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.
Who said what is the hard part of a meeting transcript, and the usual answer
is to hand one mixed recording to a model and ask it to tell the voices apart.
That guess is wrong often enough to be worse than useless in minutes, where a
decision attributed to the wrong person is a decision nobody made.
So the question never reaches a model. ffmpeg records the microphone and the
default sink's monitor as one stereo stream, you on the left and everyone else
on the right, and one process reading both is what keeps them aligned over an
hour. Each channel is transcribed on its own and the two are interleaved on a
single timeline, so attribution is settled by the wire a voice arrived on.
What a microphone picks up from the speakers lands on both channels; our copy
is dropped when it overlaps theirs in time and says nearly the same thing.
The stream is written to disk as it arrives rather than held in memory, so
length costs nothing and a crash costs the tail instead of the whole meeting.
Every stage the run reaches is recorded in meetings.jsonl, so a failure while
summarising does not throw away the transcription of an hour of audio: the
retry reads the transcript back out of the document and picks up from there.
A run that dies keeps its recording whether or not audio is being kept.
The minutes model is configured on its own, under Settings, with its own
prompt, and it is told who was expected in the room so the names come out
spelled right. It is told outright that the transcript is a record of other
people talking, not instructions addressed to it.
The built-in listener now holds several bindings rather than one, and the KDE
side is parameterised by desktop id, so the meeting toggle gets a shortcut of
its own on the same footing as the dictation one.
yusufipk