rayyyyy007/kwerimoto-engine-sounds

KweriMoto Engine Sounds

Car engine sounds with three labels: healthy, engine_knock and misfire. The dataset combines three public sources into one set of clips with one fixed split, so that different models can be compared fairly. It was made for KweriMoto, powered by EdgeKnock, a final-year project (CSC 3118, Makerere University) that detects engine faults from phone recordings.

All clips are mono, 16 kHz, 16-bit WAV, at most 10 seconds long. Shorter clips are not padded.

The code that builds the dataset: github.com/SUUNARAYMOND/kwerimoto-dataset

Splits

label	source	train	validation	test	real_test	total
engine_knock	AudioSet (real)	199	43	42	38	322
engine_knock	Revix (synthetic)	200	43	43	0	286
healthy	Hugging Face	56	12	12	0	80
healthy	Revix	190	40	41	0	271
misfire	Hugging Face	76	16	16	0	108
total		721	154	154	38	1,067

• train / validation / test (70 / 15 / 15): every label from every source is split in these shares. Clips that share a recording (the same YouTube video, car recording or knock sound) are always in the same split, so a model is never tested on a recording it heard during training.
• real_test: 38 real engine-knock recordings from the clips Google set aside for evaluating AudioSet. None of them was used to make the synthetic clips, and none shares a video with any other clip here. It answers one question: does a model trained here recognise real engine knocking? It contains only knock, so it measures how many real knocks a model catches, not false alarms. Do not train or tune on it.

How to load

With plain files (works anywhere, including Google Colab):

from huggingface_hub import snapshot_download
import pandas as pd
import soundfile as sf

root = snapshot_download("rayyyyy007/kwerimoto-engine-sounds", repo_type="dataset")
manifest = pd.read_csv(f"{root}/manifest.csv", keep_default_na=False)
train = manifest[manifest["split"] == "train"]
audio, sample_rate = sf.read(f"{root}/{train['file_name'].iloc[0]}")  # sample_rate == 16000

With the datasets library:

from datasets import load_dataset

ds = load_dataset("rayyyyy007/kwerimoto-engine-sounds")  # train, validation, test, real_test

With datasets 4 or newer, reading the audio also needs pip install torchcodec.

Files

manifest.csv                        one row per clip (all splits)
prepared/<split>/<label>/*.wav      the clips
prepared/<split>/metadata.jsonl     that split's rows of the manifest

Column	Meaning
file_name	path of the clip (in manifest.csv: from the dataset root; in metadata.jsonl: from the split folder)
label	healthy, engine_knock or misfire
split	train, validation, test or real_test
group	clips that share a recording have the same group, e.g. youtube:DqOGNij7qIE
source_dataset	revix, huggingface or audioset
original_file	file name in the source dataset
source_recording	Revix: the car recording used; others: the file itself
knock_source	Revix knock clips: the knocking sound laid on top; otherwise empty
synthetic	True for Revix knock clips, False for everything else
vehicle_tag	AudioSet clips: car or untagged (no vehicle named in Google's labels); otherwise empty
duration_s	length of the clip in seconds
original_sample_rate, original_channels, original_duration_s	format of the source file
sha256	fingerprint of the source file

Sources and credits

Source	Clips used	Label here
Revix engine-sounds dataset by Lwanga Caleb, Arinda Emmanuel and Ssempija Gideon Ethan: car clips	271 + 286	healthy, engine_knock
RashmiKarunarathne/MyDataSet on Hugging Face (revision 07bb22b): healthy_*, misfire_*	80 + 108	healthy, misfire
Google AudioSet strong labels: YouTube clips labelled "Engine knocking"	322	engine_knock

• Revix knock clips are synthetic. Revix took a real "Engine knocking" sound from AudioSet, filtered it, and laid it over a car recording. Revix's healthy clips are AudioSet clips tagged "Car".
• AudioSet clips were downloaded from YouTube. Clips tagged with another vehicle or machine (motorcycle, lawn mower, truck, ...) were left out; car-tagged and untagged clips were kept. Clips from Google's evaluation list form real_test.

How it was built

1. Gather every clip from the three sources.
2. Remove files that are missing, unreadable, completely silent, exact duplicates, or exact duplicates with conflicting labels (49 removed out of 1,116).
3. Group clips that share a recording.
4. Split whole groups 70 / 15 / 15 (fixed seed 42); Google's evaluation clips go to real_test.
5. Convert to mono 16 kHz, and cut anything longer than 10 s to one random 10 s piece.

Known issues

• Labels are not verified. They come from the source datasets. A Revix "Car" clip means a car can be heard, not that a mechanic checked the engine.
• Background voices and music are kept on purpose. Many YouTube clips have people talking, wind or music in the background (about 4 in 10 Revix clips, both healthy and knock). The final system will be used where people talk nearby. The Hugging Face clips are much cleaner, so misfire is never heard with background voices.
• Real knock clips may not all be cars. 264 of the 322 real clips name no vehicle at all; filter on vehicle_tag == "car" for the 58 that are certainly cars. Some Revix synthetic clips use a knock sound taken from a motorcycle or lawn-mower video.
• Synthetic knock clips are shorter. No Revix knock clip is a full 10 s, and 21 knock clips are under 1 s. If every clip is padded to 10 s, a model can learn "more silence means knock". Cut fixed windows from real audio instead of padding.
• Sources differ in how they sound. Revix and AudioSet clips are YouTube videos; the Hugging Face clips are cleaner recordings. Healthy comes from both kinds, misfire only from Hugging Face.

Licence

This dataset is shared for research and education. The audio belongs to its original owners: Revix and AudioSet clips come from YouTube uploads, and RashmiKarunarathne/MyDataSet states no licence. AudioSet's labels are © Google, CC BY 4.0. Check the original sources before any other use.

Citation

If you use the AudioSet clips, please cite:

• Gemmeke et al., "Audio Set: An ontology and human-labeled dataset for audio events", ICASSP 2017.
• Hershey et al., "The benefit of temporally-strong labels in audio event classification", ICASSP 2021.