---
pretty_name: Procedural Engine Sounds Dataset
task_categories:
- audio-to-audio
- audio-classification
tags:
- audio
- audio-dataset
- engine-sounds
- combustion-engine
- procedural-generation
- time-aligned
- rpm
- torque
- automotive
- nvh
- vehicle-acoustics
- sound-synthesis
- noise-free
- synthetic-audio
- annotations
- simulation
- multi-channel
size_categories:
- 10B<n<100B
license: cc-by-nc-4.0
license_url: https://creativecommons.org/licenses/by-nc/4.0/
---


# Procedural Engine Sounds Dataset (Official)

## ⚠️ Canonical Source

This is the **original and official release** of the Procedural Engine Sounds Dataset, created by Robin Doerfler:  
- Project page (start here): https://rdoerfler.github.io/procedural-engine-sounds-page/  
- Hugging Face repository: https://huggingface.co/datasets/rdoerfler/procedural-engine-sounds  
- Zenodo DOI (primary citation): https://doi.org/10.5281/zenodo.16883336

Other versions of this dataset on Hugging Face are duplicates of this repository.  
This repository should be considered the authoritative and maintained source.

## Official Links

- Project page (overview, links, citations): https://rdoerfler.github.io/procedural-engine-sounds-page/
- Zenodo (DOI, primary citation): https://doi.org/10.5281/zenodo.16883336
- Paper (EUSIPCO 2026): https://eurasip.org/Proceedings/Eusipco/Eusipco2026/pdfs/0000221.pdf
- Preprint (arXiv): https://arxiv.org/abs/2603.07584
- Analysis code: https://github.com/rdoerfler/engine-order-analysis

**Keywords:** engine sound dataset, car engine audio, combustion engine, audio synthesis, synthetic audio, NVH, vehicle acoustics, RPM, torque, annotations, automotive audio, audio generation, simulation, machine learning dataset

## Dataset Description

The Procedural Engine Sounds Dataset is a comprehensive collection of synthetically generated and annotated engine audio samples. This dataset contains procedurally generated high-resolution engine sounds free of confounding noises, with detailed time-aligned annotations, designed for research in audio processing, vehicle acoustics, and synthetic sound generation.

## Associated Publication

**Analysis-Driven Procedural Generation of an Engine Sound Dataset with Embedded Control Annotations**  
Robin Doerfler, Lonce Wyse  
In *Proceedings of the 34th European Signal Processing Conference (EUSIPCO 2026)*, Bruges, Belgium, pp. 221-225.  
Published version: https://eurasip.org/Proceedings/Eusipco/Eusipco2026/pdfs/0000221.pdf  
Preprint: https://arxiv.org/abs/2603.07584

## Dataset Details

### Dataset Summary

- **Repository:** procedural-engine-sounds
- **Version:** 1.0
- **Publication Year:** 2025
- **License:** CC BY-NC 4.0

### Research Applications

- **Audio Generation**: Train models to generate realistic engine sounds conditioned to time-varying engine operation states
- **Audio Classification**: Predict RPM and Engine Torque based on audio signals
- **Audio Analysis**: Research vehicle acoustics and engine sound patterns
- **Sound Synthesis**: Develop procedural audio generation techniques
- **Data Augmentation**: Use as augmentation material for in-cabin speech detection and recognition, noise suppression or other related tasks

### Technical Specifications

This dataset contains only audio signals - no textual or linguistic content. Both the engine sounds and annotations (RPM/torque information) are provided as audio signals at 48 kHz sample rate in WAV format.

## Dataset Structure

### Data Organization

The dataset is organized into 8 distinct sets with two categories:

**Extended Sets** (A, B, C, D):
- 3,068 files total (across 4 sets)
- ~9.83 hours of audio total
- ~12.65 GB total
- **Per set: 767 files, ~2.46 hours, ~3.16 GB**

**Standard Sets** (E, F, G, H):
- 2,867 files total (across 4 sets)
- ~9.18 hours of audio total
- ~11.82 GB total
- **Per set: ~717 files, ~2.30 hours, ~2.96 GB**

Note: Minor size differences between Standard and Extended sets reflect practical data availability during generation and have no methodological significance.

**Dataset Total: 5,935 files, ~19.01 hours, ~24.47 GB**

### File Organization

```
README.md (this file)
USAGE.txt (quick start guide)
audio/
├── A_full_set/ (767 files, ~2.46 hours, ~3.16 GB)
├── B_full_set/ (767 files, ~2.46 hours, ~3.16 GB)
├── C_full_set/ (767 files, ~2.46 hours, ~3.16 GB)
├── D_full_set/ (767 files, ~2.46 hours, ~3.16 GB)
├── E_large_set/ (717 files, ~2.30 hours, ~2.96 GB)
├── F_large_set/ (717 files, ~2.30 hours, ~2.96 GB)
├── G_large_set/ (717 files, ~2.30 hours, ~2.96 GB)
└── H_large_set/ (717 files, ~2.30 hours, ~2.96 GB)
metadata/
├── A_full_set_summary.json
├── A_full_set_stats.csv
├── B_full_set_summary.json
├── B_full_set_stats.csv
└── … (16 metadata files total)
```

### File Formats

**Audio Files:**
- Format: WAV
- Sample Rate: 48 kHz
- Channels: 4 (quad-channel)
- Bit Depth: 16 bit

**Metadata Files:**
- Summary files: JSON format
- Statistics files: CSV format (comma-separated values)

### Data Structure

Each audio file contains **4-channel audio** at 48 kHz sample rate:
- **Channel 1-2**: Stereo engine sound audio
- **Channel 3**: Engine speed (RPM × 0.0001) as continuous audio signal
- **Channel 4**: Engine torque (Nm × 0.001) as continuous audio signal

### Data Splits

No predefined train/validation/test splits are provided. 
Depending on the task, different splits are appropriate, 
and the roles of inputs and targets may vary (e.g., audio-to-annotation or annotation-to-audio). 
Users are encouraged to define splits according to their specific use case.

