flow-st8

Local voice-to-text for Windows. Press Ctrl+Win, talk, press again — your words are transcribed by Whisper and pasted wherever your cursor is. No cloud, no subscription, no audio leaving your machine.

No cloud. No subscription. No screenshots sent to anyone's server.

---

Requirements

• Windows 10/11
• Python 3.11+
• Microphone
• NVIDIA GPU recommended (RTX 20xx/30xx/40xx) — also works on CPU, slower

---

Install

git clone https://github.com/luckmattos/flow-st8.git
cd flow-st8

python install.py

install.py detects your hardware and installs the right PyTorch version automatically:

• NVIDIA GPU found → installs CUDA build (~2.7GB, transcription <1s)
• No GPU → installs CPU build (lighter, slower transcription)

On first run, Whisper downloads the model (~1.5GB, one time only).

---

Run

Double-click flow-st8.vbs — no terminal, no console window.

Or from terminal:

python main.py

A mic icon appears in the system tray. You're ready.

---

How to use

1. Press Ctrl+Win — icon turns red, recording starts
2. Talk — pause freely, silence is filtered out automatically
3. Press Ctrl+Win again — icon turns blue, transcribing
4. Text is pasted wherever your cursor is

Tray menu: right-click the icon to record, toggle autostart, or quit.

---

Choosing a model

Edit %APPDATA%\flow-st8\config.toml and change name under [model].

Model	Size	With NVIDIA GPU	CPU only	Quality
tiny	39 MB	~0.1s	1-2s	Low, hallucinates
base	138 MB	~0.3s	3-5s	Decent
small	460 MB	~0.7s	8-12s	Good — best for CPU
medium	1.5 GB	~1.5s	25-35s	Very good
large-v3-turbo	1.5 GB	~0.4-1.2s ⭐	20-45s	Excellent (default)

No GPU? Stick with base or small. Running large-v3-turbo on CPU means waiting 20-45 seconds per sentence — not practical.

Benchmarks on RTX 4060 Laptop 8GB with large-v3-turbo:

Speech duration	GPU latency	CPU latency
5s	~0.4s	~8s
15s	~1.2s	~25s
30s	~2.5s	~45s

---

Configuration

Config file: %APPDATA%\flow-st8\config.toml (auto-created on first run)

[model]
name = "large-v3-turbo"   # tiny | base | small | medium | large-v3-turbo
language = "pt"
initial_prompt = "Transcription in Brazilian Portuguese."

[hotkey]
mode = "toggle"           # toggle = press to start, press to stop
key = "ctrl+win"

[audio]
device_index = -1         # -1 = system default mic
max_seconds = 210         # 3m30s max per recording
gain = 1.8

[vad]
enabled = true
speech_threshold = 0.5

[startup]
autostart = true          # start with Windows

---

Privacy

Commercial voice dictation apps send your audio — and often screenshots of your screen — to their servers every time you speak.

flow-st8 does not:

• Send audio anywhere
• Capture screenshots
• Require internet
• Store anything outside your machine

Everything runs locally. Audio is processed and discarded.

---

Stack

Layer	Technology
Speech-to-text	OpenAI Whisper + PyTorch
Voice detection	Silero VAD v6
Global hotkey	Win32 WH_KEYBOARD_LL hook via ctypes
Audio capture	sounddevice
Text injection	pyperclip + Win32 SendInput
Tray icon	pystray + Pillow
Config	TOML via stdlib tomllib
Autostart	winreg → HKCU\...\Run

For architecture details see ARCHITECTURE.md.

---

Roadmap

• Packaged .exe installer
• Auto-updater
• Language auto-detect
• Transcription history
• Per-app custom prompts

---

Keywords: speech-to-text, voice-dictation, whisper, local, offline, privacy, windows, cuda, python, real-time, hotkey

---

Author

luckmattos - built with AI coding agents and tools. MIT license.