Cantalkie 對廣機

版本更新Release notes

最後更新:2026 年 8 月 31 日 Last updated: 31 August 2026

Cantalkie 會自己檢查更新。想手動更新:撳選單列個麥克風圖示 → 「檢查更新…」。想知自己而家用緊邊一版,撳同一個圖示, 版本號寫喺最頂。

Cantalkie checks for updates on its own. To update by hand, click the microphone icon in the menu bar and choose Check for Updates…. The version you are running is shown at the top of that same menu.

更新視窗係英文嘅 —— 嗰個係更新工具本身,唔係出咗事。

The update dialogs themselves are in English; that is the updater, not a fault.

1.2.2

2026-09-11 最新 Current

講慢啲,唔會再俾人斬開一截截 Speak slowly without your sentence being cut in half

講嘢中間停一停諗嘢,嗰句就會變成兩句,中間多咗個句號,跟住下一個字仲要變咗大階。 因為每一段都係獨立辨識嘅,而一段獨立嘅說話,辨識出嚟一定係一句完整句子嘅樣。 而家除咗睇停幾耐,仲會睇埋啲字:一段以 of、which、because 開頭,即係佢根本唔係一句嘅開頭; 一段以 a、my、is 收尾,即係佢未講完。掉句號嗰陣,順手將個大階都改返細階。

Pause mid-sentence to think and the sentence came back as two, with a full stop between them and a capital letter on the next word. Each chunk is recognised on its own, and a standalone chunk comes back looking like a complete sentence. Now the words are consulted as well as the clock: a chunk opening with of, which or because did not start a sentence, and one ending with a, my or is did not finish one. When the full stop goes, the capital goes with it.

嗯、誒、咧 —— 三個廣東話唔會用嘅字 Three Mandarin spellings that do not belong in Cantonese

「嗯」、「誒」(同「欸」)、「咧」係普通話嘅寫法。講廣東話嘅人寫嘅係 「唔」、「呀」、「呢」。而家出嚟就係後者。

嗯, 誒 (and 欸) and 咧 are Putonghua spellings. A Cantonese speaker writes 唔, 呀 and 呢, and that is what comes out now.

1.2.1

2026-09-03

揀返「原始」,記憶體即刻放返 Switch back to Raw and the memory comes back at once

1.2.0 之後,改寫模型閒置一段時間就會自動放返。但如果你用完「潤飾」之後 即刻揀返「原始」,嗰幾 GB 仲會喺度等個閒置時間到 —— 明明你已經講咗你唔要。 而家:你一揀「原始」,用唔著嗰啲就即刻放返,唔使等,亦都唔使撳「重新啟動引擎」。 由「潤飾」轉返「整理」都一樣,大嗰個模型即刻放返。

Since 1.2.0 the rewrite model is released after a period of inactivity. But if you finished with 潤飾 and switched straight back to 原始, those few GB still sat there waiting for the timer — after you had already said you did not want them. Now the release happens on the switch itself: no waiting, and no need for Restart Engine. Dropping from 潤飾 to 整理 frees the larger model the same way.

1.2.0

2026-09-03

唔再食住幾 GB 記憶體唔放 Several GB of memory, no longer held for nothing

一位測試員發現 Cantalkie 開住就長期食 5GB 以上記憶體, 就算你成日都冇用過「整理」「潤飾」。原因係負責改寫文字嗰個模型一開機就載入, 而且載入之後直到你關閉 app 為止都唔會放返出嚟。 而家:一開機唔會長期霸住,用完之後閒置五分鐘就自動放返。 真係要用嗰陣重新載入大約一秒。

A tester noticed Cantalkie sitting at 5 GB or more of memory the whole time it was open, whether or not you ever used 整理 or 潤飾. The rewrite model was loaded at launch and then never released until you quit. Now: nothing is held at launch, and after a rewrite the model is released once you have been idle for five minutes. Loading it again when you next need it costs about a second.

