SuperScreecher9000

Screech at your computer. It listens.

A local, GPU-accelerated speech-to-text tool that types wherever your cursor is. Built for accessibility, privacy, and freedom.

What It Does

Press a key, speak, release. Your words appear at cursor. Or say a command. No cloud. No API keys. No subscriptions. Just you and your computer.

Features

• Local inference - Whisper runs on YOUR machine
• GPU accelerated - Vulkan (Intel/AMD), CUDA (NVIDIA), Metal (macOS)
• Types at cursor - Works anywhere: editors, browsers, chat apps
• Voice commands - Navigation, editing, media controls, custom shell commands
• Leader words - No reserved words; say "command enter" vs just "enter"
• Punctuation - 50+ symbols via voice: "punctuation arrow" → =>
• Spell mode - NATO phonetic, letters, numbers, punctuation: "command spell alpha at bravo dot com" → a@b.com
• Shift mode - Text selection: "command shift right times five" selects 5 characters
• Hold/Release - Hold keys for gaming/accessibility: "command hold w" runs forward
• Emoji - 80+ emoji via voice: "emoji thumbs up" → 👍
• Case modes - snake_case, camelCase, PascalCase, SCREAMING_SNAKE, and more
• Math mode - Spoken math to symbols: "one plus one" → 1 + 1
• Inserts - Voice-triggered text snippets with placeholders: {date}, {shell:cmd}
• Wrappers - Wrap text by voice: "wrap quotes hello" → "hello"
• Repetition - "command backspace times five" or "command repeat three"
• Mishearing tolerance - Built-in handling for common Whisper errors (caret/carrot, colon/colin, etc.)
• Fuzzy matching - Custom commands match despite spacing/number variations
• Self-documenting - "command help" shows all commands, "command config" opens config
• Hot-reload config - Change settings without restarting
• Quiet mode - Suppress verbose output once you’re comfortable
• Multiple models - tiny (75MB) to large (3GB), pick your speed/accuracy tradeoff
• Cross-platform ready - Built with portable Rust crates

Installation

# Clone and build (Linux)
git clone https://github.com/sqrew/ss9k
cd ss9k
cargo build --release

# Run it - model downloads automatically on first launch
./target/release/ss9k

Windows Installation

Windows requires additional dependencies:

Visual Studio Build Tools with "Desktop development with C++" workload

• Download from Visual Studio Downloads
• Select "Build Tools for Visual Studio"
• In the installer, check "Desktop development with C++"

LLVM/Clang (needed for whisper-rs bindings)

• winget install LLVM.LLVM or download from LLVM Releases
• Make sure to add to PATH during installation

Rust via rustup 1.

Git from git-scm.com

Then build from "x64 Native Tools Command Prompt for VS":

git clone https://github.com/sqrew/ss9k
cd ss9k
cargo build --release

Yeah, it’s a lot. Linux is easier. But it works!

Usage

# Hold your hotkey (F12 default), speak, release
# Text appears at cursor

Voice Commands

SS9K uses a leader word (default: command) to distinguish commands from dictation:

• "command enter" → presses Enter key
• "enter" → types the word "enter"
• "command punctuation period" → types . (or "command punk period")
• "command spell alpha at bravo" → types a@b
• "command emoji fire" → types 🔥

Everything goes through the leader word. Configure it in your config:

leader = "voice"  # or "computer", "hey", whatever feels natural

Commands (say "command" + any of these):

Category	Commands
Navigation	enter, tab, escape, backspace, space, up, down, left, right, home, end, page up/down
Editing	select all, copy, paste, cut, undo, redo, save, find, close tab, new tab
Media	play, pause, next, skip, previous, volume up, volume down, mute
Utility	help (show commands), config (open config), repeat, repeat [N]

Punctuation (say "command punctuation" + any of these, or "command punk"):

Category	Options
Basic	period, comma, question, exclamation, colon, semicolon
Quotes	quote, single quote, backtick
Brackets	open/close paren, bracket, brace, angle
Symbols	plus, minus, equals, asterisk, slash, pipe, at, hash, etc.
Programming	arrow (=>), thin arrow (->), double colon, equals equals, not equals, and and, or or

Spell Mode (say "command spell" + letters/numbers/punctuation):

Input	Output
command spell alpha bravo charlie	abc
command spell capital alpha bravo	Ab
command spell one two three	123
command spell alpha at bravo dot com	a@b.com
command spell alpha space bravo	a b
command spell alpha underscore bravo	a_b

Supports: NATO phonetic (alpha-zulu), number words (zero-nine), raw letters, raw digits, space, and punctuation (dot, at, dash, underscore, slash, colon, hash, etc.). Capital modifiers: capital, cap, uppercase, upper.

