v0.1.2 Download

Parakeet TDT v3 Engine & Offline HUD skins

Your voice, written.

Dwani is a premium local push-to-talk dictation utility for macOS. Hold a key, speak your thoughts, and release. The transcript instantly streams, formatting itself perfectly into whatever application you are writing in.

Mac only — open this page on your Mac

Download Free for macOS Email link to my Mac Try Interactive Demo

brew install --cask anilraok/anvl/dwani

Requires macOS 14+ No account required 100% On-Device & Private

Works natively inside your active window

VS Code Cursor Slack Terminal Notion Obsidian Chrome

Experience Dwani in action

Select a mockup, choose your HUD overlay skin, and click-and-hold the trigger button to dictate.

1. Focused App

2. HUD Overlay Skin

VS Code - users.ts

1 import { db } from './database';

2 import { User, serializeUser } from './models';

3 // Press and hold Space to dictate code below:

Anil Rao Hey team, did you test the latest Dwani dictation engine? It's incredibly fast.

Jack Yeah, it matches punctuation perfectly. Typo rates went way down!

Dwani Bot APP Press & Hold to inject your message.

Last login: Wed May 20 15:11:27 on ttys003

anilraok@Anils-MacBook-Pro dwani % # Dictate a custom voice shortcut command

0.0s Listening...

Menu-Bar Active

Dwani Dictation

0.0s

Speak now...

Built for Speed. Designed for Privacy.

Every decision in Dwani centers around providing a zero-friction, native macOS experience.

Push-to-Talk Precision

No wake words, no delays. Hold Right ⌘ (or a custom hotkey), speak instantly, and let go. The transcription inserts exactly at your focused text cursor.

100% Local Processing

No audio files are ever sent to servers. Everything transcribes locally on your Apple Neural Engine, keeping your dictation confidential.

Spoken Keystrokes

Seamlessly weave commands. Say "press tab" or "select all" mid-sentence, and Dwani executes native keyboard shortcuts instead of writing text.

Custom Voice Actions

Bind spoken shortcuts directly to system workflows. Say "open Claude Code" to run AppleScripts, launch terminal commands, or load URLs.

Vocabulary Mapping

Train your speech. Map custom strings (e.g. heard words like "type script" or "react js") to print perfectly as "TypeScript" or "React" on output.

Per-App Formatting

Intelligent app-matching rules. Dwani formats casual capitalization for Slack, while preserving lowercase code syntax and tabs in VS Code.

Why Dwani is a complete upgrade

A direct comparison between built-in system dictation and Dwani's developer-focused voice engine.

Capability	Default macOS Dictation	Dwani (Parakeet/Nemotron)
Global Trigger	Fn double-tap (slow to start, lacks hold/release mapping)	Push-to-Talk (Hold Right ⌘-zero start latency)
Data Processing	Hybrid cloud servers (sends voice files over network)	100% offline local Neural Engine (private & offline)
App Customization	Writes identical text blocks in all applications	App-aware (formats lower-case code vs capitalized chat)
Spoken Shortcuts	Writes out words literally ("press enter")	Simulates keystroke actions ("press enter", "tab")
Workflow Hooks	None	Execute terminal commands & AppleScripts directly
Jargon Training	Manual Keyboard shortcuts list replacement rules	Phonetic custom vocabulary compilation sandbox

Uncompromised speed.
Unmatched accuracy.

Dwani ships NVIDIA's **Parakeet TDT v3** model natively integrated via FluidAudio, with support for the cutting-edge **Nemotron ASR** engine currently in parallel development. Both operate directly on the Apple Silicon Neural Engine (ANE) to provide unparalleled response times and record-low Word Error Rates (WER).

3,333×

Real-Time Factor (RTFx)

1.9%

Best Word Error Rate (WER)

Engine Performance Benchmark Arm64 Apple ANE

Parakeet TDT v3 (FluidAudio) 3,333× RTFx

Nemotron ASR (On-Device, Parallel) 2,800× RTFx

Apple Speech Engine (On-Device) 1,400× RTFx

WhisperKit large-v3 turbo 146× RTFx

*Real-Time Factor (RTFx) represents processing throughput: higher is faster. Parakeet TDT v3 and Nemotron ASR complete transcription in microseconds, using almost zero CPU cycles by utilizing the Apple Silicon Neural Engine.

Adaptive Post-Processing

Configure how Dwani reformats shorthand text or technical jargon instantly.

Heard Speech (Sounds like) Written Transcript (Pasted)

*Edit heard and written columns to see how the vocabulary compiler formats transcripts in the right sandbox.

Speech Input Simulator

Formatted Output Preview

Compiled transcript will print here...

Frequently Asked Questions

Everything you need to know about setting up and running Dwani on your Mac.

How does the global Hotkey work in macOS?

Dwani runs as a background process monitor using macOS `CGEventTap` to monitor global key releases. The default binding is **Right ⌘**, allowing push-to-talk convenience without blocking standard left ⌘ keyboard shortcuts. You can easily configure custom layouts under Settings → Hotkeys.

Does Dwani upload my voice or text data anywhere?

No. Dwani is committed to privacy. All recording buffers are transcribed locally on your computer via the Apple Silicon Neural Engine (or local CPU APIs for Intel). The application has no tracking telemetry and functions completely offline.

What is the difference between Parakeet and Apple Speech engines?

Apple Speech is the built-in system recognizer, which requires no extra files and streams partials instantly. Parakeet TDT v3 is a state-of-the-art ~600MB CoreML model packaged via FluidAudio. It provides lower Word Error Rates, processes multilingual audio, and performs significantly better on code blocks or industry-specific vocabulary.

Why does Dwani request Accessibility and Input Monitoring permissions?

Accessibility API permissions allow Dwani to discover which application is active, read the focused text boundaries, and programmatically inject text when you release the hotkey. Input Monitoring permits the app to monitor keydown and keyup states globally.

Get Dwani today.

Experience dictation at the speed of thought. 100% offline, private, and highly customizable for developer workflows.