打一個詞唔會再自動加句號 A single word no longer gets a full stop

講「電郵」出「電郵。」,講「確認」出「確認。」—— 用嚟填 search 格、檔案名、人名嗰陣,個句號永遠係錯。 而家:完整句子照舊有標點,但淨係一個詞或者一小段、入面又冇逗號嗰啲, 後面嗰粒句號會除返。

Say 電郵 and you got 電郵。; say 確認 and you got 確認。 Dictation is often filling a field — a search box, a file name, a person's name — where a full stop is simply wrong. Full sentences keep their punctuation; a short fragment with no comma inside it now loses the trailing mark.

出咗韓文、日文假名嗰啲,會即刻重新聽過 Korean or kana in the output is now re-decoded

講「用戶」,出「영국。」—— 唔係偶然,係次次都咁。 冇人用廣東話、普通話或者英文講嘢會講出韓文字母、假名、泰文、西里爾字母。 而家一見到呢啲字,嗰段會自動重新解碼一次,重新解碼唔乾淨就唔用。

Say 用戶, get 영국。 — not occasionally, but every time. Nobody dictating Cantonese, Mandarin or English produces Hangul, kana, Thai or Cyrillic, so a chunk that comes back in one of those scripts is now decoded a second time, and the retry is kept only if it comes back clean.

「聽唔清楚」唔會再用嚟講一件我哋自己做錯咗嘅事 "Didn't catch that" is no longer used for our own fault

有一種情況下,個 app 其實聽到你講嘢,但係佢自己將啲字掉咗, 然後同你講「沒有聽清楚」。咪係話麥克風有事囉 —— 兩樣都唔係真。 而家佢會講返實情,同埋俾你即場交一份問題報告。

There was a case where the app heard you perfectly, discarded the words itself, and then told you it had not caught them — which blamed your microphone for our fault. It now says what actually happened and offers you the problem report.

「重新啟動引擎」而家係中文 Restart Engine is now in Chinese too

介面預設係繁體中文,但成個選單入面就得呢一行係英文 —— 偏偏佢就係引擎死咗之後最常要撳嘅嗰一行。

The interface opens in 繁體中文, yet this one line stayed in English — and it is the line you most often need when the engine has stopped.

1.1.11

2026-08-31

「係」唔會再寫成「系」 係 no longer comes out as 系

「唔係」「係咪」「真係」—— 以前大概每三個「係」就有一個寫咗做「系」。 呢個係語音模型自己嘅問題,唔關你部機事,亦都唔會喺簡體輸出度出現。 而家改返,同時保護住真係要用「系」嗰啲詞:系統、系列、中文系。 喺一個 6,933 句嘅公開測試集上面改咗 477 處,一處都冇改壞。

「唔係」, 「係咪」, 「真係」— until now, roughly one 係 in three came back as 系. That is the speech model, not your Mac, and it never happened in Simplified output. Now corrected, with a protect list for the words that genuinely take 系 — 系統, 系列, 中文系. Measured across a public 6,933-utterance set: 477 corrections, none of them wrong.

「真係好」唔會再變「真繫好」 真係好 no longer becomes 真繫好

呢個更加古怪:轉繁體嗰個工具本身將「系好」當咗普通話「繫好安全帶」嘅「繫好」, 所以「真系好」「唔系好」出咗「真繫好」「唔繫好」。喺我哋嘅測試入面, 「真係好」係俾呢個問題搞到最多嘅一句。

Stranger than the one above: the Traditional-Chinese converter itself reads 系好 as the Mandarin 繫好 of 繫好安全帶, so 真系好 and 唔系好 came out as 真繫好 and 唔繫好. 真係好 was the single most-affected phrase in our testing.

「睇吓」「試吓」嗰個「吓」 The 吓 in 睇吓 and 試吓

以前會出「睇嚇」「試嚇」——「嚇」係嚇親人嗰個嚇,唔係「睇吓啦」嗰個。 而家分得返開,「嚇到」「嚇死」「唔好嚇我」照舊唔會郁。

睇吓 and 試吓 used to come back as 睇嚇 and 試嚇 — 嚇 as in to frighten, not the softening particle. Told apart now, with 嚇到, 嚇死 and 唔好嚇我 left alone.