Shift Mode (say "command shift" + direction for text selection):

Input	Effect
command shift right	Select 1 character right
command shift right times five	Select 5 characters right
command shift word left	Select word left
command shift home	Select to start of line
command shift end	Select to end of line
command shift page down	Select a page down

Supports: left, right, up, down, word left, word right, home, end, page up, page down, tab, enter.

Hold/Release (for gaming, accessibility, or modifier keys):

Input	Effect
command hold w	Hold W key (run forward)
command hold shift	Hold Shift modifier
command release w	Release W key
command release all	Release all held keys

Supports: all letters (a-z), modifiers (shift, control/ctrl, alt, meta/super/win), arrows (up, down, left, right), and common keys (space, enter, tab, escape, backspace).

How it works: Hold mode rapidly presses the key (configurable via key_repeat_ms). All held keys press together, so "hold shift" + "hold w" works for sprint+move.

Tip: Use hold for games ("command hold w" to run), accessibility, or any situation where you need a key pressed continuously.

Emoji (say "command emoji" + name):

Input	Output
command emoji smile	😊
command emoji thumbs up	👍
command emoji fire	🔥
command emoji blue heart	💙
command emoji crab	🦀
command emoji poop	💩

80+ emoji available: faces, gestures, hearts (all colors), animals, objects, symbols. Say "command emoji rust" for 🦀.

Case Modes (say "command mode" + mode name):

Mode	Effect	Example Output
snake	snake_case	hello_world
camel	camelCase	helloWorld
pascal	PascalCase	HelloWorld
kebab	kebab-case	hello-world
screaming	SCREAMING_SNAKE	HELLO_WORLD
caps	ALL CAPS	HELLO WORLD
lower	lowercase	hello world
math	spoken math → symbols	one plus one → 1 + 1
off	normal (default)	hello world

Mode persists until changed. Say "command mode snake", then dictate naturally—all text becomes snake_case. Say "command mode off" to return to normal.

Tip: Great for coding—"mode snake" for Python, "mode camel" for JavaScript, "mode pascal" for type names.

Math Mode converts spoken math to symbols:

Input	Output
one plus one	1 + 1
five times three	5 * 3
x greater than y	x > y
open paren a plus b close paren	( a + b )
three point one four	3 . 1 4

Supports: numbers 0-20, operators (+, -, *, /, =, %, ^), comparisons (>, <, >=, <=, !=, ==), parentheses/brackets/braces, decimals, and common homophones (to→2, for→4).

Inserts (say "command insert" + name):

Define text snippets in your config and insert them by voice:

[inserts]
email = "you@example.com"
sig = "Best regards,\nYour Name"
header = "// Created: {date}\n// Author: Your Name"
branch = "{shell:git branch --show-current}"

Input	Output
command insert email	you@example.com
command insert header	// Created: 2026-01-17\n// Author: Your Name
command insert branch	main (or current branch)

Placeholders:

• {date} → 2026-01-17
• {time} → 13:52
• {datetime} → 2026-01-17 13:52
• {timestamp} → Unix timestamp
• {iso} → ISO 8601 format
• {shell:command} → output of any shell command
• \n → newline, \t → tab

The {shell:...} placeholder is powerful—pull in git info, environment variables, clipboard contents, API responses, anything shell can do.

Wrappers (say "command wrap" + name + text):

Define text wrappers in your config and wrap dictated text:

[wrappers]
quotes = '"'
parens = "(|)"
fire = "🔥"
div = "<div>|</div>"
bold = "**|**"

Input	Output
command wrap quotes hello world	"hello world"
command wrap parens check this	(check this)
command wrap fire awesome	🔥awesome🔥
command wrap div content here	<div>content here</div>

If the wrapper value contains |, it splits into left/right. Otherwise, the value is used on both sides.

Repetition (add "times N" to any command, or use "repeat"):

Input	Effect
command backspace times five	Delete 5 characters
command down times ten	Move down 10 lines
command repeat	Repeat last command once
command repeat three	Repeat last command 3 times

Works with number words (one-twenty) or digits. Handles common mishearings like "to"→2, "for"→4.

Mishearing tolerance: SS9K handles common Whisper transcription errors automatically:

• caret → also matches "carrot", "karet"
• colon → also matches "colin", "cologne"
• asterisk → also matches "asterix", "astrix"
• tilde → also matches "tilda", "squiggle"
• And many more built-in.

Custom commands (from config) work without a leader word.