### Metadata Structure

#### Summary Files (.json)
Per-set statistics including:
- `num_files`: Number of audio files in set
- `total_duration_*`: Duration in seconds/minutes/hours
- `total_size_gb`: Storage size in GB
- `rpm_distribution`: Statistical distribution (min, max, mean, std, percentiles)
- `torque_distribution`: Statistical distribution (min, max, mean, std, percentiles)

#### Statistics Files (.csv)
Per-file metrics with columns:
- `filename`: Audio file name
- `samplerate`: Sample rate (48 kHz)
- `duration_sec`: File duration in seconds
- `size_MB`: File size in megabytes
- `rpm_min/max/mean/std`: RPM statistics for the file
- `torque_min/max/mean/std`: Torque statistics for the file

### Data Access

Each audio file is a standard WAV file containing a 4-channel audio array at 48 kHz sample rate. When loaded, you receive the raw multichannel audio data from which RPM and torque information can be extracted from channels 3 and 4 respectively.

### Audio Signal Encoding

- **RPM Signal**: Channel 3 contains engine speed values in RPM scaled by 0.0001 (multiply by 10,000 to get actual RPM)
- **Torque Signal**: Channel 4 contains torque values in Newton meters scaled by 0.001 (multiply by 1,000 to get actual Nm)
- **Engine Audio**: Channels 1-2 contain the stereo procedural engine sound

## Technical Requirements

To work with this dataset, you will need:
- Audio processing software capable of reading multi-channel WAV files
- Programming languages: Python (recommended with librosa, soundfile, or scipy), MATLAB, R, or similar
- For metadata: JSON and CSV reading capabilities

## Dataset Creation

### Source Data

All audio samples are synthetically generated using procedural audio synthesis techniques as described in the associated paper.
Dataset results were thoroughly analyzed and compared to real world recordings to verify representativeness and similarity regarding engine order magnitudes and harmonic deviations.

### Annotations

Annotations were created during the generation process, with additional manual verification for quality assurance.

## Considerations for Using the Data

### Social Impact of Dataset

This dataset enables research in:
- Automotive audio simulation
- Vehicle sound design
- Audio processing algorithms
- Synthetic data generation techniques

### Discussion of Biases

As a synthetic dataset, it reflects the biases inherent in the procedural generation algorithms and may not capture all real-world engine sound variations.

### Other Known Limitations

- Limited to procedurally generated sounds
- May not represent all engine types or acoustic environments
- Generated with empirically determined synthesis parameters, hence contains fictional engine types and exhaust pipe configurations

## License and Usage

### License

This dataset is released under CC BY-NC 4.0 license (Creative Commons Attribution-NonCommercial 4.0 International).

**Original Author:** Robin Doerfler  
This dataset was first published via Zenodo and Hugging Face.

**Attribution Required**: When using this dataset, please cite the Zenodo release as the primary source, and the EUSIPCO 2026 paper when referring to the generation method:

DOI: https://doi.org/10.5281/zenodo.16883336

For convenience, this repository provides direct access to the dataset and documentation.

### Citation

```bibtex
@dataset{doerfler_2025_procedural_engine_sounds,
  author       = {Doerfler, Robin},
  title        = {Procedural Engine Sounds Dataset},
  month        = {August},
  year         = 2025,
  publisher    = {Zenodo},
  version      = {1.0},
  doi          = {10.5281/zenodo.16883336},
  url          = {https://doi.org/10.5281/zenodo.16883336}
}

@inproceedings{doerfler2026analysis,
  author       = {Doerfler, Robin and Wyse, Lonce},
  title        = {Analysis-Driven Procedural Generation of an Engine Sound Dataset with Embedded Control Annotations},
  booktitle    = {Proceedings of the 34th European Signal Processing Conference (EUSIPCO)},
  year         = 2026,
  address      = {Bruges, Belgium},
  publisher    = {EURASIP},
  pages        = {221--225},
  isbn         = {978-94-645936-3-1},
  url          = {https://eurasip.org/Proceedings/Eusipco/Eusipco2026/pdfs/0000221.pdf}
}

```

## Contact

For questions or issues regarding the dataset, please use the Hugging Face repository discussion page, the project page at https://rdoerfler.github.io/procedural-engine-sounds-page/, or refer to the associated publication.

## Acknowledgments

This dataset was created through procedural audio synthesis leveraging established principles from engine acoustics research, including engine order analysis, extended harmonic-plus-noise synthesis methodologies, and exhaust system resonance modeling. The synthesis methodology builds upon decades of foundational research in vehicle acoustics and internal combustion engine sound modeling. Special thanks to the digital signal processing and vehicle acoustics research communities for their foundational work that made this dataset possible.