README.md

---
license: cc-by-sa-4.0
task_categories:
  - audio-classification
pretty_name: Engine Sound Windows (YouTube-derived, metadata-only)
tags:
  - audio
  - automotive
  - engine-sounds
  - youtube
  - weak-supervision
size_categories:
  - 1M<n<10M
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train.parquet
      - split: test
        path: data/test.parquet
dataset_info:
  features:
    - name: window_id
      dtype: string
    - name: youtube_id
      dtype: string
    - name: engine_class
      dtype: string
    - name: start_sec
      dtype: float64
    - name: end_sec
      dtype: float64
    - name: contains_target
      dtype: bool
    - name: quality_flag
      dtype: bool
    - name: split
      dtype: string
  splits:
    - name: train
      num_examples: 1511908
    - name: test
      num_examples: 372069
  download_size: 14615747
  dataset_size: 14615747
---

# Engine Sound Windows (YouTube-derived, metadata-only)

Timestamps and weak labels for training an engine-configuration audio classifier (v-twin vs.
inline-4 vs. flat-6, etc.) from short audio windows. **This dataset does not contain audio.**
Each row points at a public YouTube video id plus a `(start_sec, end_sec)` window; you fetch
and slice the audio yourself (see [Reconstructing audio](#reconstructing-audio) below).

## Why metadata-only

The source audio was collected by searching YouTube (via [`yt-dlp`](https://github.com/yt-dlp/yt-dlp))
for engine-sound terms and downloading matching videos. The dataset author does not hold
redistribution rights to that audio, so — following the precedent set by
[AudioSet](https://research.google.com/audioset/), [MusicCaps](https://huggingface.co/datasets/google/MusicCaps),
and [FSD50K](https://zenodo.org/records/4060432) for exactly this situation — only the video id,
window timestamps, and derived labels are published here, not audio bytes. This also means the
`cc-by-sa-4.0` license above covers only this repository's metadata (ids, timestamps, labels); the
underlying YouTube videos remain under their original creators' copyright and are not
relicensed or redistributed by this dataset in any form.

## Dataset structure

| Column | Type | Meaning |
| --- | --- | --- |
| `window_id` | string | Unique id for this window (`winNNN`), stable across the whole corpus |
| `youtube_id` | string | 11-character YouTube video id (`https://www.youtube.com/watch?v=<youtube_id>`) |
| `engine_class` | string | Engine configuration label, e.g. `v8_flat`, `i4_diesel`, `single_two_stroke` |
| `start_sec` / `end_sec` | float | Window bounds within the source video, in seconds |
| `contains_target` | bool | See [Label semantics](#label-semantics) |
| `quality_flag` | bool | See [Label semantics](#label-semantics) |
| `split` | string | `train` or `test` — assigned per source video, so every window from one video stays in the same split |

Windows are 2.0 seconds long with 1.0 second of step between consecutive windows (50% overlap),
confirmed directly from the underlying manifest's timestamps.

## Label semantics

`contains_target` and `quality_flag` are **model-derived, not human-verified** — they come from
running [`panns_inference`](https://github.com/qiuqiangkong/audioset_tagging_cnn)'s
`AudioTagging` model (trained on AudioSet) over each window and thresholding two sets of its 527
class scores:

- `contains_target = True` when the window's max score across a set of engine/vehicle AudioSet
  classes exceeds `0.585` — i.e. an engine sound was likely detected.
- `quality_flag = True` when the window's max score across a set of background-noise/contamination
  AudioSet classes exceeds `0.2` — i.e. contamination was likely detected.

**`quality_flag = True` is a caution flag, not an endorsement** — despite the name, it does not
mean the window is good quality. Treat both columns as weak, noisy supervision (useful for
filtering or as auxiliary features) rather than ground truth.

`engine_class`, by contrast, comes from which search query the source video was found under —
also not independently verified per-video (see [Known limitations](#known-limitations)).