Tip: Use aliases to shorten the leader: "cmd" = "command" → say "cmd enter"

Configuration

Create ~/.config/ss9k/config.toml:

model = "small"              # tiny, base, small, medium, large
language = "en"              # ISO 639-1 code
threads = 4                  # whisper inference threads
device = ""                  # audio device (empty = auto-detect)
hotkey = "F12"               # see supported hotkeys below
hotkey_mode = "hold"         # hold (release to stop) or toggle (press again to stop)
toggle_timeout_secs = 0      # auto-stop after N seconds in toggle mode (0 = no timeout)
leader = "command"           # leader word for commands (or "voice", "computer", etc.)
key_repeat_ms = 50           # key repeat rate for hold mode (ms between presses)
quiet = false                # suppress verbose output (set true once comfortable)

[commands]
"open terminal" = "kitty"
"open browser" = "$BROWSER"  # supports $ENV_VAR expansion
"screenshot" = "flameshot gui"
"workspace one" = "i3-msg 'workspace 1'"  # fuzzy matches "work space 1", "Workspace One", etc.

[aliases]
"taping" = "typing"          # fix consistent misrecognitions
"come and" = "command"       # common Whisper mishearing

[inserts]
email = "you@example.com"
sig = "Best regards,\nYour Name"
header = "// Created: {date}\n// Author: {shell:git config user.name}"
branch = "{shell:git branch --show-current}"

[wrappers]
quotes = '"'
parens = "(|)"
brackets = "[|]"
fire = "🔥"
div = "<div>|</div>"

Supported hotkeys: F1-F12, ScrollLock, Pause, PrintScreen, Insert, Home, End, PageUp, PageDown, Num0-Num9

Config hot-reloads when you save - no restart needed.

Models

Model	Size	Speed	Accuracy	Use Case
tiny	75MB	Fastest	Basic	Quick notes, low-end hardware
base	142MB	Fast	Good	General use
small	466MB	Medium	Great	Recommended default
medium	1.5GB	Slow	Excellent	When accuracy matters
large-v3	3GB	Slowest	Best	Maximum quality

Models auto-download on first launch. Change model in config to switch.

Hardware

Minimum:

• Any x86_64 CPU
• 2GB RAM (+ model size)
• Microphone

Recommended:

• Modern CPU (last 5 years)
• GPU with Vulkan support
• 8GB RAM
• Decent microphone

Tested on:

• Intel i5-6500 + HD Graphics 530 (works, ~15s inference on medium)
• Your machine (probably faster)

GPU Backends

Build with the appropriate feature:

cargo build --release --features vulkan  # Intel/AMD (Linux/Windows)
cargo build --release --features cuda    # NVIDIA (requires CUDA toolkit)
cargo build --release --features metal   # macOS

Known Issues

• Wayland: Global hotkeys don’t work (Wayland security model). Use X11.

Comparison with Alternatives

There are several voice control tools out there. Here’s how SS9K compares:

Feature	SS9K	Talon	Dragon	Dictara	Voxtype	Nerd Dictation
Platforms	Linux, Windows	All	Windows, Mac	All	Linux	Linux
Price	Free	Freemium	$200-500+	Free	Free	Free
Speech Engine	Whisper.cpp	Own/Whisper	Proprietary	Whisper	Whisper.cpp	VOSK
Fully Local	✅	✅	✅	⚠️ Optional	✅	✅
Voice Commands	✅ Full	✅ Scripting	⚠️ Basic	❌	⚠️ Punctuation	❌
Case Modes	✅ 8 modes	✅	❌	❌	❌	❌
Math Mode	✅	❌	❌	❌	❌	❌
Insert Snippets	✅ + {shell:}	✅ Python	❌	❌	❌	❌
Text Wrappers	✅	✅	❌	❌	❌	❌
Custom Commands	✅ TOML	✅ .talon	⚠️	❌	❌	❌
GPU Acceleration	✅	✅	❌	✅	✅	❌
Learning Curve	Low	High	Medium	Low	Low	Low
Setup Complexity	Low	High	Medium	Low	Medium	Low

TL;DR:

• Want power-user features with low learning curve? → SS9K (you’re here)
• Want maximum power and don’t mind learning? → Talon
• Need professional/medical dictation on Windows? → Dragon
• Want minimal Linux-only dictation? → Voxtype or Nerd Dictation
• Want mac dictation with a GUI? → Dictara

SS9K sits in the sweet spot: Talon-level features without the scripting complexity. Everything is configured in a simple TOML file, yet you get case modes, math mode, shell integration, custom commands, and more.

Why?

For people who:

• Have RSI or carpal tunnel
• Have motor disabilities
• Think better out loud
• Want local/private speech-to-text
• Are tired of cloud subscriptions
• Rage-smashed your keyboard into pieces

5GB for a fully offline accessibility tool is nothing. Your voice stays on your machine. Just remove unused models to save space if it’s important.

License

MIT. Do whatever you want.

Credits

Built with:

• whisper-rs - Whisper.cpp Rust bindings
• cpal - Cross-platform audio
• rdev - Global hotkey capture
• enigo - Keyboard simulation
• rubato - Audio resampling
• chrono - Date/time for insert placeholders
• arc-swap - Lock-free config hot-reload
• notify - File watching

Built by sqrew + Claude. The screech is real. 🦀