Audio version — Estimated duration: 9 min 14 sec

Configuration

All settings in Whisper Dictate are variables at the top of the script (lines 28-44 in whisper-dictate-v2.py). You can change any of them with a text editor before running the script.

Config Location

Open whisper-dictate-v2.py and look for the section starting with:

# ── Config ────────────────────────────────────────────────────────────────────
HOTKEY_CODE  = "KEY_RIGHTALT"
MODEL_SIZE   = "medium"
...
MIN_DURATION = 0.3
# ─────────────────────────────────────────────────────────────────────────────

Everything above the separator line is configurable.

Quick Reference Table

Variable	Default	What It Controls
HOTKEY_CODE	"KEY_RIGHTALT"	Which keyboard key triggers recording
MODEL_SIZE	"medium"	Whisper model size (accuracy vs speed)
SAMPLE_RATE	16000	Audio capture rate (don’t change unless you know what you’re doing)
DEVICE	"cuda"	GPU ("cuda") or CPU ("cpu") for transcription
COMPUTE_TYPE	"int8_float16"	How model weights are stored (memory usage vs speed)
LANGUAGE	None	Spoken language (None = auto-detect)
SHOW_NOTIFY	True	Desktop pop-ups on/off
MIN_DURATION	0.3	Minimum recording time in seconds (filters accidental taps)

---

HOTKEY_CODE — Choose Your Hotkey

Default: "KEY_RIGHTALT"

This is the key you hold to dictate. Any key on your keyboard can be used.

Common Hotkeys

Code	Physical Key	Notes
KEY_RIGHTALT	Right Alt	Default — convenient for thumb
KEY_RIGHTCTRL	Right Ctrl	Also easy for thumb
KEY_CAPSLOCK	Caps Lock	You won’t accidentally hit it
KEY_F9	F9	Function row — unlikely to conflict
KEY_RIGHTMETA	Right Super/Win	Usually opens app menu — beware conflicts
KEY_LEFTMETA	Left Super/Win	Usually opens app menu — beware conflicts
KEY_COMPOSE	Compose/Menu key	Often unused
KEY_LEFTCTRL	Left Ctrl	May conflict with Ctrl+C, Ctrl+V, etc.

How to Find Your Key Code

python3 -m evdev.evtest

This opens an interactive key tester. Press the key you want to use and look for the KEY_* name printed to the terminal. For example, pressing the Right Alt key shows:

Event: time 12345.678901, type 1 (EV_KEY), code 100 (KEY_RIGHTALT), value 1

The code name (KEY_RIGHTALT) is what you put in your config.

Best Practices

• Avoid modifiers used by applications — Left Ctrl conflicts with copy/paste shortcuts, Super/Windows key opens menus
• Use a key you can press with your non-dominant hand — You’ll be holding it for several seconds while speaking
• Avoid keys you type frequently — You don’t want to accidentally start recording while typing normally

---

MODEL_SIZE — Accuracy vs Speed

Default: "medium"

Whisper comes in several sizes. Larger models are more accurate but slower and use more VRAM (video memory).

Model Comparison

Size	Parameters	Speed (RTX 3050)	VRAM	Accuracy	Best For
tiny	39 million	~1 second	~1 GB	Lowest	Quick commands, low-end GPUs
base	74 million	~2 seconds	~1 GB	Medium	Balanced on limited hardware
small	244 million	~3 seconds	~2 GB	Good	Everyday use on 4GB GPUs
medium	769 million	~5 seconds	~5 GB	Great	Default — best for 6GB GPUs
large-v3	1.55 billion	~10 seconds	~10 GB	Best	Maximum accuracy, high-end GPUs

How to Choose

1. Check your GPU memory: Run nvidia-smi and look at the “Memory-Usage” for your GPU.
2. Pick the largest model that fits: If you have 6GB VRAM like an RTX 3050, medium is the sweet spot. large-v3 requires 10GB+.
3. CPU users: Use tiny or base — transcription takes 5-20x longer on CPU.

Storage

Models are cached in the ./models/ directory in your project folder. Space used:

Model	Disk Space
tiny	~300 MB
base	~600 MB
small	~2 GB
medium	~6 GB
large-v3	~12 GB

---

DEVICE — GPU or CPU

Default: "cuda"

Controls whether transcription runs on your NVIDIA GPU or on your CPU.

