wispr-flow

Local AI dictation for macOS. Speak → text appears in any app. Fully private — nothing leaves your machine.

How it works

• Whisper (via MLX, Apple Silicon optimised) transcribes your speech
• By default, raw transcription is pasted instantly
• Say "delta, ..." to run the utterance through a local LLM for cleanup/rephrasing
• Push-to-talk: hold a key while speaking, release to paste

Requirements

• macOS (Apple Silicon recommended)
• Ollama running locally with a model pulled
• Python 3.11+

Setup

# 1. Clone
git clone https://github.com/Xaid-vfx/wispr-flow.git
cd wispr-flow

# 2. Create venv and install dependencies
python3 -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# 3. Pull a model for LLM rewriting (only needed for delta prefix)
ollama pull dolphin3

Run

# Push-to-talk (hold Right ⌘ while speaking, release to paste)
.venv/bin/python main.py --hotkey --hotkey-key right_cmd

# With a custom prompt to improve Whisper accuracy
.venv/bin/python main.py --hotkey --hotkey-key right_cmd --prompt "Transcribing technical notes."

# Continuous VAD mode (terminal output only)
.venv/bin/python main.py

# Use a more accurate Whisper model
.venv/bin/python main.py --hotkey --hotkey-key right_cmd --model large-v3

Grant Accessibility permission when prompted (required for auto-paste).

Usage

What you say	What happens
Anything	Pasted as raw Whisper output
delta, <text>	LLM rewrites <text> before pasting

Options

Flag	Default	Description
--hotkey	off	Push-to-talk mode
--hotkey-key	right_option	Key to hold (right_cmd, right_ctrl, etc.)
--model	medium.en	Whisper model (tiny.en → large-v3)
--llm-model	dolphin3	Ollama model for delta rewriting
--prompt	—	Initial prompt to bias Whisper
--debug	off	Show raw Whisper transcript
--threshold	0.015	VAD energy threshold

Tests

.venv/bin/python -m pytest tests/ -q