Use is limited to internal research and evaluation; commercial use,
redistribution, resale, and sublicensing are not permitted.
You will not use the audio to identify individuals or in any way that
invades privacy.
Full terms are in the accompanying LICENSE file.
Need the full dataset, a larger curated subset, or commercial rights? See "Going further" below, or reach us at data@prosoundeffects.com.

Pro Sound Effects — AI/ML Sample Dataset

The Most Comprehensive Audio Dataset for AI & Machine Learning.

A free, downloadable evaluation slice of the Pro Sound Effects (PSE) Audio Dataset: professionally recorded, 100% human-annotated, UCS-organized sound effects with clean, fully-owned commercial rights. This sample lets you benchmark PSE data against whatever you're using today before committing to anything.

Built and licensed by Pro Sound Effects — recording, annotating, and licensing professional audio since 2004.

TL;DR for researchers

What it is: A representative 102-clip / 18-minute subset spanning 28 categories, with the same uniform metadata schema used across the full 1.2M+ audio dataset.
Why it's different from crowdsourced/public audio: recordings by Oscar-winners' and masters of craft, uniform hand-tagged metadata (not noisy auto-labels), consistent recording quality, and unambiguous commercial rights — PSE owns the library outright, so there's no provenance or licensing gray area to inherit into your model.
Publishing: The Research License explicitly permits publication with citation (see Going further). Cite as below.
License (this sample): non-commercial evaluation only — see LICENSE.

What's in this sample


Clips	`102`
Total duration	`~18 mins`
Categories covered	`28 — top ones are Swooshes, Fight, Animals, Destruction, Electricity, Glass, Ambience, Explosions`
Format	`WAV · 48 / 96 / 192 kHz · 16- and 24-bit · mono, stereo, and some multichannel`
Metadata fields	`file_name, Keywords, Description, Category, SubCategory, Channels, BitDepth, SampleRate, Duration, Library`

Each row pairs an audio file with structured, human-written metadata. A metadata.csv (or metadata.jsonl) maps every file to its tags so the dataset loads directly with 🤗 datasets.

The full private PSE Audio Dataset (what this sample is drawn from)

1.2M+ exclusive, discrete sounds — growing toward 3M+ by 2028
5,000+ hours, ~5.8 TB of audio
650+ categories, optimized for the Universal Category System (UCS)
100% human-annotated with uniform metadata (100K+ editorial hours)
Full rights owned by PSE — private, not aggregated, not scraped, not crowdsourced

Built for AI use cases

The library is structured to support, among others:

Speech recognition & voice AI — assistants, transcription, voice auth
Environmental sound recognition — accessibility, assistive tech, IoT, transportation
Active noise cancellation & audio separation — comms, broadcast, restoration
Generative audio & text-to-sound — text-to-SFX, creative AI, music generation
Audio classification & tagging — moderation, intelligent classification
Dynamic retrieval & RAG — adaptive playback, audio search

Why the rights & provenance matter

Training-data provenance is now a real legal and reputational exposure for AI teams. PSE is built to remove that risk:

Full ownership of rights — cleanly licensable for commercial use, unlike Creative-Commons-mixed or scraped sources
Artists are compensated — licensing revenue flows back to the creators
Affiliations: Fairly Trained · Dataset Providers Alliance · Human Artistry Campaign · Content Authenticity Initiative

"Human creativity is the foundation of AI's existence. Our mission is to help creators bring ideas to life — ethically and at scale." — Douglas Price, CEO

Quickstart

from datasets import load_dataset

# Requires a Hugging Face token (gated dataset); accept the terms once on the Hub.
ds = load_dataset("ProSoundEffects/pse-audio-sfx-dataset", split="train")
print(ds)
print(ds[0])  # {'audio': {...}, 'category': ..., 'description': ..., ...}

Full PyTorch DataLoader examples, metadata walkthroughs, and integration recipes live in the companion GitHub repo: <<FILL: https://github.com/ORG/pse-audio-sample>>

Going further

This sample is intentionally small — enough to validate quality and fit. When your prototype shows the data earns its place, here are the paths:

Bigger curated subset / category analysis — tell us your use case and we'll scope a tailored slice. Email data@prosoundeffects.com.
Research License (non-commercial R&D) — publication permitted with citation. Indicative terms: $2,500/mo · $10,000/yr · $25,000 one-time buyout (program one-pager). Starting at $15,000/yr for ongoing programs.
Commercial licensing — Annual, Perpetual, or Revshare structures, scoped to your stage and deployment. Custom GenAI / non-GenAI terms.
Custom curation & bespoke recording — if you need coverage we don't have off the shelf.

Start here: prosoundeffects.com/machine-learning-ai · prosoundeffects.com/data · book a call

Citation

@misc{prosoundeffects_aiml_sample,
  title  = {Pro Sound Effects --- AI/ML Sample Dataset},
  author = {{Pro Sound Effects}},
  year   = {2026},
  howpublished = {Hugging Face Hub},
  url    = {https://huggingface.co/datasets/ProSoundEffects/pse-audio-sfx-dataset}
}

Contact

Data / technical: data@prosoundeffects.com
Web: prosoundeffects.com/data

Downloads last month: 32

Total file size:

312 MB