> ## Documentation Index
> Fetch the complete documentation index at: https://mumbli-feat-proof-of-use.mintlify.site/llms.txt
> Use this file to discover all available pages before exploring further.

# Voice to Text for Mac — Mumbli

> The fastest way to turn speech into text on your Mac. Hold Fn, speak, get clean text in any app. AI-powered transcription with custom vocabulary support.

# Voice to text for Mac

Mumbli turns your Mac into a voice-first machine. Hold **Fn**, speak naturally, and polished text appears right where your cursor is — in any application.

No dictation box. No copy-paste. No app-specific plugins. Just speak and type.

## Why Mumbli

**Apple Dictation** is built into macOS, but it has limits: no polishing, no custom vocabulary, no way to handle filler words or self-corrections. It works fine for short phrases but falls apart for longer dictation.

Mumbli fills the gap:

* **Filler word removal** — "um", "uh", and verbal tics are cleaned out automatically
* **Self-correction handling** — Say "at 4 pm, actually 3 pm" and get "at 3 pm"
* **Grammar and punctuation** — Light cleanup without rewriting your voice
* **Custom vocabulary** — Add names, jargon, and technical terms that speech engines get wrong
* **Multi-language** — Language detection is automatic, no switching needed

## How it works

<Steps>
  <Step title="Hold Fn to record">
    Press and hold the Fn key. A small overlay appears near your cursor showing that Mumbli is listening. Or double-tap Fn for hands-free mode.
  </Step>

  <Step title="Speak naturally">
    Talk at your normal pace. Audio is captured at 16 kHz mono and sent to an AI transcription service.
  </Step>

  <Step title="Get polished text">
    The raw transcription passes through a lightweight LLM that cleans it up while preserving your voice. Text is injected at the cursor via the Accessibility API.
  </Step>
</Steps>

## Speed matters

Mumbli offers three engine configurations:

| Engine       | Transcription     | Polish            | End-to-end     |
| ------------ | ----------------- | ----------------- | -------------- |
| **Standard** | ElevenLabs Scribe | GPT-5.4 Nano      | \~3-5 seconds  |
| **Fast**     | Groq Whisper      | Groq Llama 3.1 8B | \~0.5-1 second |
| **Deepgram** | Deepgram Nova-3   | GPT-5.4 Nano      | \~2-4 seconds  |

The Fast engine delivers sub-second transcription — faster than you can move your hands back to the keyboard.

## Custom vocabulary

Speech engines struggle with proper nouns, brand names, and technical terms. Mumbli lets you add a custom vocabulary list in Settings. These words are:

1. Sent as hints to the transcription engine, biasing it toward correct spellings
2. Injected into the polishing prompt so the LLM fixes any remaining errors

Real-world benchmark: accuracy jumped from **36% to 100%** across 11 vocabulary instances.

## Works everywhere

Mumbli uses the macOS Accessibility API to inject text at your cursor. This means it works in:

* Browsers (Chrome, Safari, Firefox, Arc)
* Email clients (Mail, Outlook, Gmail)
* Messaging apps (Slack, Discord, Messages, Telegram)
* IDEs (VS Code, Xcode, IntelliJ)
* Note-taking apps (Notes, Obsidian, Notion)
* Terminals and command lines
* Any app with a text field

No integrations to configure. No plugins to install.

## Permissions

Mumbli needs three macOS permissions to work:

| Permission           | Why                          |
| -------------------- | ---------------------------- |
| **Microphone**       | Audio capture for dictation  |
| **Accessibility**    | Injecting text at the cursor |
| **Input Monitoring** | Detecting Fn key presses     |

You'll be prompted on first launch. All processing uses your own API keys — nothing is stored on external servers.

<Card title="Download Mumbli" icon="download" href="https://github.com/fireharp/mumbli/releases/latest">
  Free and open source. macOS 13.0+ required.
</Card>