A diagnose app for detect dysarthria with ML and Korean vocalization for Korean
- AVFoundation: Record user speech as 44.1kHz PCM WAV
- RosaKit: Extract Mel Spectrogram from raw audio signal
- Metal: Hardware-accelerated GPU pipeline for rendering spectrograms
- Firebase Firestore: Save and sync patient's diagnostic results
- PyTorch Mobile: Execute lightweight
.ptlmodels directly on iOS - Hierarchical Classification: Sequential inference pipeline (T00 to T03) for deep dysarthria analysis
graph TD
%% Audio Processing Pipeline
subgraph AudioPipeline [🎙️ Audio Preprocessing]
Mic[AVFoundation / AVAudioRecorder]
WAV[44.1kHz PCM WAV]
Rosa[RosaKit]
Mel[Mel Spectrogram]
Metal[Metal GPU Pipeline]
Mic -->|Record Voice| WAV
WAV -->|Signal Processing| Rosa
Rosa -->|Extract Feature| Mel
Mel -->|Hardware Accelerate| Metal
end
%% AI Inference Pipeline (Hierarchical)
subgraph AIPipeline [🧠 PyTorch Mobile Inference]
T00[Model T00: Binary Detection]
T01[Model T01: Type Classification 1]
T02[Model T02: Type Classification 2]
T03[Model T03: Final Severity]
Metal -.->|Input Tensor| T00
T00 -->|Hierarchical Flow| T01
T01 --> T02
T02 --> T03
end
%% Client & Backend
subgraph ClientBackend [📱 Client & ☁️ Firebase]
UI[UI & State Management]
AR[ARKit / Facial Tracking]
PDF[CoreGraphics / PDF Report]
Firestore[(Firebase Firestore)]
AR <-->|AR Speech Train| UI
T03 -->|Classification Result| UI
UI -->|Generate Report| PDF
UI <-->|Save Result| Firestore
end
| Layer | Original | 2026 Pick | Reason |
|---|---|---|---|
| Audio recording | AVAudioRecorder + manual WAV header | AVAudioRecorder (native WAV, no manual header) | Fewer silent bugs |
| Spectrogram | RosaKit | vDSP + Accelerate (already imported) | No CocoaPods dep, same performance |
| GPU rendering | Metal + custom shaders | Metal (keep) | Already well-done |
| ML inference | PyTorch Mobile Nightly | CoreML .mlpackage + Neural Engine |
2-5x faster, non CocoaPods, stable API |
| Crypto | CryptoSwift@main | Apple CryptoKit | Hardware-accelerated, no dep |
| Credentials | UserDefaults | Keychain via KeychainAccess (SPM) | Security baseline |
| Build system | CocoaPods + SPM hybrid | SPM only | After CoreML migration |
| Concurrency | Completion handlers + @MainActor blocking | Swift Concurrency + background actors | Prevents UI freeze |
| State machine | switch progress multi-call |
Single async throws inference function |
Safe, testable API |
Show Contents
Check your latest inspection results at a glance.
Check your dysarthria trends with graph
Check your inspection results history by date.
Export Inspection results as PDF
Inspect your dysarthria with speeching by word, sentence, paragraph, semi-free speech and free speech powered by ML
Get your inspection results powered by 95% or up accuracy ML Model, it's fast, accurate
Train your dysarthria with AR
Change your information, delete your data on server, change your disease info, sign out, and secession
Dysarthria Checker is compatible with these devices.
iPhone 14 Pro Max
iPhone 14 Pro
iPhone 14 Plus
iPhone 14
iPhone 13 Pro Max
iPhone 13 Pro
iPhone 13
iPhone 13 mini
iPhone 12 Pro Max
iPhone 12 Pro
iPhone 12
iPhone 12 mini
iPhone 11 Pro Max
iPhone 11 Pro
iPhone 11
iPhone Xs Max
iPhone Xs
iPhone XR
iPhone SE (3rd-Generation) (AR Train not supported.)
iPhone SE (2nd-Generation) (AR Train not supported.)
iPad Pro 12.9 (6th-Generation)
iPad Pro 11 (4th-Generation)
iPad Pro 12.9 (5th-Generation)
iPad Pro 11 (3rd-Generation)
iPad Pro 12.9 (4th-Generation)
iPad Pro 11 (2nd-Generation)
iPad Pro 12.9 (3rd-Generation)
iPad Pro 11 (1st-Generation)
iPad Pro 12.9 (2nd-Generation) (AR Train not supported.)
iPad Pro 10.5 (AR Train not supported.)
iPad Air (5th-Generation) (AR Train not supported.)
iPad Air (4th-Generation) (AR Train not supported.)
iPad Air (3rd-Generation) (AR Train not supported.)
iPad mini (6th-Generation) (AR Train not supported.)
iPad mini (5th-Generation) (AR Train not supported.)
iPad (10th-Generation) (AR Train not supported.)
iPad (9th-Generation) (AR Train not supported.)
- Required iOS/iPadOS 16 or up.