Value	Hardware	Speed
"cuda"	NVIDIA GPU (requires CUDA toolkit)	Fast
"cpu"	System CPU	5-20x slower

GPU Notes

• You need an NVIDIA GPU with at least 4GB VRAM for small or larger models
• CUDA toolkit must be installed: nvidia-smi should show driver version
• If you get “CUDA out of memory” errors, try a smaller model or switch to CPU

CPU Notes

• Works on any system, even without a GPU
• medium model takes ~1-2 minutes on a modern CPU
• tiny model takes ~5-10 seconds on CPU
• The script will still work for dictation, just with a delay between releasing the key and text appearing

---

COMPUTE_TYPE — Memory Quantization

Default: "int8_float16"

This controls how model weights are stored in GPU memory. Different formats trade off VRAM usage against speed and accuracy.

Type	VRAM Saved	Speed	Notes
"int8"	Most	Fast	Slight accuracy loss
"int8_float16"	Medium	Fast	Default — good balance, safe for RTX 3050
"float16"	Some	Fastest	May cause NaN/infinity errors on some GPUs
"float32"	None	Slowest	Full precision, rarely needed outside research

If you see NaN errors or incorrect output: Switch to "int8_float16" or "int8". Some GPUs have issues with full float16.

If you have limited VRAM and need a larger model: Use "int8" — it reduces model size the most.

---

LANGUAGE — Auto-Detect or Fixed

Default: None

None = automatic language detection. Whisper listens to the first 30 seconds of audio and determines the language on its own. This adds ~0.5 seconds to transcription.

Set a specific language code to skip detection and improve accuracy:

Code	Language
"en"	English
"hi"	Hindi
"mr"	Marathi
"es"	Spanish
"fr"	French
"de"	German
"zh"	Chinese
"ja"	Japanese
"ko"	Korean

When to set a language:

• You always speak the same language
• Auto-detection adds noticeable delay
• Whisper confuses your language with a similar one (e.g., Ukrainian with Russian)

Full list: https://github.com/openai/whisper/blob/main/whisper/tokenizer.py

---

SHOW_NOTIFY — Desktop Notifications

Default: True

Controls whether desktop pop-ups appear for recording status.

Value	Behavior
True	Show notifications (requires libnotify-bin)
False	No notifications, terminal output only

Notification events:

• Model loading (startup)
• “Ready” (model loaded)
• “Recording…” (hotkey pressed)
• “Transcribing…” (hotkey released)
• “Typed” (text injected)
• “Copied to clipboard” (fallback used)
• “Nothing detected” (too short or no speech)

---

MIN_DURATION — Accidental Tap Filter

Default: 0.3

The minimum recording duration in seconds. If you press and release the hotkey in less than this time, nothing is transcribed.

Why this exists: You might accidentally brush the hotkey while typing normally. Without this filter, every accidental tap would trigger a transcription of silence or background noise.

Recommendations:

• 0.3 — Good default. Filters accidental taps, doesn’t interfere with intentional use
• 0.5 — More aggressive filtering, requires a deliberate press
• 0.1 — Minimal filtering, captures even brief utterances

---

Internal Variables (Not User-Configurable)

These are set inside the code and affect behavior:

Variable	Default	Purpose
_stream_interval	2.0	Seconds between live preview transcriptions during recording
_stream_idx	0	Tracks which audio chunks have been previewed; resets each recording
_lock	threading.Lock()	Thread safety lock — prevents race conditions

You can change _stream_interval if you want more or less frequent live previews. A shorter interval (e.g., 1.0) gives more real-time feedback but uses more GPU time. A longer interval (e.g., 5.0) conserves GPU resources but shows less feedback.

---

Best Practices Summary

Goal	Settings
Fastest transcription	MODEL_SIZE = "tiny", COMPUTE_TYPE = "int8_float16", LANGUAGE = "en"
Highest accuracy	MODEL_SIZE = "large-v3", COMPUTE_TYPE = "float16", LANGUAGE = "en"
Low VRAM usage	MODEL_SIZE = "tiny", COMPUTE_TYPE = "int8"
CPU-only system	DEVICE = "cpu", MODEL_SIZE = "tiny"
Quiet operation	SHOW_NOTIFY = False

Now see common-issues if you run into issues, or components to understand how each config option is used internally.