Speecher is an edge-native voice-to-text transcription application designed to facilitate communication between speaking individuals and those who are deaf or hard of hearing. The app runs entirely on-device using local Whisper models for real-time speech recognition with zero cloud dependency.

• ✅ 100% Offline & Private - All processing happens on-device, no internet connection required
• 🎤 Real-time Edge Inference - Convert speech to text instantly using local models
• 🇵🇹 Portuguese Language Support - Built for Portuguese speech recognition, other languages available through the Whisper CPP Model
• 📱 Clean, Accessible Interface - Large text display for easy reading
• 🔄 Simple Recording Controls - One-tap recording with visual feedback
• ⚡ Metal GPU Acceleration - Optimized edge performance on iOS devices
• 🔒 Privacy-First - No data leaves your device, complete edge computing solution

The app leverages edge computing principles with:

• SwiftUI - Modern iOS user interface
• Whisper.cpp - Local edge speech recognition engine - the small size model works offers a good balance of size (500MB) with speed (tested in iPhone PRO MAX 14)
• AVFoundation - On-device audio recording and processing
• Metal - GPU acceleration for edge model inference
• Local Storage - Edge model deployment with no external dependencies

• iOS 16.4+ (iPhone 13+) / macOS 13.3+
• Xcode 15+
• ~500MB edge storage space for the small Whisper model
• Microphone permissions

git clone --recursive https://github.com/lazaroborges/speecher.git
cd speecher

First, build the Whisper.cpp XCFramework for edge deployment on iOS:

cd whisper.cpp
./build-xcframework.sh

This will create the necessary framework files optimized for edge computing.

Download the Portuguese-optimized model for edge inference in the desired size:

cd whisper.cpp/models
./download-ggml-model.sh <model_size>

This will download ggml-<model_size>.bin (~1.4GB) for local edge deployment.

1. Open speecher.xcodeproj in Xcode
2. Right-click on the Resources/models/ folder in the project navigator
3. Select "Add Files to 'speecher'"
4. Navigate to whisper.cpp/models/ and select ggml-<model_size>.bin
5. Ensure "Add to target: speecher" is checked
6. Click "Add"

1. In Xcode, select your project in the navigator
2. Go to your target's "General" tab
3. Under "Frameworks, Libraries, and Embedded Content", click "+"
4. Click "Add Other..." → "Add Files..."
5. Navigate to whisper.cpp/ and add the generated whisper.xcframework
6. Set "Embed & Sign" for the framework

The app requires microphone access for edge processing. The Info.plist should include:

<key>NSMicrophoneUsageDescription</key>
<string>Speecher needs microphone access to transcribe your speech to text using edge computing.</string>

1. Launch the app - The Whisper model will load automatically
2. Tap the blue microphone button - Start recording
3. Speak clearly - The recording indicator will show red
4. Tap the red stop button - End recording and process speech
5. View transcription - Large text appears in full-screen mode
6. Record again or clear - Use the control buttons to continue

You can deploy different Whisper models based on your edge computing needs:

To deploy a different edge model:

# Download the desired edge model
cd whisper.cpp/models
./download-ggml-model.sh small

# Update WhisperState.swift to use the new edge model
# Change "ggml-<model_size>.bin" to "ggml-small.bin" in modelUrl

speecher/
├── ContentView.swift          # Main UI with edge recording interface
├── WhisperState.swift         # Edge speech recognition state management
├── LibWhisper.swift          # Whisper.cpp at the edge integration layer
├── Recorder.swift            # Edge audio recording functionality
├── RiffWaveUtils.swift       # WAV file processing for edge inference
├── speecherApp.swift         # Edge app entry point
└── Resources/
    └── models/               # Edge model storage
        └── ggml-<model_size>.bin   # Deployed edge model file

The app is configured for Portuguese edge processing by default. To change the language for edge inference, modify LibWhisper.swift:

// Line 35 in LibWhisper.swift
params.language = pt  // Change "pt" to your language code for edge processing

Supported model language codes: en, es, fr, de, it, pt, ru, zh, etc.

Adjust threading for optimal edge computing performance in LibWhisper.swift:

// Line 25 - Modify thread count based on edge device capabilities
let maxThreads = max(1, min(8, cpuCount() - 2))

• Verify ggml-<model_size>.bin is properly deployed in the app bundle
• Check edge model file size (~500MB for small model)
• Ensure edge model is added to Xcode target

• Speak clearly and close to device during edge processing
• Minimize background noise for better edge inference
• Consider deploying a larger edge model (large-v3)
• Check microphone permissions for edge recording

• Deploy a smaller edge model (small or base)
• Close other apps to free memory for edge computing
• Ensure device has sufficient storage for edge models

• Update Xcode to latest version for edge framework support
• Clean build folder (⌘+Shift+K)
• Verify whisper.xcframework is properly linked for edge deployment

1. Fork the repository
2. Create a feature branch for edge improvements
3. Make your edge computing changes
4. Test thoroughly on edge devices
5. Submit a pull request

This edge computing project uses the MIT License. The Whisper.cpp library is also under MIT License.

• OpenAI Whisper - Original speech recognition model
• whisper.cpp - Efficient C++ implementation for edge computing
• ggerganov - whisper.cpp author and edge computing pioneer

For edge computing issues and questions:

1. Check the edge troubleshooting section above
2. Search existing GitHub issues for edge-related problems
3. Create a new issue with detailed edge deployment information

Made with ❤️ in Brazil for the deaf and hard of hearing community using computing