yusufipk
2cfbbb2d99
whisper-server is started on --inference-path /v1/audio/transcriptions, which is exactly the path api.py already builds for the hosted providers, and llama-server answers /chat/completions the way OpenRouter does. So the local half is one more base URL rather than a second code path: worker.py, filetranscribe.py and meeting.py are untouched, and dictation, subtitles and meetings all work here on the first try. Three findings worth naming, none of them in the new code: whisper.cpp cuts segments on tokens, which in Turkish lands inside a word about as often as between two. Pasted raw that gives "akraba değ\niller."; in a subtitle it gives a cue reading "değ". Whisper marks the start of a word with a leading space, so a piece that does not begin with one continues the word above it. A small model will repeat the transcript until the context is full, and every one of those tokens is a second of somebody waiting: measured at 206 seconds, and 25 with a ceiling on the reply. Hosted models are left alone, where the same runaway is rare and a ceiling would cut the minutes short. A server outlives SIGTERM and SIGKILL holding its model in memory. Signals are now turned into an event Qt delivers, since Qt blocks in C where a Python handler never runs, and a pid file lets the next start sweep up what a SIGKILL left behind. The minutes keep their own provider rather than following cleanup's. The two jobs are not the same size: a 4B model here will strip the filler words out of a dictation and will not write up an hour long meeting. The suite runs offline now: a test that reaches the network says so instead of quietly going there.
Dikte
Press Ctrl+Space, talk, press again. The recording is transcribed on this
machine by default, a model cleans it up (dropping the uhs, the restarts, the
missing punctuation), and the result lands in your clipboard and is pasted into
whatever window you were typing in.
Built for KDE Plasma 6 on Wayland. No dependencies beyond system packages: just the Python standard library and PyQt6.
 |
 |
 |
 |
Install
sudo pacman -S --needed pipewire-audio wl-clipboard ydotool ffmpeg python-pyqt6
systemctl --user enable --now ydotool # needed for auto-paste
./install.sh # or: ./install.sh "Ctrl+Alt+Space"
dikte # the settings window opens on first run
On Ubuntu/GNOME X11, recording uses PulseAudio and clipboard/paste use the X11 tools instead:
sudo apt install pulseaudio-utils xclip xdotool ffmpeg
install.sh adds the dikte command, a menu entry and an autostart entry. The
settings window installs a GNOME or KDE global shortcut.
Speech to text and cleanup each pick a provider in the settings window. Both can
run here, on whisper.cpp and llama.cpp: the program and the model are downloaded
from that window (checksummed, into ~/.local/share/dikte), so nothing has to be
installed first and nothing leaves the machine. The hosted alternatives want a
key: OpenAI or OpenRouter for speech to text, OpenRouter for cleanup
(google/gemini-3.5-flash-lite), so a single OpenRouter key can cover both. They fall back to OPENAI_API_KEY and OPENROUTER_API_KEY, and are
stored in ~/.config/dikte/config.json, mode 600. Cleanup can be switched off,
in which case the raw transcript is pasted, and a thinking model's effort can be
set next to it.
Using it
| What | How |
|---|---|
| Start / stop recording | Ctrl+Space, or click the tray icon |
| Cancel a recording | Tray menu → Cancel recording, or dikte cancel |
| Speak a command to an agent | Tray menu → Ask Claude, or dikte ask |
| Start / end a meeting | Tray menu → Record a meeting, or dikte meeting |
| Settings | Tray menu → Settings, or dikte settings |
| Reload after an update | Tray menu → Restart, or dikte restart |
| Quit | Tray menu → Quit, or dikte quit |
An indicator in the screen corner shows a red dot, a live waveform and the
elapsed time, then the stage it is on. It never takes focus. Pressing
Ctrl+Space again while Dikte is still working does nothing; nothing queues up.
A dictation and a command to the agent do wait on each other for the microphone,
which is one device, but for nothing else: each has its own indicator, and the
second one stacks above the first while both are up.
Everything the settings window holds has a verb of its own too, so a script or
an agent can work the whole thing: dikte record --seconds 8 says back what was
said, dikte transcribe talk.mp4 --srt writes subtitles, and the settings, the
history and the meetings are there beside them. dikte --help lists them, they
all take --json, and only the ones needing the microphone need the application
running.
