clear
On-device speech enhancement for podcasters, video creators and voice apps.
Messy recording in, clean audio out. Reduce noise, clean up audio, normalize volume. Takes noisy 48 kHz mono or stereo audio (meeting recorders, bluetooth microphones, mobile device built-in mics, a laptop in a coffee shop) and returns rich, present, podcast-ready sound.
Runs loop-free on Apple Silicon via Core ML (iOS 17+, macOS 14+) and on Android via ONNX Runtime.
Sound
Trained to deliver a rich, present, close-miked podcast sound.
- Denoised. HVAC, keyboard clicks, mouse rustle, mic bumps, room hum, laptop fans, coffee shop background — all pulled down without chewing consonants.
- Dereverbed. Untreated bedrooms, offices and hotel rooms come out sounding closer to a treated studio. The model does not add reverberation of its own.
- Warm and present. Low-mids brought forward so voice sits comfortably in a mix rather than sounding thin or distant.
- Sibilance-safe. No harsh peaks introduced when cleaning up S / T / F consonants.
- No pumping or musical-noise artefacts. Trained with a large detail-preservation loss so breaths, plosives and vocal texture stay intact.
Variants
Two variants ship, sharing the exact
spec / feat_erb / feat_spec → spec_enhanced I/O contract.
clear-studio
The default. Quiet, studio-like character; silences sit close to true zero.
Best for solo podcasts, tutorials, voiceover, video demos, screen recordings, and anything that wants a clean broadcast feel.
| File | Purpose | Size |
|---|---|---|
clear-studio.mlpackage |
iOS / macOS Core ML source (fp16 + INT6 palette-6, iOS 17 target) | 9.2 MB |
clear-studio.mlmodelc |
Compiled Core ML, bundled inside the iOS SDK | 9.3 MB |
clear-studio.mlmodelc.zip |
On-demand download variant, same bytes | 8.8 MB |
clear-studio.onnx |
Android / cross-platform ONNX (fp16 weights, fp32 I/O) | 24 MB |
clear-studio.pt |
PyTorch checkpoint, for research and re-export | 46 MB |
clear-natural
Preserves room tone, breath, and lip texture.
For treated podcast studios, intentional voiceover, interviews where the room is part of the take, and remote guest recordings where absolute silence would sound wrong.
| File | Purpose | Size |
|---|---|---|
clear-natural.mlpackage |
iOS / macOS Core ML source | 9.2 MB |
clear-natural.mlmodelc |
Compiled Core ML | 9.3 MB |
clear-natural.mlmodelc.zip |
On-demand download variant | 8.8 MB |
clear-natural.onnx |
Android / cross-platform ONNX | 24 MB |
clear-natural.pt |
PyTorch checkpoint | 46 MB |
Measured performance
clear-studio on iOS 27.0, 5-min mono, full production pipeline (audio
in → enhanced audio out, R128 mastering + limiter, re-encode).
| Device | .all |
.cpuOnly (default) |
|---|---|---|
| iPhone 17 Pro | 109x | 114x |
| iPhone 16 Pro | 83x | 95x |
.cpuOnly measured highest on both devices in the release sweep and is
the SDK default.
Callers can still opt into .all or .cpuAndNeuralEngine when a
specific workload benefits.
Deployment target
- Apple — iOS 17.0+, iPadOS 17.0+, macOS 14.0+, visionOS 1.0+, via Core ML.
- Android — Android 8.0+ via ONNX Runtime.
- Other platforms — the ONNX file runs anywhere ONNX Runtime goes: Linux, Windows, Web (ORT-Web), Python, JVM.
Integration
- iOS / macOS.
clear-swift— Swift package; both variants bundled, works offline. - Android / JVM.
clear-kotlin— Kotlin SDK. - JavaScript / TypeScript.
@desert-ant-labs/clear— npm package for Node and browser.
Direct low-level use. Load the .mlpackage with Core ML or the
.onnx with ONNX Runtime. Feed spec / feat_erb / feat_spec (DFN3
input convention), read spec_enhanced, then ISTFT back to the time
domain. The SDKs handle STFT / features / ISTFT / R128 mastering for
you.
See also: desertant.com/models/clear.
What it's good for
- Meeting recorders. Zoom, Teams, Meet, Detail exports — single or multi-speaker.
- Bluetooth microphones. AirPods, Sony, headset mics.
- Mobile devices. iPhone and Android built-in microphone recordings, voice notes, field recordings.
- Laptop built-in microphones. MacBook and PC built-in mics.
- Untreated rooms. Bedrooms, hotel rooms, kitchens, coffee shops.
Whenever the pitch is messy recording in, clean audio out.
What it is not
- Not a general-purpose audio denoiser. Speech is the target; music, effects, and non-vocal signals get pulled down as noise.
- Not a source separator. Overlapping speakers stay overlapping.
- Not a voice changer, cloner, or transcription model.
Keywords
speech enhancement · noise suppression · dereverberation · speech denoising · reduce noise · clean up audio · normalize volume · turn a recording into studio sound · messy recording in clean audio out · podcast audio · voice cleanup · meeting recorder cleanup · bluetooth microphone cleanup · mobile device audio · built-in microphone · on-device audio · edge ML · Core ML · ONNX · iOS speech enhancement · Android speech enhancement · real-time speech enhancement · DFN3 · DeepFilterNet · studio sound · podcast sound · TCN · distilled model · Apple Neural Engine · ANE
- Downloads last month
- -