## Engine classes

34 engine classes, 3,877 source videos, after exclusions below:

| Engine class | Files | Windows | Train | Test |
| --- | --- | --- | --- | --- |
| `2_rotor` | 141 | 43121 | 36452 | 6669 |
| `h12` | 10 | 4863 | 3026 | 1837 |
| `h2` | 92 | 56012 | 40333 | 15679 |
| `h4` | 112 | 40657 | 32253 | 8404 |
| `h6` | 97 | 84341 | 70911 | 13430 |
| `i2_180` | 127 | 88771 | 73716 | 15055 |
| `i2_180_two_stroke` | 41 | 9162 | 6487 | 2675 |
| `i2_270` | 169 | 134456 | 110712 | 23744 |
| `i2_360` | 63 | 24789 | 18944 | 5845 |
| `i2_360_two_stroke` | 44 | 12124 | 11617 | 507 |
| `i3` | 133 | 57496 | 45909 | 11587 |
| `i4` | 394 | 168553 | 140949 | 27604 |
| `i4_crossplane` | 134 | 68221 | 54634 | 13587 |
| `i4_diesel` | 76 | 31606 | 27721 | 3885 |
| `i5` | 115 | 29789 | 22419 | 7370 |
| `i5_diesel` | 52 | 21759 | 19513 | 2246 |
| `i6` | 157 | 49570 | 40540 | 9030 |
| `i6_diesel` | 52 | 29830 | 11117 | 18713 |
| `single_four_stroke` | 101 | 42045 | 29771 | 12274 |
| `single_two_stroke` | 94 | 44185 | 35973 | 8212 |
| `v10_72` | 156 | 63068 | 51501 | 11567 |
| `v12` | 112 | 37882 | 26548 | 11334 |
| `v2_45` | 138 | 165815 | 144387 | 21428 |
| `v2_90` | 97 | 83045 | 75219 | 7826 |
| `v4` | 80 | 31392 | 22601 | 8791 |
| `v4_two_stroke` | 41 | 9750 | 7735 | 2015 |
| `v6_120` | 72 | 31924 | 24099 | 7825 |
| `v6_60` | 211 | 60753 | 48324 | 12429 |
| `v6_90_even` | 76 | 43334 | 36442 | 6892 |
| `v8_cross` | 265 | 140562 | 111631 | 28931 |
| `v8_diesel` | 126 | 47716 | 30052 | 17664 |
| `v8_flat` | 180 | 81189 | 65364 | 15825 |
| `v8_voodoo` | 65 | 34937 | 24634 | 10303 |
| `vr6` | 54 | 11260 | 10374 | 886 |

Class sizes are heavily imbalanced (10 to 394 files per class) — account for this when
sampling/weighting during training.

## Known limitations

- **Weak, auto-derived labels.** `contains_target`/`quality_flag` come from an AudioSet-trained
  tagger's thresholded scores, not human review (see [Label semantics](#label-semantics)).
- **Class imbalance.** File counts per class range from 10 (`h12`) to 394 (`i4`).
- **All classes have representation in both train and test sets** (no class has zero test windows
  after conflict exclusions).
- **Cross-class conflicts.** Cross-referencing every video id against every engine_class it was
  scraped under found videos that had been pulled into **more than one conflicting engine_class**
  (almost certainly multi-engine compilation/comparison videos caught by more than one search
  query). All windows sourced from any of these videos were dropped entirely (99,105 of
  1,983,082 rows, 5.0%) rather than guessing which label was correct. This also resulted in
  9 engine classes being dropped entirely from the export (they had no videos that didn't also
  appear in at least one conflicting class). The class table above reflects the final, post-conflict
  list.
- **Link rot.** Since only YouTube ids are published (see [Why metadata-only](#why-metadata-only)),
  some fraction of source videos will become unavailable over time as creators delete or
  privatize them — unlike a self-hosted audio dataset, this one can shrink on its own.
- **`engine_class` isn't independently verified per video** beyond the cross-class-conflict
  check above — a video could still be mislabeled by its original search query in a way that
  doesn't produce a detectable cross-class conflict (e.g. a single video mislabeled but never
  scraped under any other class).

## Reconstructing audio

For a given row, download the source video's audio and trim to the window:

```bash
yt-dlp -f bestaudio -x --audio-format m4a \
  "https://www.youtube.com/watch?v=<youtube_id>" -o source.m4a

ffmpeg -i source.m4a -ss <start_sec> -to <end_sec> -c copy window.m4a
```

For batch reconstruction, group rows by `youtube_id` first so each video is downloaded once
regardless of how many windows come from it.

## License and usage

The labels, timestamps, and ids in this repository are released under `cc-by-sa-4.0`. This does
**not** extend any rights to the underlying YouTube video content, which remains the property of
its original creators — this dataset does not redistribute, host, or relicense that audio.
Commercial users should independently verify their own right to use any audio they fetch via the
ids in this dataset.