What it does
-
Silence never reaches the API. Handed near-silence, a transcription model invents a sentence instead of returning nothing ("Thanks for watching", or in Turkish "Altyazı M.K."). A recording is dropped when nothing rose 10 dB above that recording's own noise floor for at least 0.3 s, which is also what removes steady fan noise however loud, or when its loud end sits below -55 dBFS. The indicator reports the level it measured, which is what you calibrate the threshold against.
-
Misheard words are repaired. Speech models fail phonetically on proper nouns, so the cleanup model is asked to fix those from context, and to leave the word alone when the context does not make the intended one clear. The names you list under Cleanup rules go to the transcription model as a hint and to the cleanup model as a glossary, which is what lets it recognise "kuber netis":
raw ıı bugün şey kuber netis üzerinde çalışan servisleri güncelledim yani sonra grafanada bir panel açtım hani ve pay kut ile arayüzü şey bitirdim işte result Bugün Kubernetes üzerinde çalışan servisleri güncelledim. Sonra Grafana'da bir panel açtım ve PyQt ile arayüzü bitirdim. -
A failed cleanup is never silent. The raw transcript is still pasted so the dictation is not lost, but the indicator turns amber with the reason instead of looking like a normal run.
-
A dictation can be a command instead. Its own shortcut sends the transcript to Claude Code (
claude -p) rather than pasting it, and pastes back what comes of it: the answer, or a sentence saying what was done. It is the session you would have opened yourself, so your skills and connected services are there, which is what makes "put that in my calendar on Thursday at three" a thing you can say to a window that is not Claude. Codex (codex exec) runs the same way, and OpenRouter is there as a plain question-and-answer fallback for a machine with neither CLI on it. Provider, model, permissions and working directory are under Settings → Agent, and commands close together stay in one conversation. -
Meetings are recorded from the microphone and the speaker output at the same time, which settles who said what by the channel a voice arrived on instead of guessing at it. The two sides are transcribed separately and interleaved into one timestamped transcript, and a second model, configured under Settings → Meeting along with its own instruction, turns that into minutes: decisions, action items, open questions. They land in
~/.local/share/dikte/meetingsand in Settings → Minutes. A run that fails keeps its recording, and a retry resumes from the transcript it already paid for. -
Audio and video files run through the same models under Settings → Audio file, optionally with
[mm:ss]timestamps, chunked through ffmpeg when long, and saved as.txtor as.srtsubtitles; their cleanup follows its own rules, written for subtitles, so the lines keep their place and nothing is shortened. -
History of every dictation under Settings → History, with a size limit and right-click to delete.
-
Turkish and English interface, following the system locale by default.
The global shortcut needs one logout
KWin only reads kglobalshortcutsrc at startup, so the shortcut install.sh
writes will not fire until you log out and back in. Until then, Settings →
Shortcut → built-in listener reads /dev/input and catches the combination
itself. The difference: it does not swallow the key, so Ctrl+Space also reaches
the focused application (some editors will pop up autocomplete). The listener
needs your user in the input group: sudo usermod -aG input $USER.
Layout
dikte.py entry point, tray icon, state machine
cli.py the command line: every verb, and what it answers with
ipc.py one request and one reply over the local socket
audio.py PCM capture: pw-record for dictation, ffmpeg for a meeting
meeting.py channel split, speaker labelling, cleanup, minutes
assistant.py running a dictation through Claude Code, Codex or OpenRouter
api.py transcription and cleanup on any provider (stdlib only)
ggml.py whisper.cpp and llama.cpp here: fetch, verify, keep serving
hub.py what GitHub and Hugging Face have on offer today
worker.py transcribe → clean up → clipboard → paste
vad.py deciding whether a recording holds speech at all
filetranscribe.py file transcription: ffmpeg, chunking, timestamps
overlay.py the corner indicator
settings_ui.py settings window
hotkey.py KDE shortcut installation and the evdev listener
paste.py wl-clipboard and ydotool wrappers
i18n.py the string table
The indicator is drawn through XWayland, because a Wayland client cannot place a
window in a screen corner; dikte.py sets QT_QPA_PLATFORM=xcb for that.
License
GPL-3.0, see LICENSE.