選單而家講得出自己係邊個版本 The menu can now tell you which version it is

選單第一行以前淨係得一個 build 時間,所以「你用緊邊個版本?」呢條問題, 個 app 自己答唔到。而家寫住 Cantalkie 1.1.11 (23), 同「檢查更新…」睇嘅係同一組數字。

The first line of the menu showed only a build timestamp, so "which version are you on?" was a question the app could not answer. It now reads Cantalkie 1.1.11 (23) — the same two numbers the updater quotes.

1.1.10

2026-08-21

⌘V 而家貼得到 —— 由第一個版本開始就貼唔到 ⌘V now works — it never had

授權碼、問題報告個輸入格、詞庫、設定入面每一個輸入框, 由頭到尾都貼唔到嘢,打字先得。原因唔係嗰啲格, 係 Cantalkie 住喺選單列、冇menu bar,而 macOS 嘅 ⌘V 要經 menu 先派得到落去。 而家補返。授權碼隔籬亦都加咗個「貼上」掣, 佢仲識喺你複製嘅成行字入面淨係揀返條 key 出嚟。

The licence key, the problem-report box, your vocabulary, every field in Settings — none of them could be pasted into, only typed. Not the fields' fault: Cantalkie lives in the menu bar with no menu of its own, and on macOS ⌘V is delivered through a menu. Fixed. There is also a Paste button beside the licence key, which picks the key out of a whole copied line.

「電郵支援」㩒落去唔會再乜都唔發生 Email support no longer does nothing

如果你部 Mac 冇設定好郵件程式(好多人淨係喺瀏覽器睇 Gmail), 㩒落去係完全冇反應嘅 —— 冇窗、冇錯、乜都冇。而家個地址 一定會複製到剪貼簿,貼落你平時睇電郵嗰度就得。

If your Mac has no working mail route — and plenty of people only ever read Gmail in a browser — the button did nothing at all: no window, no error. The address is now always copied to the clipboard; paste it wherever you read your mail.

試用期完咗嗰個授權視窗,之前㩒唔到「完成」 The licence window was cut off when the trial ended

個視窗高度寫死咗,而試用期完咗嗰陣內容會長啲 —— 最底一行(「電郵支援」同「完成」)跌咗出窗外。而家高度跟返內容。

The window had a fixed height, and the expired state is taller — the bottom row, with Email support and Done in it, fell off the end. It now sizes to its content.

「書面語」唔會再寫「同他」 書面語 no longer writes 同他

講「同佢講聲,我遲少少到」,以前出「同他説一下」—— 每個字都啱,但香港人唔會咁寫,應該係「跟他說一聲」。而家亦都識執你講嘢途中嘅更正: 「我今晚 send 俾你,唔係,其實聽朝先得」出「我明早會發送給你」, 唔會再將「唔係,其實」原封不動搬入去。

「同佢講聲」used to come back as 「同他説一下」— every character correct, and not what a Hongkonger writes. It is 跟他說一聲. Written mode also resolves a spoken self-correction now, the way Polish already did.

數字:唔會再幫你「填準」一個你冇講過嘅數 Numbers: no more inventing a figure you never said

講「五六百蚊」,以前會出「600蚊」—— 一個你從來冇講過嘅準確數字。而家「五六百蚊」照樣係「五六百蚊」。 同一批仲修好咗:「幾十個」唔會變「幾10個」、「二點五公斤」唔會變「2點5公斤」、 「開一個會議」唔會變「開1個會議」、「三點五十分」而家出「3點50分」。

Say 「五六百蚊」 (five or six hundred) and it used to come back as 600 — a precise figure you never said. Fixed, along with three phrases that were coming back half in digits and half in characters.

1.1.9

2026-08-20

出事嗰陣,個 App 而家識自己開口 When something breaks, the app now says so

引擎開唔到、模型載入唔到、或者上次死過機 —— 以前你要自己估, 而我哋要靠你喺 WhatsApp 講返一兩句先估到發生咩事。而家個 App 會問你 要唔要寄一份報告,你亦可以隨時喺選單列揀「回報問題…」。

The engine failing to start, a model failing to load, a crash last time — you used to be left guessing, and we were left piecing it together from a line or two on WhatsApp. The app now offers to send a report, and 「回報問題…」 is in the menu whenever you want it.

