Whisper hears in thirty second windows and decides for itself where one cue
ends and the next begins. A chunk that starts in the middle of a sentence can
answer with one cue per window, twenty seconds of text at a time, for the whole
rest of the chunk: a twenty five minute recording was fine until 20:00, which
was where the second cut fell, and ran on in paragraphs from there. Sending the
same audio in one request instead of three gives cues of two and a half seconds
throughout.
The cuts were only ever there for the upload limit, and we were the ones
walking into it: ffmpeg opened a 24 MB m4a into 48 MB of uncompressed WAV, over
the 25 MB the APIs take, so the file had to be cut every ten minutes. As mp3 it
is 9 MB, and an hour of speech goes in one request. A server on this machine is
still handed the WAV, where nothing is uploaded and the encoder would only cost
quality. How long a chunk may be is now measured from the encoded file rather
than assumed from a bitrate.
Where a file still has to be cut, the chunks overlap by a whisper window and
stitch() drops the telling that was cut short, keeping the one that heard the
sentence whole. Meetings, which upload the WAV itself and so still cut every
ten minutes, get the same stitching.
Cleanup was the one step with only one place to run. Speech to text has three
providers behind a setting and the agent has three behind another, but the
model that drops the "eee"s out of a sentence was always a request to
OpenRouter, which meant a second key on a machine that already pays for a model
and already hands whole dictations to it as commands. Claude Code and Codex can
rewrite a sentence as easily as they can put something in your calendar, and
now they may.
cleanup.py is where that choice lives, so worker, the file transcriber and the
meeting all ask the same question rather than each building the same OpenRouter
request. What comes out of a CLI that failed is a CleanupError, which is an
ApiError, because to the chain a cleanup that failed is a cleanup that failed
however it was run: the raw transcript is still pasted and the reason still
shows in the corner, unchanged.
Neither CLI is given anything it does not need for the job. No tools, no MCP
servers, no session to resume, and the home directory rather than wherever the
agent is pointed, since a project's instructions have opinions about how text
should be written and none of them are about this transcript. The transcript
goes in fenced the same way the OpenRouter call fences it, because it is
material rather than an instruction however much of it reads like one. Claude
takes the cleanup rules as its whole system prompt; Codex has no system prompt
of its own, so they ride in front of the text, and its answer is read from the
file it writes on the way out rather than from a stdout that also carries a
header, its thinking and a token count.
The cost is seconds. OpenRouter answers in about one, a CLI in six or seven,
because each one opens a whole session to do it. That is the trade the box
says out loud, and the default has not moved: OpenRouter cleans up until you
say otherwise.
Codex's two lowest thinking levels now ask for "low". "minimal" was its bottom
rung until the newer models replaced it with "none", and each of them answers
the other's word with a 400, which the agent has been quietly hitting too.
In the settings window the model box belongs to whoever is chosen rather than
meaning three different things in turn, since an OpenRouter id and a Claude
alias do not belong in the same field, and under it is the same "found it or
not" line the agent tab has. dikte doctor asks about the program instead of the
key when a CLI does the cleaning, and the history records which model actually
did it.
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