license: cc-by-4.0
task_categories:
- audio-classification
language:
- en
tags:
- footstep-detection
- footstep-audio
- sound-event-detection
- audio-classification
- acoustic-recognition
- walking-sounds
- human-activity-recognition
- smart-home
- acoustic-biometrics
- footstep-biometrics
- person-identification
- surface-classification
- foley
- foley-synthesis
- environmental-sound
- real-world-audio
- WAV
- audio-dataset
- field-recordings
- PAD
size_categories:
- 1K<n<10K
pretty_name: Footstep Detection Audio Dataset
modality:
- audio
Footstep Detection Dataset — 50 Hours of Real Footstep Audio
50 hours of real footstep audio recordings for training footstep detection, sound event detection, and audio classification models. 166 manually verified files captured in natural indoor and outdoor conditions, with per-file metadata on surface, footwear, location, and background noise. The largest publicly listed footstep audio dataset — 3–5× larger than academic benchmarks (AFPILD: 10h, AFPID-II: 14h).
Contact us and share your feedback — receive additional samples for free! 😊
Key Highlights
- 50 hours of real-world footstep audio
- 166 manually verified files — every recording reviewed for clear footstep audibility
- Indoor + outdoor capture conditions
- 6 surface categories annotated per file
- 6 footwear categories annotated per file
- No synthetic audio, no augmentation, no AI-generated content
- Smartphone-first recordings (matches real deployment conditions)
Use This Dataset For
- Footstep detection — binary or multi-class footstep classifiers for smart home, security, and IoT
- Sound event detection (SED) — footstep as a target class in AudioSet-style models
- Acoustic person identification — biometric models recognizing individuals by walking sound
- Walking surface classification — distinguishing footsteps across floor materials
- Activity recognition — elderly care, fall detection, ambient assisted living
- Foley generation — training V2A models for walking sounds in games and animation
Dataset Structure
footstep-detection-dataset/
├── audio/
│ ├── rec_001.wav
│ ├── rec_002.wav
│ └── ... (158 WAV + 8 M4A files)
├── metadata.csv
└── README.md
metadata.csv schema
| Field | Type | Values |
|---|---|---|
file_id |
string | unique recording ID |
filename |
string | path to audio file |
duration_sec |
float | 10–100 seconds |
sample_rate |
int | 48000 (majority), 44100, 16000 |
channels |
int | 1 (mono) or 2 (stereo) |
format |
string | wav, m4a |
surface |
string | wood_laminate, tile, carpet, concrete_asphalt, stairs, other |
footwear |
string | barefoot, slippers, sandals, sneakers, dress_shoes_boots, other |
location |
string | indoor, outdoor |
noise_level |
string | low, medium, high |
device_class |
string | smartphone, laptop, tablet |
Dataset Statistics
| Metric | Value |
|---|---|
| Total duration | 50 hours |
| Total files | 166 |
| WAV files | 158 |
| M4A files | 8 |
| File duration range | 10–100 sec |
| Sample rates | 48 kHz / 44.1 kHz / 16 kHz |
| Surface categories | 6 |
| Footwear categories | 6 |
| Capture conditions | indoor + outdoor |
How This Compares to Academic Footstep Audio Datasets
| Dataset | Duration | Footstep samples | Metadata |
|---|---|---|---|
| Axon Labs Footstep Detection | 50 hours | 166 files | Surface + footwear + noise + location |
| AFPILD | 10 hours | 40 subjects | Location only |
| AFPID-II | 14 hours | 41 subjects | Clothing + shoes |
| FSD50K | <1h equivalent | 921 samples | None (label only) |
| ESC-50 | <0.1h equivalent | 40 samples | None (label only) |
| PURE | 14 minutes | 14 samples | 5 subjects |
Quick Start — Loading with 🤗 Datasets
from datasets import load_dataset
dataset = load_dataset("AxonData/footstep-detection-dataset")
print(dataset)
sample = dataset["train"][0]
print(sample["audio"]) # audio array + sampling_rate
print(sample["surface"]) # e.g. "wood_laminate"
print(sample["footwear"]) # e.g. "sneakers"
print(sample["noise_level"]) # e.g. "low"
Quick Start — PyTorch DataLoader
import torch
import torchaudio
from datasets import load_dataset
ds = load_dataset("AxonData/footstep-detection-dataset", split="train")
def collate(batch):
waveforms = [torch.tensor(item["audio"]["array"]) for item in batch]
labels = [item["surface"] for item in batch]
return waveforms, labels
loader = torch.utils.data.DataLoader(ds, batch_size=8, collate_fn=collate)
Sample vs Full Version
This HuggingFace repository contains a sample subset for evaluation. The full 50-hour dataset is licensed for commercial use through Axon Labs.
Full version of dataset is available for commercial usage — leave a request on our website Axonlabs to purchase the dataset 💰
What Makes This Dataset Unique
- Largest footstep audio corpus available commercially — 3–5× larger than the most cited academic alternatives
- Manually verified, not scraped — every file reviewed for clear footstep audibility
- Real smartphone recordings — matches deployment conditions for smart speakers, phones, wearables
- Structured metadata across 4 dimensions — supports filtered training and multi-task learning
- Backed by a biometric AI specialist — Axon Labs builds datasets used by 21% of iBeta 2025 certified companies
Two Dataset Versions Available
- Sample Version — open subset for EDA, evaluation, and proof-of-concept (this repo)
- Full Version — 50 hours of audio with complete metadata, licensed for commercial training
Contact us to choose the version that fits your project.
FAQ
Q: What's the largest publicly available footstep audio dataset? This one — 50 hours of curated recordings, 3–5× larger than AFPILD (10h) or AFPID-II (14h), which are the most cited academic benchmarks in the field. Sound event datasets like FSD50K and ESC-50 contain footsteps only as a small subset (under 1,000 samples).
Q: Can I use this dataset for footstep biometrics / acoustic person identification? Yes. The dataset is well-suited for footstep biometrics research, especially as a pre-training corpus. For per-subject identification tasks, we can collect additional per-subject sessions on request through our custom data collection service.
Q: What surfaces and footwear are covered? 6 surface types (wood/laminate, tile, carpet, concrete/asphalt, stairs, other) and 6 footwear types (barefoot, slippers, sandals, sneakers, dress shoes/boots, other). Every file is labeled across both dimensions.
Q: Is the data ethically collected? Yes. All recordings were captured with explicit participant consent and processed in accordance with GDPR. Full documentation of consent and provenance is available for the commercial version.
Citation
If you use this dataset in your research, please cite:
@misc{axonlabs2026footstep,
title = {Footstep Detection Audio Dataset},
author = {Axon Labs},
year = {2026},
url = {https://axonlab.ai/dataset/footsteps-audio-dataset/}
}
keywords: footstep audio dataset, footstep sound dataset, footstep detection dataset, sound event detection, audio classification dataset, acoustic person identification, footstep biometrics, walking surface classification, foley dataset, environmental sound dataset, real-world audio dataset, smart home audio, activity recognition
Visit us at Axonlabs to request a full version of the dataset for commercial usage.