佢永遠唔會自己寄,而且寄乜你睇得一清二楚 It never sends by itself, and you read every word first

撳「寄出」之前,成份報告會完整咁列晒喺個窗度 —— 你睇到嘅,就係我哋收到嘅。入面係版本、機型、模型狀態、權限, 同兩個紀錄檔嘅最後 200 行。你講過嘅字唔會喺入面,除非你自己剔咗 「連埋我最後一次講嘅結果」嗰格 —— 佢預設熄咗。詞庫入面嘅字(好多時係人名) 同你部機嘅使用者名稱寄出之前會遮走,錄音永遠唔會寄。唔想上網就撳 「改為存落桌面」。報告 90 日之後自動刪。

Before you press send, the entire report is laid out in the window — what you see is what we get: version, machine, model states, permissions, and the last 200 lines of two logs. Your words are not in it unless you tick the box for your last dictation, which starts unticked. Terms from your own vocabulary — often people's names — and your Mac's username are masked before anything leaves. Audio never leaves at all. There is a Save to Desktop button if you would rather not upload. Reports are deleted after 90 days.

收工嗰陣唔會再彈個「有冇問題?」出嚟嘈你 No more false alarm on the way out

閂 App 會順手熄咗個引擎,而個 App 見到引擎死咗就當出事 —— 結果你一 quit 佢就問你要唔要回報。引擎自己重開得返嗰啲情況一樣會誤報。 而家佢會等十五秒再問一次個狀態:捱得過十五秒都仲係壞嘅,先至值得回報。

Quitting kills the engine, the app noticed the engine had died, and asked whether you wanted to report it — on the way out. A restart that healed itself nine seconds later triggered the same thing. It now waits fifteen seconds and asks the state again.

1.1.8

2026-08-19

「書面語」由頭到尾冇行過 —— 修好咗 書面語 had never actually run. Fixed.

如果你揀過「書面語」或者「潤飾」,出返嚟仍然係口語 —— 唔係你撳錯掣,係佢真係冇行過。 今次逐句量度過先出:講「個 proposal 我今晚 send 俾你,唔係,其實聽朝先得」, 「潤飾」而家會執行埋你嗰個更正,出返「個 proposal 我聽朝 send 俾你」—— 舊嘅版本會將成句原封不動咁交返俾你。

If you ever chose 書面語 or Polish and got spoken Cantonese back, you were not pressing the wrong keys — it genuinely never ran. Now measured case by case: say 「個 proposal 我今晚 send 俾你,唔係,其實聽朝先得」 and Polish acts on the correction, giving 「個 proposal 我聽朝 send 俾你」. The previous version handed the whole sentence straight back.

設定精靈唔會再示範一個你部機行唔到嘅功能 The walkthrough no longer demos what your Mac cannot run

頭三步練習入面,教「書面語」嗰一步需要一個 5 GB 模型, 而安裝嗰陣淨係會落 2.5 GB 嗰個 —— 即係每一部機喺裝機當日都做唔到嗰一步, 但個畫面仲要扮成功。而家:模型未齊就唔會出嗰一步,完成頁會講返點解跳過; 你落咗個模型之後,佢會補返嗰一課俾你,順便即場試一次。

The practice step that teaches 書面語 needs the 5 GB model, and setup only ever downloads the 2.5 GB one — so on the day you installed, that step could not work, and the screen claimed success anyway. Now it is skipped until the model is there, the last page says why, and the lesson comes back to you when the download finishes, with a live try.

Caps Lock 唔會再死喺度 Caps Lock can no longer be left dead

兩個唔同嘅情況,兩個都修好咗。開 App 之前 Caps Lock 著咗燈, 以前會整部機打大階而冇得熄 —— 而家開之前會先熄返佢。 用「強制結束」閂咗個 App,以前 Cantalkie 冇機會將粒掣還返俾系統, Caps Lock 之後喺所有 App 都冇反應,要重開機先得 —— 而家正常閂、登出、 甚至系統送出嘅結束訊號,都會還返粒掣。

Two separate cases, both fixed. Caps Lock already on when the app launches used to leave the whole Mac in uppercase with no way out. And force-quitting the app used to strand the key remap, so Caps Lock did nothing in every app until you rebooted. Quitting, logging out and system termination all hand the key back now.

