邊個講咗乜，同點樣確定Who said what, and how it is decided
最後更新：2026 年 9 月 12 日 Last updated 12 September 2026
一個待辦擺錯咗人名，比擺「未定」更差。
因為會有人讀到，然後唔去做 —— 佢見到唔係自己個名。 所以呢個 App 對「唔知」嘅處理，唔係估，係照寫「未定」。
An action against the wrong name is worse than an action against nobody.
Because somebody reads it and does not do it — it has someone else's name on it. So when this app does not know, it does not guess. It writes "Unassigned".
會議紀錄工具最容易出事嘅地方唔係聽錯字，係擺錯人。 呢版講埋個 App 實際上用緊嘅規則，同埋你可以點樣自己睇住佢執行。
The place a meeting-notes tool does real damage is not mishearing a word — it is attaching the wrong person to a sentence. This page sets out the rules the app actually applies, and how you can watch them fire.
邊把聲先算證明到What counts as proof of a voice
你自己嗰條聲軌係肯定嘅Your own track is certain
你支咪係獨立一條聲軌。喺嗰條聲軌講嘅嘢，一定係你講 —— 呢個唔使估。
Your microphone is its own channel. Anything on it is yours; there is nothing to infer.
分得出嘅聲，會有個一致嘅編號，唔會有個估返嚟嘅名A separated voice gets a consistent label, not a guessed name
對面嗰邊嘅人會被分開做唔同把聲。喺你話畀個 App 知邊個係邊個之前， 佢哋一路都係「Speaker 1」、「Speaker 2」—— 前後一致，但冇名。
The far side is separated into distinct voices. Until you say who they are, they stay "Speaker 1", "Speaker 2" — consistent throughout, and unnamed.
要擺個名落去，門檻係實數Attaching a name has an actual threshold
如果你之前幫某把聲改過名，個 App 會留低一小段聲做參考。 下次開會，要同時滿足兩個條件先會將個名擺返落去：嗰段參考聲有 七成或以上落喺同一個聲音群組，而且嗰個群組冇第二把已知嘅聲撞埋。
兩個人把聲似到黐埋一齊點算？個 App 會當係「分得唔夠細」， 自動加多一個人數再分過，最多再試兩次。如果試完都仲係分唔開, 嗰個群組就冇名 —— 唔會二揀一。
If you have named a voice before, a short reference sample is kept. At the next meeting a name goes back on only if two things hold: that reference falls 70% or more inside one voice cluster, and no second known voice touches that same cluster.
When two people's voices do collapse into one cluster, the app treats that as having split the audio too coarsely — it raises the speaker count and separates again, up to twice more. If they still will not come apart, that cluster is left unnamed. It does not pick one.
寫完之後仲有一道關A second pass over the finished minutes
寫紀錄嗰個語言模型收到嘅指示係唔准作人名。「大致上唔會作」對一份待辦清單嚟講 唔算一個標準，所以份紀錄寫完之後仲要再過一次。每個待辦嘅負責人會逐個檢查：
The model that writes the minutes is told not to invent an owner. "Mostly does not" is not a standard an action list can be held to, so the finished minutes are checked again. Every owner is tested:
- 係你自己 → 保留（錄音證明到）。
- 係一個分開咗嘅 Speaker 編號 → 保留（係一個真實、前後一致嘅人，只係未有名）。
- 係一個共用聲道上面嘅名，但證明唔到佢應承過 → 改做「未定」。
- 成份逐字稿都搵唔到呢個名 → 咁佢唔喺個會入面，亦冇人提過佢，即係個模型作嘅 → 改做「未定」。
- It is you → kept. The recording proves it.
- It is a separated speaker label → kept. That is a real, consistent person who simply has no name yet.
- It is a name on the shared channel who cannot be shown to have volunteered → changed to Unassigned.
- It is a name that appears nowhere in the transcript → they were not in the room and were not mentioned in it, so it came from the model → changed to Unassigned.
錄壞咗會出聲It tells you when the recording failed
2026 年 8 月 18 號，一個 63 分鐘嘅會議入面，支咪錄咗 64 秒就死咗。 當時乜提示都冇。份紀錄照寫出嚟，內容全部嚟自一邊, 但係讀落好似係成個會議咁 —— 呢種失敗係靜㷫㷫嘅，而呢個先係最危險嗰種。
而家停止錄音嗰陣，個 App 會比較兩條聲軌嘅長度，唔對路就會即刻彈個提示, 同時喺份文件最頂寫清楚。佢亦都會重新啟動死咗嘅咪、頂得住你中途換耳機或者音效裝置, 爆音亦都會提你。
On 18 August 2026 a microphone stopped 64 seconds into a 63-minute meeting. Nothing warned anyone. The minutes were written from one side of the conversation and read as though they were the whole of it. That kind of failure is silent, which is what makes it the dangerous one.
The app now compares the length of the two tracks when you press Stop and says so — in a dialog, and at the top of the document. It also restarts a stalled microphone, survives the audio device changing underneath it, and warns you about clipping.
點樣自己驗證How to check this yourself
上面每一條規則行過之後都會寫低。開 Console.app,
或者直接睇 ~/Library/Logs/CantalkieMeetings.log，搵呢兩行：
Every rule above writes a line when it fires. Open Console.app, or read
~/Library/Logs/CantalkieMeetings.log directly, and look for these two:
voices: 1 cluster(s) held two known voices — left unnamed attribution guard: 2 unverifiable owner(s) on the shared channel, 1 name(s) absent from the transcript
第一行係話：有兩把已知嘅聲分唔開，所以嗰個人冇被安上名。 第二行係話：份紀錄原本有幾個負責人，過唔到關，已經改做「未定」。 呢啲唔係錯誤訊息 —— 係啲規則喺度做嘢。
The first says two known voices would not separate, so nobody was named there. The second says the minutes came back with owners that did not survive the check and were changed to Unassigned. These are not error messages — they are the rules working.
呢版冇寫嘅嘢What this page does not claim
我哋喺度冇擺一個準確率百分比。 我哋內部試過，結果好好，但係嗰次嘅原始輸出冇留低。 呢個網站有條規矩：一個數字追唔返去一次有紀錄嘅測試，就唔會擺上網 —— 語音辨識嗰版係咁做，呢版都一樣。
重新做一次、留低原始輸出之後，數字就會擺返落呢版。
There is no accuracy percentage on this page. We have tested this internally and the result was good, but the raw output from that run was not kept. This site has a rule: a figure that cannot be traced back to a recorded run does not go on a public page — it is how the speech-recognition page is built, and this page is held to the same standard.
When the test is run again with its output kept, the numbers will appear here.
唔好將呢版當成保證唔會出錯。語音辨識會聽錯字, 分聲音會分錯，兩者都會。呢版講嘅係另一件事：個 App 唔會將一個佢證明唔到嘅名, 寫成好似證明咗咁。
None of this is a promise that nothing goes wrong. Speech recognition mishears words and speaker separation makes mistakes; both do. What this page describes is a different thing: the app will not present a name it cannot support as though it could.