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.
Two things a real command turned up.
A job can run for ten minutes. "Report on the front page" is not a question
with an answer a second later, and having the corner narrate every tool it
touches for those ten minutes is worse than saying nothing. So that indicator
can be clicked away. It is the only one that can: a dictation is over in
seconds, and an indicator that swallows a click meant for the window underneath
has to earn it. The work carries on; what is muted is the progress, not the
outcome, which shows up whether or not the box was sent away. A faint cross on
the right says the box can be clicked, because a feature nobody can see is not
one. The next run starts visible again.
Muting leaves the state alone rather than setting it to something hidden, so no
later repaint puts the box back on the screen behind its own back.
Thinking effort is a setting now, offered once rather than three times: how hard
to think is one thing to want, and only the rungs differ. Claude takes it as
--effort, Codex as a model_reasoning_effort override, OpenRouter in the
reasoning field it already understood for cleanup. A level a provider does not
have lands on the nearest one it does, so "maximum" is xhigh on Claude and high
on Codex rather than an error or a silent drop. Left alone, nothing is sent and
each model does what it would have done.
Everything the last commit built assumed one agent was installed, which is a
poor assumption to bake into a dictation tool. So the provider is a setting, and
what it selects is one of three quite different things.
Claude Code and Codex are the same shape: a CLI, streaming JSONL, a session id
to resume, tools that reach the machine and whatever is connected to it. They
share the runner. What differs is spelled out where it differs, which is more
than the flag names: Codex has no system prompt to append, so the instruction
rides in front of the command with a rule between them; it confines its commands
in a sandbox rather than asking about them, so the permission setting is a
sandbox mode; and `-s` is not accepted by `exec resume`, so both settings go
through `-c` overrides, which are.
OpenRouter is the odd one and is meant to be. No tools, no files, no calendar:
it can say what the capital of Peru is and not what is in your diary, and the
settings box says so rather than letting it be discovered. It also has no
session to resume, so the conversation is kept here and resent, capped at 24
messages.
A stored conversation names the provider that made it, and is ignored by any
other: none of them can pick up another's thread, and a stale id would otherwise
fail every command until the timeout cleared it.
The interface calls the thing by its name, which in Turkish means the suffix has
to agree with it: Claude'a but Codex'e, Claude'u but Codex'i. A name dropped
into a sentence through t() cannot be inflected by that sentence, so it arrives
inflected, from a small table in i18n. English takes the name as it is and keeps
the preposition in the sentence.
whisper-1's verbose response carries a start and an end for every segment,
and the file tab was reading only the start, to build the [mm:ss] prefix.
Keeping the end as well is all an SRT needs.
The text stays the authority on wording and the segments on timing; they
meet at that prefix, which the cleanup model is already told to leave alone.
So a transcript that went through cleanup still turns into properly timed
subtitles. A line whose stamp matches no segment runs until the next line
starts, a line with no stamp at all joins the cue above it, and an end that
would run into the next cue is trimmed back.
The button is dead until a timestamped run finishes, because without
timestamps there are no segments to time anything with.
OpenRouter mirrors OpenAI's /audio/transcriptions field for field, so one
multipart request serves both providers and only the key, the base URL and
the model id change. That puts a dozen speech models behind the key that was
already there for the cleanup, and makes a single OpenRouter key enough for
the whole chain.
Its transcription endpoint takes no hint field, so the words from Cleanup
rules are not sent there; they still reach the cleanup model as a glossary.
Timestamps switch to openai/whisper-1, the namespaced id of the only model
that returns segment times.
The API tab is now grouped by role rather than by service, because a key no
longer belongs to a single job: both keys sit at the top, the two jobs below.
Each provider keeps its own model, so switching back and forth does not
overwrite the other one's.
The cleanup prompt now asks the model to fix words the transcriber misheard
when the context makes the intended one clear, and to leave them alone when
it does not. Speech models fail phonetically on proper nouns, and that is
exactly what context can recover.
The names you enter for the transcription hint are handed to the cleanup
model as a glossary too. Knowing the spelling is what lets it recognise
"kuber netis" as Kubernetes.
A failed cleanup used to be almost invisible: the raw transcript was pasted
and a progress line flashed by, so a rejected key looked exactly like
working dictation for days. It now leaves the indicator amber with the
reason, sends a notification, and records the error in the history. HTTP
401, 402 and 429 are reported as what they are, naming the service.
Also:
- Settings can test the OpenRouter key, not just the OpenAI one
- Tray menu and CLI gained Restart, which re-execs in place
- Defaults saved into the config by older versions are recognised by their
fingerprint and dropped, so an untouched prompt keeps getting improvements
- The IPC socket is user-only; Qt puts it in /tmp
The screenshots were downscaled to 430 px wide, which made the UI text
blurry. Restore them at native 1292 px as lossless WebP, which is also
half the size of the original PNGs (72 KB against 155 KB for the largest).
Rewrite every em dash in prose, comments, docstrings and interface strings
as ordinary punctuation.
Ctrl+Space starts and stops a recording. The audio goes to OpenAI for
transcription, a model on OpenRouter strips the fillers and restores
punctuation, and the result is copied and pasted into the focused window.
Only the Python standard library and PyQt6 — HTTP, multipart uploads and
WAV writing are all hand-rolled.
- pw-record captures raw 16 kHz mono PCM with a live level meter
- the corner indicator is drawn through XWayland, since a Wayland client
cannot position its own window
- silence is caught before it costs an API call, relative to each
recording's own noise floor, plus a filter for the stock phrases models
invent when handed silence
- audio and video files can be transcribed too, optionally with [mm:ss]
timestamps, chunked through ffmpeg for long inputs
- global shortcut installs as a KDE custom shortcut, with an evdev
listener as a fallback until the session is restarted
- Turkish and English interface, following the system locale by default
yusufipk