Whisper Voice — Live Speech-to-Text Mac App

Goal

Build and run a native macOS menu bar app that captures live microphone audio, transcribes it offline using WhisperKit (on Apple Silicon), and auto-types the text wherever the cursor is.

Inputs

Name	Type	Required	Description
model_size	string	No	Whisper model: tiny, base (default), small
language	string	No	"en" (default) or "hi" for Hindi mode
chunk_duration	float	No	Seconds per audio chunk (default: 3.0)

Process

1. Build the app

cd AiwithDhruv_Voice/WhisperAiwithDhruv
swift build

2. Run the app

swift run WhisperAiwithDhruv
# Or open in Xcode: open Package.swift → Cmd+R

3. First launch setup

1. Grant microphone permission when prompted
2. Grant Accessibility in System Settings → Privacy → Accessibility
3. Wait for model download (~140MB for base model)

4. Usage

• Cmd+Shift+Space — Toggle recording on/off
• Click mic icon in menu bar for controls
• Speak — text auto-types at cursor position
• Toggle Hindi mode for Hindi/Hinglish input

Outputs

Name	Type	Description
transcribed_text	string	Live transcribed text typed at cursor
history	array	Last 50 transcription entries in menu bar

Edge Cases

• No mic: Shows error in menu bar dropdown
• Accessibility denied: Auto-type disabled, manual copy from history
• Silence: VAD skips silent chunks (energy-based threshold)
• Hallucinations: Filters common Whisper artifacts ("Thank you.", "...")
• Model not downloaded: Shows download progress bar

Environment

• macOS 14+ (Sonoma)
• Apple Silicon (M1/M2/M3/M4)
• Xcode 15+ (for building)
• No API keys needed (fully offline)

Schema

Inputs

Name	Type	Required	Description
model_size	string	No	tiny / base / small
language	string	No	en / hi
chunk_duration	float	No	2.0 - 8.0 seconds
silence_threshold	float	No	0.002 - 0.05

Outputs

Name	Type	Description
transcription	string	Live text output
auto_typed	boolean	Whether text was injected at cursor

Credentials

Name	Source
None	Fully offline, no API keys

Composable With

video-edit (add transcription captions), send-telegram (send transcriptions to phone)

Cost

Free — runs entirely on-device. Model download is one-time (~140MB for base).