VelociDictate

AI-Powered Hands-Free Technical Dictation

VelociDictate is a cross-platform speech-to-text application designed for technical professionals. It combines OpenAI's Whisper for local audio transcription with Anthropic's Claude for intelligent correction of technical terminology, acronyms, and domain-specific language.

The Problem

General-purpose speech-to-text consistently fails on technical content:

• "oh SPF" instead of OSPF
• "B G P" instead of BGP
• "ten dot zero dot zero dot one" instead of 10.0.0.1
• "V LAN one hundred" instead of VLAN 100
• "eight oh two dot one Q" instead of 802.1Q

VelociDictate solves this with a two-stage pipeline: Whisper handles the speech recognition, Claude corrects the technical terms.

Features

Hands-Free Operation Voice-activated recording with customizable start/stop phrases. Say "start now" to begin dictating, "stop now" to transcribe and paste. No keyboard or mouse required.

Technical Accuracy Domain-specific vocabulary correction for networking, software development, IT operations, project management, and creative writing. Extensible with custom vocabulary packs.

Auto-Paste Mode Transcriptions paste directly into whatever application has focus. Dictate into emails, chat windows, IDEs, terminals, or documents without switching context.

Local Processing Whisper runs entirely on your machine. Audio never leaves your system. Only the text transcript is sent to Claude for refinement.

System Tray Application Runs quietly in the background. Tray icon indicates state: green (ready), orange (listening for wake word), red (recording).

---

How It Works

Microphone → Wake Word Detection → Recording → Transcription → Refinement → Auto-Paste
              (Whisper tiny)                    (Whisper small)  (Claude API)

1. Wake Word Detection - A lightweight Whisper "tiny" model continuously monitors for voice commands
2. Recording - Audio captured from start phrase until stop phrase detected
3. Transcription - Whisper "small" model converts speech to text
4. Refinement - Claude corrects technical terminology using domain-specific prompts
5. Output - Text copied to clipboard and optionally pasted into the active window

---

Installation

Requirements

• Python 3.12+
• Anthropic API key
• Microphone

Install from PyPI

pip install velocidictate

Install from Source

git clone https://github.com/scottpeterman/velocidictate.git
cd velocidictate
pip install -e .

Linux Dependencies

sudo apt install libportaudio2 portaudio19-dev xdotool xclip

API Key Setup

export ANTHROPIC_API_KEY='sk-ant-...'

Or configure through Settings → API Keys after launching.

---

Usage

Starting the Application

velocidictate

A system tray icon appears. Right-click for the menu.

Hands-Free Dictation

1. Right-click tray → Enable Hands-Free Mode
2. Right-click tray → Settings → Output → Enable Auto-paste
3. Focus on any text field (email, chat, IDE, document)
4. Say "start now"
5. Dictate naturally
6. Say "stop now"
7. Text appears in the focused field

Manual Mode

• Click tray icon to start recording
• Click again to stop and transcribe
• Or use the global hotkey (default: Ctrl+Shift+R)

Example

You say:

"Start now. Configure the BGP neighbor at 10.0.0.1 with remote AS 65001 period. Enable OSPF on VLAN 100 comma area zero period. Stop now."

Result:

Configure the BGP neighbor at 10.0.0.1 with remote AS 65001. Enable OSPF on VLAN 100, area 0.

---

Screenshots

System Tray Menu

Wake Word Settings

Vocabulary Manager

Help - Dictation Commands

---

Vocabulary Packs

VelociDictate includes domain-specific vocabulary correction. Select your domain in Settings → Vocabulary.

Built-in Domains

Domain	Description
networking	BGP, OSPF, VLAN, firewalls, Cisco/Juniper/Arista terminology
development	Python, JavaScript, APIs, Docker, Kubernetes, cloud platforms
it-general	Windows, Linux, Active Directory, helpdesk, sysadmin
project-management	Agile, Scrum, SAFe, Kanban, Jira, stakeholder terminology
creative-writing	Latin phrases, Shakespeare, literary terms, classical references
general	Basic punctuation and capitalization without domain focus

Extended Vocabulary Packs

Download comprehensive vocabulary packs with hundreds of terms and their common misrecognitions:

Settings → Vocabulary → Download Vocabulary Packs

Or manually from: https://github.com/scottpeterman/velocidictate

Custom Vocabulary

Create your own vocabulary file (one entry per line):

OSPF, oh SPF, OSP F, O S P F
BGP, B G P, beep, bee gee pee
Kubernetes, K 8 S, K8s, kube, kubernetes

Point to it in Settings → Vocabulary → Additional Custom Vocabulary.

