AADD

AADD

Detects 12 kinds of audio degradation in speech recordings every 20 ms, such as noise, reverb, clipping and codec artefacts. Meant for cleaning speech datasets and auditing recordings.

Made by gertelrina on Hugging Face. These notes are Mindket's, written from its model page and model card; the model's own files and license are on Hugging Face.

About

What it is: AADD, short for Automatic Audio Degradation Detection, checks speech recordings and marks which of 12 problems are present and where. Examples include noise, reverb, clipping, packet loss, hum and clicks, plus changes to pitch and speed. Several can show up at once.

Who it suits: Teams that build speech datasets for recognition or synthesis, and engineers who audit recordings for quality problems.

Industry | Who uses it | How they use it | Example
Speech AI | ML engineers | Filter noisy or damaged clips out of a training set | A report on a folder of recordings
Telecom | Audio engineers | Spot packet loss and codec artefacts in call audio | A call checked in 0.1 s segments
Media | Sound editors | Find clipping, clicks or hum in recordings | A podcast session scanned for hum
Research | Audio researchers | Study degradation detection with a Whisper-based model | A test on noise and reverb benchmarks

If you only need to remove noise from audio, you can skip it, since this tool detects problems and does not repair them.

Where to use it:

• The aadd command line tool, on Windows, macOS or Linux with Python
• Python programs, through the Detector class
• Speech dataset pipelines for ASR and TTS

Use case | Who | What they get back
Dataset cleaning | ML engineers | An HTML report on a folder of recordings
Frame-level detection | Audio engineers | A probability for each of 12 classes every 20 ms
Segment export | Analysts | A CSV file of detections per fixed time segment

How to set it up:

• Python and pip
• The aadd package from the project's GitHub page, which holds the code and command line
• Room for the checkpoint, which downloads on first use

1. Install the Hugging Face hub tools with pip.
2. Download the checkpoint with the hf command, or let the tool fetch it on first use.
3. Install the aadd package from its GitHub page.
4. Run aadd scan on a file or folder.

pip install -U huggingface_hub
hf download gertelrina/aadd

How to use it: Run aadd scan on a folder with a report option, or on one file with a finer resolution. The default model is v3.1. In Python, create a Detector and call predict_file.

aadd scan my_dataset/ --report report.html              # downloads this checkpoint on first use
aadd scan call.wav --resolution 0.1 --segments per_100ms.csv        # v3.1 is the default model

from aadd import Detector

det = Detector("v3", resolution=0.1)       # segments of 0.1 s (1.0 by default); 20-ms frames in res.frame_probs
res = det.predict_file("call.wav")
res.detected                                # {'noise': 0.97, 'reverb': 0.88}
res.frame_probs.shape                       # (n_frames_20ms, 12)

Tool | What it does
aadd scan | Scans a file or folder and writes a report or a segment file
Detector | Python class that loads the model
predict_file | Returns detected classes and frame probabilities
resolution | Sets the segment length in seconds

Example: You run aadd scan on call.wav with a resolution of 0.1 and a segments file. You get a CSV with detections per 100 ms, so you can see where noise or reverberation starts and ends.

For newcomers and experts: Start by scanning one short recording and opening the report. The default threshold is conservative on mild noise, so lower it when filtering datasets.

The model uses the encoder of Whisper-small, with 88M parameters, fine-tuned end to end. For every 20 ms frame it gives one score per class, working on 8-second windows, and sliding windows let it cover recordings of any length. The latest version cut false music detections in babble from cafes and buses, with noisy clips flagged as music falling from 42% to 28% on VoiceBank-DEMAND.

Give it a recording
Whisper-small encoder reads 8 s windows
One score per class per 20 ms frame
A report or segment file lists what it found

Good to know:

• License: MIT
• Cost to run: none named
• Data it reaches: the audio files you scan, and Hugging Face for the checkpoint
• Not for: other languages, since training covered English and Russian
• Limits: it reports problems but does not repair audio, and mild noise may need a lower threshold

Title: ROC-AUC for noise or reverberation in the card's benchmarks
VoiceBank-DEMAND noise: 0.992
DNS 2020 noise: 0.998
DNS 2020 reverberation: 1.000
Dynamic-SUPERB noise: 0.992

On Hugging Face, 13 November 2026

• Task: audio classification
• Library: pytorch
• Downloads in the last 30 days: 73
• Likes: 9
• License: mit
• Languages: en, ru
• Based on: openai/whisper-small
• Tags: pytorch, aadd, audio, speech, audio-quality, degradation-detection, data-curation, sound-event-detection, whisper, audio-classification

Links

• Model: https://huggingface.co/gertelrina/aadd
• Files: https://huggingface.co/gertelrina/aadd/tree/main

Get the model

Download its files with Hugging Face's command line:

pip install -U huggingface_hub
hf download gertelrina/aadd

Then follow the usage in its model card. Nothing in this kit runs by itself.