喺 Finder 撳個 App 唔會再乜都唔發生 Clicking the app now opens something

Cantalkie 冇窗口,佢住喺選單列 —— 所以喺「應用程式」度撳佢, 以前係真係乜都唔會出,同死咗機一模一樣。而家會彈設定出嚟, 同埋閃一閃選單列個圖示話你知佢喺邊。

Cantalkie has no window — it lives in the menu bar — so clicking it in Applications used to do nothing at all, which looks exactly like a hang. It now opens Settings and blinks the menu bar icon to show you where it lives.

免提模式 Hands-free

㩒一下 ⌥⇧Space 開始講,再㩒一下停 —— 唔使一直㩒住個掣。 講長嘢嗰陣先有用。錄音期間邊緣會一直亮住,選單列圖示會變紅, 10 分鐘之後自動停:一個冇人記得開過嘅麥克風,一定要自己閂。

Tap ⌥⇧Space to start, tap it again to stop — no holding a key through a long passage. The edge glow stays on and the menu bar icon turns red throughout, and it stops itself after 10 minutes: a microphone nobody remembers opening has to close by itself.

其他Also

  • 冇開「輔助使用」嘅時候,啲字會入剪貼簿 —— 嗰個提示而家有粒掣, 直接帶你去開,唔使自己搵。
  • 功能行唔到嘅時候會直接同你講「行唔到,上面啲字係未經處理嘅」, 唔會再扮做咗。
  • 記憶體 12 GB 以下嘅 Mac,練習嗰幾步嘅講解換成貼合你部機嘅版本 —— 唔會再聲稱一啲你部機做唔到嘅嘢。
  • Without Accessibility the text lands on the clipboard — that message now carries a button straight to the setting instead of just naming it.
  • When a feature cannot run it says so — "the text above is unedited" — rather than pretending it worked.
  • On Macs under 12 GB the practice steps explain what that machine actually does, instead of claiming work no model did.

1.1.7

2026-08-15

「整理」以前其實冇改過你嘅字 —— 修好咗 Clean-up was doing nothing. Fixed.

出廠設定用嗰個模型,我哋第一次認真量佢,先發現佢係原文入、原文出: 231 個字入去,231 個字出返嚟,一個字都冇改過。 即係話,如果你之前揀過「整理」而覺得同「原始」冇分別 —— 你冇睇錯。 換咗模型之後,同一段字由 231 字執到 150 字:um、uh、呃、嗯、 講到一半重新開過嗰啲全部清走,句子斷返開,標點補返。

The model shipped as the default turned out to hand the transcript straight back — 231 characters in, 231 characters out, nothing changed. So if you picked Tidy and it read exactly like Raw, you were not imagining it. With the new model the same passage goes from 231 characters to 150: English and Cantonese fillers gone, false starts gone, sentences broken properly, punctuation added.

唔使再裝 Ollama Ollama is no longer needed

以前想用文字整理、潤飾同書面語,要去第二個網站裝多一個軟件, 再開「終端機」打指令 —— 由零到用得晒,一共 8 個步驟、兩個網站、5.6 GB。 而家全部收埋喺 Cantalkie 入面,App 自己攞需要嘅檔案, 一次 download 就完。改寫模型亦由 9.6 GB 縮到 5.15 GB。

Tidy, Polish and 書面語 used to require a separate program from another website plus a Terminal command — eight steps, two websites, 5.6 GB in total. All of that now happens inside Cantalkie. The app fetches what it needs itself, and the rewrite model dropped from 9.6 GB to 5.15 GB.

跟你部機記憶體決定行邊啲功能 Features now match your Mac's memory

記憶體細過 12 GB 嘅 Mac,唔會再見到「潤飾」同「書面語」。 我哋喺 8 GB 機實測過:行呢兩個功能會逼部機用 swap,連本來好順嘅語音輸入都會拖慢。 與其畀你撞完先知,不如唔 offer。

「整理」照用,會換成一個唔食記憶體嘅版本 —— um、uh、呃、嗯、 重複咗嘅開頭一樣清走。每一個灰咗嘅選項下面都會寫明原因, 唔會靜靜哋灰咗畀你估。

On Macs with less than 12 GB of memory, Polish and 書面語 are no longer offered. Measured on an 8 GB machine, running them forces the Mac into swap and slows down dictation itself — so we would rather not offer them than let you find out the hard way. Tidy still works, using a version that needs no extra memory. Every disabled option states its reason instead of just going grey.