---

Configuration

Config File Location

Platform	Path
Linux	~/.config/velocidictate/config.toml
Windows	%APPDATA%\velocidictate\config.toml
macOS	~/Library/Application Support/velocidictate/config.toml

Settings Dialog

Right-click tray → Settings

API Keys - Claude and OpenAI credentials

Whisper - Model size (tiny/base/small/medium/large-v3), device (cpu/cuda)

Wake Word - Start/stop phrases, detection sensitivity

Output - Auto-paste toggle, notifications

Vocabulary - Domain selection, custom vocabulary files

Hotkeys - Global keyboard shortcuts

---

Voice Commands

Speak these words to insert punctuation:

Say	Result
period	.
comma	,
question mark	?
exclamation point	!
colon	:
semicolon	;
new paragraph	(blank line)
new line	(line break)
open paren	(
close paren	)
quote / open quote	"
end quote / close quote	"
dash	-

---

Architecture

Components

Component	Technology	Purpose
Audio Capture	sounddevice	Microphone input
Wake Word	Whisper tiny	Voice command detection
Transcription	Whisper small (faster-whisper)	Speech to text
Refinement	Claude API	Technical term correction
Output	pyperclip, xdotool	Clipboard and auto-paste
GUI	PyQt6	System tray application

Design Decisions

Two Whisper Models: The tiny model (~75MB) runs continuously for wake word detection with minimal CPU impact. The small model (~460MB) provides higher accuracy for actual transcription but only runs after recording stops.

Claude Refinement: Speech-to-text engines consistently fail on technical terminology. Claude's language understanding corrects domain-specific terms that would otherwise require extensive post-processing rules.

Voice-Activated: True hands-free operation enables dictation while typing, reading documents, or working in any application without switching focus.

---

Troubleshooting

No Audio Captured

# List available devices
velocidictate --list-devices

# Check PulseAudio/PipeWire
pactl list sources short

Wake Word Not Detected

• Speak clearly with a brief pause before the phrase
• Reduce background noise
• Increase sensitivity in Settings → Wake Word
• Try different wake phrases

Auto-Paste Not Working (Linux)

# Install xdotool
sudo apt install xdotool

# For Wayland, xdotool may not work
# Use X11 or install wtype

High CPU Usage

Expected in hands-free mode (tiny model runs continuously). Disable hands-free mode when not dictating. Consider CUDA acceleration if available.

Claude Errors

• Verify API key: echo $ANTHROPIC_API_KEY
• Check quota at console.anthropic.com
• Refinement is optional; raw Whisper output still works without Claude

---

Command Line Options

velocidictate                    # Launch GUI
velocidictate --debug            # Enable debug output
velocidictate --list-devices     # List audio devices

---

Project Structure

velocidictate/
├── audio/
│   ├── capture.py           # Microphone recording
│   └── wake_word.py         # Voice-activated detection
├── transcription/
│   └── whisper_local.py     # Whisper inference
├── refinement/
│   └── claude.py            # Claude API integration
├── output/
│   └── clipboard.py         # Clipboard and auto-paste
├── ui/
│   ├── tray.py              # System tray application
│   ├── settings_dialog.py   # Settings interface
│   ├── help_dialog.py       # Help documentation
│   └── config.py            # Configuration management
└── vocabularies/
    └── networking.txt       # Default vocabulary

---

License

GPLv3 - This project uses PyQt6 which is licensed under GPL.

Author

Scott Peterman

Links

• Repository: https://github.com/scottpeterman/velocidictate
• Vocabulary Packs: https://github.com/scottpeterman/velocidictate
• Issues: https://github.com/scottpeterman/velocidictate/issues