vaani 0.2.0

macOS menu bar app that captures voice, transcribes, and enhances text with AI

Vaani · वाणी

वाणी — Sanskrit for speech, voice; the goddess of language and learning.

Voice to polished text, right at your cursor — anywhere on macOS.

Vaani is a macOS menu bar app that listens when you hold a hotkey, transcribes your speech with OpenAI Whisper, enhances it with Claude AI, and pastes the result directly at your cursor. No switching apps, no copy-pasting — just speak and the text appears.

---

The Problem

Typing is slow. Dictation on macOS gives you raw, unedited speech dumps — filler words, broken sentences, no punctuation. Third-party dictation tools are either expensive subscriptions, cloud-locked, or produce output that still needs manual cleanup.

Professionals who write a lot — engineers, writers, product managers, support teams — spend significant time translating their thoughts into polished prose. The gap between what you think and what ends up on screen is real friction.

Why We Built Vaani

We wanted voice to be a first-class input method, not an afterthought. The goal: speak naturally, get back something you'd actually send or commit. Vaani sits silently in your menu bar and activates on a global hotkey — no window to focus, no app to switch to. It works in any text field, terminal, IDE, browser, Slack, email client, or document editor.

The name "Vaani" (वाणी) comes from Sanskrit, meaning speech or voice — the goddess of language and learning.

---

Quick Start

pip install vaani
python -m spacy download en_core_web_sm
vaani start

On first launch, Vaani walks you through entering your API keys and granting macOS permissions. Then:

1. Hold Alt (or your configured hotkey) and speak
2. Release — Vaani transcribes and enhances your speech
3. Polished text appears at your cursor in ~2–4 seconds

API keys needed: OpenAI (transcription) · Anthropic (enhancement). Keys are stored in macOS Keychain — never written to disk in plaintext.

---

How It Works

Hold hotkey → Mic capture → VAD trims silence → Gain normalization
   → OpenAI Whisper (transcribe) → Claude Haiku (enhance) → Paste at cursor

Every step runs in the background. The menu bar icon shows your current state: idle, recording, or processing.

Pipeline Detail

Step	Technology	What it does
Audio capture	sounddevice / PortAudio	Streams mic input at 16kHz mono
Voice activity detection	Silero VAD (PyTorch)	Strips silence; handles whisper-level audio
Gain normalization	RMS-based	Amplifies quiet audio so VAD works on whispers
Transcription	OpenAI Whisper API	Accurate STT across accents and background noise
Enhancement	Anthropic Claude Haiku	Polishes grammar, tone, and structure
Output	pynput + pbcopy/pbpaste	Saves clipboard → pastes → restores clipboard
Name formatting	spaCy NER (en_core_web_sm)	Detects person names, optionally prefixes with @

---

Enhancement Modes

Switch modes from the menu bar dropdown:

Mode	What it does
Cleanup	Fix grammar and remove filler words with minimal rewrites
Professional	Formal rewrite for business communication, emails, and docs
Casual	Friendly, conversational tone for chats and informal writing
Bullets	Convert your speech into organized bullet points

---

Requirements

• macOS 12+ (uses native menu bar, Keychain, clipboard APIs)
• Python 3.10+
• API Keys: OpenAI · Anthropic

---

macOS Permissions

Grant these in System Settings → Privacy & Security on first run:

Permission	Why
Microphone	Audio recording (auto-prompted on first use)
Accessibility	Simulating Cmd+V to paste text
Input Monitoring	Detecting the global hotkey from any app

---

Configuration

Config file: ~/.vaani/config.yaml

hotkey: "alt"                   # Global hotkey to hold while speaking
active_mode: professional       # Default enhancement mode
sounds_enabled: true            # Audio feedback on start/stop
vad_threshold: 0.05             # Lower = more sensitive (good for whispers)
sample_rate: 16000              # Audio sample rate (Hz)
max_recording_seconds: 600      # Auto-stop after 10 minutes
stt_model: whisper-1            # OpenAI transcription model
llm_model: claude-haiku-4-5-20251001  # Anthropic enhancement model
microphone_device: null         # null = system default, or set device index
paste_restore_delay_ms: 100     # How long to wait before restoring clipboard
launch_at_login: false          # Start Vaani automatically on login

Configuration reloads automatically when the file changes — no restart needed.

---

Custom Prompts

Override any prompt by creating files in ~/.vaani/prompts/:

~/.vaani/prompts/
├── system.txt          # Override the base system prompt
├── context.txt         # Add personal context (your writing style, name, role)
└── modes/
    ├── cleanup.txt     # Override the cleanup mode prompt
    ├── professional.txt
    ├── casual.txt
    └── bullets.txt

User prompts take priority over built-in defaults. Use context.txt to tell Vaani about you — your name, your company, common terms you use — so the output matches your voice.

---

Privacy

Data	Where it goes	How it's stored
Audio	Sent to OpenAI for transcription	Not stored by Vaani
Transcribed text	Sent to Anthropic for enhancement	Not stored by Vaani
API keys	macOS Keychain only	Never written to disk in plaintext
Transcription history	Local SQLite database	AES-256 encrypted (Fernet)

No data is retained on Vaani's servers because there are no Vaani servers. All cloud calls go directly from your machine to OpenAI and Anthropic under your API account, subject to their data retention policies.

---

Architecture

┌─────────────────────────────────────────────────────────────┐
│  Main Thread (macOS requirement)                            │
│  ┌─────────────┐   ┌───────────────────────────────────┐   │
│  │ HotkeyListener│  │ VaaniMenuBar (rumps event loop)   │   │
│  │ (pynput)    │   │ Status icon · Mode selector       │   │
│  └──────┬──────┘   └──────────────────────────────────-┘   │
│         │ on_press / on_release                              │
└─────────┼───────────────────────────────────────────────────┘
          │
          ▼
┌─────────────────────────────────────────────────────────────┐
│  StateMachine: IDLE → RECORDING → PROCESSING → IDLE        │
└─────────────────────────────────────────────────────────────┘
          │
          ▼  (daemon threads)
┌──────────────────────────────────────────────────────────┐
│  AudioRecorder → process_audio() → transcribe() → enhance() → paste_text()  │
│  sounddevice     Silero VAD        Whisper API    Claude     pynput           │
│                  + gain norm       + WAV encode   Haiku      + pbcopy         │
└──────────────────────────────────────────────────────────┘
          │
          ▼
┌─────────────────────────────┐
│  HistoryStore (SQLite)      │
│  Fernet-encrypted records   │
└─────────────────────────────┘

---

Development

git clone https://github.com/ankushbhardwxj/vaani
cd vaani
python -m venv .venv && source .venv/bin/activate
pip install -e ".[test]"
python -m spacy download en_core_web_sm

# Run tests
pytest

# Run in foreground (with live logs)
vaani start --foreground

Running Tests

pytest                  # Run all tests
pytest -v               # Verbose output
pytest tests/test_audio.py  # Single module

---

License

MIT — see LICENSE for details.