SpeechDemo

SpeechDemo

A minimal SwiftUI macOS app demonstrating the Qwen3 Speech library: mic recording with ASR transcription and text-to-speech synthesis with playback.

Requirements

• macOS 15+ (Sequoia)
• Apple Silicon (M1 or later)
• Xcode 16+

Build & Run

1. Open Examples/SpeechDemo/ in Xcode (File > Open, select this folder)
2. Xcode resolves the SPM dependency on the parent Qwen3Speech package automatically
3. Select the SpeechDemo scheme and press Run

Entitlements (Xcode project only)

If running from an Xcode project, add these entitlements to the signing config:

• com.apple.security.app-sandbox — required for App Sandbox
• com.apple.security.device.audio-input — microphone access for recording
• com.apple.security.network.client — outbound network for model downloads

The entitlements file is provided at SpeechDemo/SpeechDemo.entitlements.

Command-line build

cd Examples/SpeechDemo
swift build

Note: The command-line build produces an executable, not a sandboxed .app bundle. Microphone access requires running from Xcode with proper entitlements, or granting Terminal microphone permission in System Settings > Privacy & Security > Microphone.

Usage

The app has three tabs:

Dictate

1. Select an ASR engine: Parakeet TDT (CoreML, ~400 MB) or Qwen3-ASR (MLX, ~400 MB)
2. Click Load to download the model (first run only — cached in ~/Library/Caches/qwen3-speech/)
3. Hold the Record button, speak, then release
4. The transcription appears below — click Copy to copy to clipboard

Speak

1. Click Load Qwen3-TTS to download the model (~1 GB, first run only)
2. Type or paste text into the text field
3. Select a language from the dropdown
4. Click Synthesize — audio plays automatically when ready

Echo (Voice Pipeline)

1. Click Load Models — downloads Silero VAD (CoreML), Parakeet TDT (CoreML), and Qwen3-TTS (MLX)
2. Click Start to begin the voice pipeline
3. Speak into the mic — the pipeline detects speech, transcribes it, and speaks it back
4. Full-duplex: you can interrupt the agent while it's speaking (barge-in)
5. Click Stop to end the session

The Echo tab uses the SpeechCore voice pipeline with AEC (acoustic echo cancellation) for full-duplex operation.

Models downloaded on first run

Model	Size	Cache location
Parakeet TDT (CoreML INT8)	~500 MB	~/Library/Caches/qwen3-speech/aufklarer_Parakeet-TDT-v3-CoreML-INT8/
Qwen3-ASR (MLX 4-bit)	~400 MB	~/Library/Caches/qwen3-speech/aufklarer_Qwen3-ASR-0.6B-MLX-4bit/
Qwen3-TTS (MLX 4-bit)	~1 GB	~/Library/Caches/qwen3-speech/aufklarer_Qwen3-TTS-12Hz-0.6B-Base-MLX-4bit/
Qwen3-TTS Tokenizer	~650 MB	~/Library/Caches/qwen3-speech/Qwen_Qwen3-TTS-Tokenizer-12Hz/
Silero VAD v6.2.1 (CoreML)	~1.2 MB	~/Library/Caches/qwen3-speech/