Compact
DeepFilterNet3
Noise removal · Best Quality
A low-complexity speech-enhancement network that filters noise in the frequency domain while preserving voice detail.
Runs when you pick Best Quality denoise.
Local AI
The AI runs on your computer. Here’s how.
Smart Cleanup is powered by neural models — but unlike most “AI” tools, none of them live in the cloud. They download once and run on your own hardware, so confidential client audio never leaves the building. Here’s the architecture, model by model.
Your recording
On-device models
Cleaned audio
your machine · no network
Why it has to be local
Court recordings are privileged. Sending them to a third-party AI service — even briefly — means client audio touches someone else’s servers, logs, and retention policy. DepoAudio’s answer is to never make that trade: the models come to your machine instead of your audio going to theirs. No account, no upload, no network connection required — it works on a sealed laptop in a deposition room.
How it works
Bundled ONNX Runtime
Models execute through ONNX Runtime shipped with the app — no Python, no separate install, no service phoning home.
Download on demand
Install only the models you use, straight from Settings. Skip the ones you don’t — nothing is forced into the download.
SHA-256 verified
Every model is checksum-verified before it’s ever run, so a corrupted or tampered file is rejected.
Optional by design
No ONNX Runtime on your system? AI features simply hide — DepoAudio still converts audio normally.
The models
Six models cover the cleanup pipeline. Install the ones you need; each runs only at the step it powers.
Small
Fast denoise
Noise removal · Fast
A lightweight spectral suppressor for instant cleanup when you don’t need the deepest pass.
Runs when you pick Fast denoise.
Medium
DCCRN+
Echo & reverb
A deep complex convolutional recurrent network that strips room reverb and echo for natural-sounding speech.
Runs when De-reverb is on.
Medium
FlashSR
Clarity · bandwidth extension
A one-step audio super-resolution model that rebuilds the high-frequency detail lost on phone and narrow-band recordings.
Runs when Enhance detects narrow-band audio.
≈300 KB
DNSMOS
Quality scoring
A non-intrusive MOS estimator that predicts a 1–5 quality score with no clean reference needed.
Runs on Scan to rate the recording.
≈38 MB
Speaker detection
Turn detection
A segmentation model that marks where each speaker starts and stops, counts speakers, and measures speech-to-silence ratio.
Runs when Turn detection is on.
Alongside the neural models, the chain includes plain DSP steps that need no model at all — high-pass filter, loudness normalize to −16 LUFS, de-clip, auto-level, silence trim, and fade.
Hardware acceleration
DepoAudio detects the fastest path on your machine and uses it automatically — falling back to CPU when nothing else is available, so the features always work.
Apple Neural Engine
Apple-silicon Macs
AMD Ryzen AI
XDNA NPUs
Intel AI Boost
Core Ultra NPUs
GPU
Discrete or integrated
CPU
Always-available fallback
The privacy guarantee
During processing DepoAudio makes zero network calls. Temporary files created while a model runs are deleted the moment the operation finishes. Nothing about your audio — not the file, not a transcript, not a fingerprint — is transmitted anywhere. The only time the app touches the network at all is an optional check for app updates, which sends nothing but the current version number.
Smart Cleanup · v0.8.0
Clean up a recording — on your machine, in seconds.
Free · Windows 10/11 & macOS 12+ · see what Smart Cleanup does