SplitSafe - Audio ML Leakage Auditor

SplitSafe - Audio ML Leakage Auditor

Checks the train, validation and test splits of an audio dataset for leaked copies and shared speakers, using audio fingerprints that need no GPU. For people who build audio machine learning models.

Made by boobalanit on Hugging Face. These notes are Mindket's, written from its Space page and README; the app's own files and license are on Hugging Face.

About

What it is: SplitSafe audits the train, validation and test splits of an audio machine learning dataset. It looks for leakage, where the same recording sits on both sides of a split, and for the same speaker appearing on both sides. Catching this early saves GPU time that would otherwise go into misleading benchmark scores.

Who it suits: It suits people who build or evaluate audio models and want to trust their benchmark numbers, and anyone who reviews such datasets before they are used.

Industry | Who uses it | How they use it | Example
Machine learning | Audio model builders | Check the splits before training starts | Audit a speech dataset for repeated clips
Research | Benchmark authors | Confirm that test audio is not in the training set | Check a new test split before publishing it
Speech technology | Dataset curators | Look for a speaker who appears on both sides of a split | Find a voice shared by train and test
Software | Developers | Read the fingerprinting code in the splitsafe folder | Reuse the idea in their own pipeline

You can skip it if your data is not audio, or if you have no train, validation and test split to compare.

Where to use it:

• A web browser: open the Space page on Hugging Face, where it runs on zero-a10g hardware.
• Windows, macOS or Linux: run a copy with Python.

Use case | Who | What they get back
Find leaked audio | Dataset owners | A check for copies across train, validation and test
Spot shared speakers | Speech researchers | A check for one speaker on both sides of a split
Run it privately | Developers | A local copy started with python app.py

How to set it up: The hosted page needs nothing from you. A local copy needs these two things first.

• Python on Windows, macOS or Linux.
• pip, which installs the packages.

1. Install the huggingface_hub package with pip.
2. Download the Space's files into a folder with the hf command.
3. Move into that folder.
4. Install Gradio 5.20.0, then the packages listed in requirements.txt.
5. Start the app with Python.

pip install -U huggingface_hub
hf download boobalanit/splitsafe --repo-type space --local-dir splitsafe
cd splitsafe
pip install gradio==5.20.0
pip install -r requirements.txt
python app.py

How to use it: Online, open the Space page. On your own computer, run the last command above inside the downloaded folder. Then give the app your audio splits and read what it reports about leaked clips and shared speakers.

Tool | What it does
Audio fingerprinting | Matches clips by their sound pattern, with no GPU
Leakage check | Finds copies across train, validation and test
Speaker check | Finds a speaker present on both sides of a split

Example: You have a speech dataset split into train, validation and test. You give the splits to the app. It is built to flag a clip that was re-encoded or made quieter and still sits in both train and test, so you can remove it before you spend GPU time on training.

For newcomers and experts: If you are new, think of it as a duplicate finder for audio data. Start with a small dataset and see what it flags. A clean split should come back with nothing flagged.

If you know the field, it uses deterministic landmark constellation fingerprinting, so there is no learned model and no GPU in the matching. It still catches copies that were re-encoded, resampled, made quieter, trimmed or cut into short sub-clips. A separate guard looks for a speaker present on both sides of a split.

You provide the audio splits
Each clip gets a fingerprint
Fingerprints are compared across splits
Leaks and shared speakers are reported

Good to know: The fingerprinting is signal processing only, so results are repeatable. The README does not describe the input format or the output layout.

• License: mit
• Cost to run: none named
• Data it reaches: the audio you give it

On Hugging Face, 13 November 2026

• Likes: 1
• Built with: gradio
• License: mit

Links

• Space: https://huggingface.co/spaces/boobalanit/splitsafe
• Files: https://huggingface.co/spaces/boobalanit/splitsafe/tree/main

Try it

Open the Space in your browser, or copy it to your own Hugging Face account from its page. Nothing in this kit runs by itself.