Known behaviours

Things you may notice while using Macalla, what causes them, and what to do about them.

Things you may notice while using Macalla, what causes them, and what to do about them. Most come from your hardware or the operating system rather than from Macalla itself; a few come from the fixed rules Macalla uses to tidy your words. Every figure states the machine it was measured on.

Apple Intelligence leaves most of a long Chinese, Japanese, Korean or Vietnamese meeting unsummarised

What you notice. You ask for a summary of a meeting held in Chinese, Japanese, Korean or Vietnamese, with Apple Intelligence chosen on the Enhancement card. Most of the summary is missing: any part of the meeting longer than about 1,900 characters can come back with an error, and one longer than about 3,800 always does. A short meeting, or a short part at the end of a long one, is still summarised.

Why it happens. Apple’s on-device model can take 4096 tokens at a time, counting the instructions, the transcript and its reply together. That is Apple’s figure, not a measurement of ours. The instructions leave about 3,800 tokens. Macalla plans to use about half of those, about 1,900, for each part of the transcript and leave the rest for the reply, so it cuts a meeting into parts of up to about 5,750 characters: about 1,900 tokens where a token is three or four characters, as Apple describes English. In Chinese, Japanese, Korean and Vietnamese a token is about one character, so the same part is nearer 5,750 tokens. A part longer than about 3,800 characters cannot fit at all, even before the reply. One longer than about 1,900 characters fits only if the reply is short, and fails whenever the reply needs more than the room left. Counting each part’s tokens before sending it would fix this, and is planned for a later phase.

What to do. For these meetings, choose a cloud model with a larger context on the Enhancement card. It summarises them in parts of the same size.

A phrase you repeat on purpose can lose a copy

What you notice. You say a short phrase more than once on purpose, such as counting “one two three one two three one two three”, and the pasted text has one copy fewer: “one two three one two three”.

Why it happens. When people restart a phrase they often say it twice, as in “we need to we need to ship”, and Macalla removes the extra copy. From the words alone it cannot tell a stumble from a repeat you meant, so it treats every repeat the same way:

• A run of one to three words said again straight away is removed in pairs: two copies become one, three become two, and four usually become two.
• A single word said twice is removed only if it is a short everyday word, such as “the”, “to”, “it” or “we”. “very very good” and “azure azure” stay as you said them.
• The copies must follow each other directly. A comma, dash or full stop between them keeps every copy.

What to do. When a repeat matters, put punctuation between the copies, for example by saying “comma” after each one, or check the pasted text.

Bluetooth headset microphones make playback thinner while you dictate

What you notice. You are listening to music or a podcast through Bluetooth headphones, such as AirPods, and they are also the microphone Macalla listens through. When you start dictating, the audio cuts out for a moment and comes back thinner, sometimes a little louder. About three seconds after you finish, it returns to normal.

Why it happens. A Bluetooth headset has two ways of working. For playback it uses a high-quality stereo mode. As soon as any app records from its microphone, the headset switches to a headset mode that carries your voice and the audio at the same time, at lower quality. It switches back once recording stops. This is how Bluetooth headsets work. Another dictation app we tested behaves the same way, and we know of no way to avoid it while the headset’s own microphone is recording. Macalla is open source: if you find one, an issue or a pull request is welcome.

What to do. Keep the headphones as your output and choose your Mac’s built-in microphone, or another wired or USB microphone, on Macalla’s Microphone card. Playback then stays as it is while you dictate.

What we measured. On a MacBook Pro (Apple M1 Pro), macOS 26.6.2, with AirPods Pro, on 12 September 2026:

• Recording from the AirPods microphone switched their playback from 48,000 Hz to 24,000 Hz within a fifth of a second of recording starting.
• After recording stopped, playback was back at 48,000 Hz about two seconds later, and audio resumed about three seconds later.
• With the MacBook’s built-in microphone chosen instead, playback stayed at 48,000 Hz throughout.
• Another dictation app recording from the same AirPods produced the same switch.

While the headset is in that mode its microphone also runs at 24,000 Hz. Macalla converts every microphone to 16,000 Hz before recognition, so that rate is enough for the speech engine. Whether Bluetooth’s compression of the microphone audio changes how accurately your words are recognised has not been measured.

