER, and the referenced ShipsEar / DeepShip split methods, as applicable.