其他Also

  • 一口氣講完一大段而中間冇停 —— 而家會補返標點,唔會出返一大串冇標點嘅字。
  • 開機之後第一次用「整理」快咗:模型喺 App 開嗰陣就預熱,唔使你等。
  • 收音收唔到聲嗰陣,會直接叫你去「系統設定 → 聲音 → 輸入」, 唔會再淨係講一句「聽唔到」。
  • 更新過程本身穩陣咗:download 途中個選單唔會再卡住喺 0%。
  • A long sentence spoken in one breath now comes back punctuated instead of as one unbroken run.
  • The first clean-up after launch is faster — the model warms up when the app starts rather than when you are waiting for it.
  • When no sound is reaching the microphone, the app now points you at System Settings → Sound → Input instead of only saying it did not catch that.
  • Updating itself is steadier: the menu no longer freezes at 0% during a download.

更新之後第一次用「整理」,會 download 大約 2.6 GB 嘅模型,淨係一次。(同你第一次裝 App 嗰 2.5 GB 語音模型係兩件事。) 「潤飾」同「書面語」第一次揀嘅時候,會再 download 一個 5.15 GB 嘅模型。

The first time you use Tidy after updating, about 2.6 GB is downloaded, once. (That is separate from the 2.5 GB speech model downloaded on first install.) Polish and 書面語 fetch a further 5.15 GB the first time you choose them.

1.1.1 – 1.1.6

2026-08-08 → 2026-08-12 測試期 Beta

呢一段係封閉測試期間出嘅幾個版本,主要係執 bug。

A run of releases during the closed beta, mostly fixes.

  • 喺 macOS 26 以下嘅 Mac 開唔到語音引擎 —— 已修正。 我哋打包嗰陣夾錯咗一個只有最新系統先載入到嘅檔案入去, 所以喺 macOS 14/15 上面,個語音引擎一用到 GPU 就即刻死, 選單只會寫住「Engine stopped unexpectedly」。一位測試員貼返段錯誤訊息先至捉到。 之後加咗一個出廠檢查,同一個錯出唔到街。
  • 要另外裝軟件先行到嘅功能,唔會再扮正常。 文字整理、潤飾同書面語以前如果冇裝到需要嘅嘢,會靜靜哋原文照返, 睇落好似個功能壞咗。改咗之後會直接變灰,下面寫住原因。
  • 書面語改用香港嘅書面中文用詞(項目、電郵、軟件、質素、你 而唔係 您), 唔會再出台灣或者內地嘅講法;亦唔會再自作主張軟化你原本講嘅嘢。
  • Fixed: the speech engine would not start on any Mac below macOS 26. A file that only the newest system can load was packaged in by mistake, so on macOS 14 and 15 the engine died the moment it touched the GPU and the menu simply read "Engine stopped unexpectedly". A tester sent back the error text, which is how it was found. A factory check now blocks that class of mistake.
  • Features that need an extra component no longer pretend to work. Tidy, Polish and 書面語 used to hand the text back unchanged when the component was missing, which read as the feature being broken. They are now disabled with the reason stated.
  • 書面語 now writes Hong Kong written Chinese rather than Taiwan or Mainland phrasing, and no longer softens what you actually said.

1.1.0

2026-08

第一個公開發佈嘅版本:離線廣東話/普通話/英文語音輸入, 㩒住 Caps Lock 講嘢,放開就打字入去你游標嗰個位。已經過 Apple 公證。

The first public release: offline Cantonese, Mandarin and English dictation. Hold Caps Lock, speak, let go, and the text lands wherever your cursor is. Notarised by Apple.

舊版本點算? Older versions

我哋只會派最新嗰個版本。如果你想試返舊版, 電郵我哋,講低你想要邊個版本號同點解。

Only the current version is distributed. If you need an older build, email us with the version number and why.