Without headphones, the other side of a call can appear under You as well

What you notice. In a live meeting on a call, some of what the other people said appears twice: once under Others, and again under You.

Why it happens. A live meeting hears your microphone and the sound your Mac plays as two separate streams. When the call plays through your Mac’s speakers, your microphone hears it too, so the same words reach both streams. Macalla does not remove the repeats, because text that repeats what someone else said can also be something you really said, such as reading a point back to agree to it. Removing it would delete your words without telling you. How often this happens, and with which microphones, has not been measured.

What to do. Wear headphones for calls you record as a live meeting. The call then reaches Macalla only as Others.

Disconnecting or switching headphones can end a live meeting early

What you notice. A live meeting stops on its own and the Files page says why: for example that a device went away, or that the call’s sound changed, as it does when headphones connect. The transcript up to that moment is kept.

Why it happens. Macalla hears each side through a device, and when a device disappears, or the sound your Mac plays changes format, it cannot go on hearing that side correctly. Rather than carry on with audio it can no longer trust, it ends the meeting and keeps everything it had already heard. A short break in the sound, too brief to change anything, is filled with silence instead, so the meeting carries on and each side keeps its place in time.

What to do. Choose your microphone and headphones before you start a meeting, and leave them connected until you stop it. If a meeting ends early, start a new one: the text you already have stays on the Files page until you copy, save or clear it.

What we saw. On a Mac mini (Apple M1), macOS 27.0, on 23 September 2026: disconnecting AirPods during a live meeting ended it, and the words heard before the disconnect were kept.

Over a long live meeting, the two sides’ timings can drift slightly apart

What you notice. Late in a long meeting, a reply under Others can appear a moment before or after the words under You that it answers.

Why it happens. Your microphone and the sound your Mac plays are timed by two different clocks, and no two clocks run at exactly the same speed. Over a long meeting the small difference adds up. Macalla does not correct for it. How far apart the two sides drift has not been measured.

What to do. Nothing needs doing: the words are unaffected, only the order of two turns that are close together. When a reply looks out of place late in a long meeting, read it with the turn next to it.

A live meeting from the command line hears only your microphone

What you notice. macalla live-meeting records you, but not the other side of a call playing on your Mac.

Why it happens. macOS decides who may record the sound your Mac plays by the app that started the program. For a command run in a terminal, that is the terminal, and a terminal is not set up to ask for that permission, so macOS gives it silence. The command says it records your microphone only.

What to do. Record calls from the Files page of the desktop app, which asks for the permission itself.

What we saw. On a Mac mini (Apple M1), macOS 27.0, on 21 September 2026: a program run from a terminal received the sound the Mac was playing as silence, and macOS’s own log showed it refusing the terminal, which had not asked for the permission.

Quitting while a live meeting records can end in a crash report

What you notice. You quit Macalla during a live meeting. The next time it opens, the meeting is offered back on the Files page as expected, but macOS may also have shown a report that Macalla quit unexpectedly. We have not seen this happen: it is worked out from the code.

Why it happens. When you quit, Macalla waits a few seconds for the meeting to finish hearing its last piece, and saves what was heard before it closes. If that piece is still being heard when the time is up, the speech engine can still be running as Macalla closes, and that can end in the crash report. What was heard is already saved by then.

What to do. Stop the meeting before you quit. Then nothing is still being heard when Macalla closes. If you do see the report, the meeting is still offered back the next time Macalla opens.

Dictating while several files are being heard can be slower

What you notice. You drop several files on the Files page and dictate while they are being heard. The dictation takes longer than usual to appear.

Why it happens. Macalla hears the files on the same speech engine that dictation uses, rather than loading a second copy of it, so a dictation can wait behind the part of a file the engine is hearing. We have not measured how much longer it takes.

What to do. Wait for the files to finish, or press Stop on the Files page. Files already saved stay saved.

Changing the speech engine while several files are being heard

What you notice. You change the speech engine while the Files page is hearing several files. The recordings that were still waiting are not heard, and each says the engine was changed.

Why it happens. A batch of files is heard on the one engine it started with, so none of its files is heard on a different engine without a word. A batch never loads a second engine of its own; the file being heard when you switch finishes on the old one. Transcript files, which need no engine, are still read and saved.

What to do. Choose or drop the recordings that were not heard again, and they are heard on the new engine.