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 entirely on the Apple Neural Engine via Core ML and on Android via ONNX Runtime. The Core ML assets use the iOS 16 model format; the Swift SDK currently supports iOS 17+ and macOS 14+.

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. Their Core ML artifacts share the exact planar spec / feat_erb / feat_spec → spec_enhanced I/O contract and use a fixed batch of four independent two-second chunks. The ONNX artifacts retain their original DFN3 layout.

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.mlmodelc	ANE-optimized Core ML (fp16 compute + 6-bit weight palette, iOS 16 target)	9.0 MB
clear-studio.mlmodelc.zip	Same compiled model, zipped	8.6 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.mlmodelc	ANE-optimized Core ML (fp16 compute + 6-bit weight palette, iOS 16 target)	9.0 MB
clear-natural.mlmodelc.zip	Same compiled model, zipped	8.6 MB
clear-natural.onnx	Android / cross-platform ONNX	24 MB
clear-natural.pt	PyTorch checkpoint	46 MB

Performance

The Core ML variants are optimized for the Apple Neural Engine. MLComputePlan confirms that all 492 model operations run on ANE.

clear-studio, whole SDK pipeline on a 60-second clip, best of three:

Device	Realtime factor
iPhone 16 Pro	302x
MacBook Pro (M5)	345x

On iPhone 16 Pro, first-ever model loading takes approximately 3.4 seconds while Core ML compiles the ANE program. Cached launches load in approximately 62 ms; applications should warm the model in the background.

Deployment target

• Core ML model — iOS/iPadOS 16.0+ model format; ANE placement depends on hardware and OS.
• Swift SDK — iOS 18+, macOS 15+, tvOS 18+, visionOS 2+.
• Android — API 24+ (arm64-v8a, x86_64), via LiteRT.
• Other platforms — the .tflite runs wherever LiteRT does: Linux, Windows, and the browser through LiteRT.js. The ONNX files are kept for runtimes the SDK does not cover.

Integration

All three SDKs live in desert-ant-core and ship one version together.

• iOS / macOS. Swift Package Manager: depend on desert-ant-core and take its Clear product. The model is downloaded on first use and cached.
• Android / JVM. ai.desertant:clear on Maven Central.
• JavaScript / TypeScript. @desert-ant-labs/clear — WebAssembly + LiteRT.js in the browser, a prebuilt native core in Node.

Direct low-level use. Load the .mlmodelc with Core ML and feed planar fp16 tensors: spec (4,2,481,200), feat_erb (4,1,32,200), and feat_spec (4,2,96,200). Read spec_enhanced (4,2,481,200), then ISTFT back to the time domain. Batch elements are independent. The ONNX files keep the original DFN3 layout for cross-platform runtimes. Integrations must handle STFT, feature extraction, layout conversion, ISTFT, and mastering around the Core ML call.

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