這不是「語音轉文字」工具 —— 它是把轉完之後、藥名和數字被聽錯的醫學逐字稿救回來的修復工具。繁中醫學、越用越準、絕不亂編。
This is not a speech-to-text tool — it rescues medical transcripts after transcription, fixing the misheard drug names and numbers. Built for Traditional Chinese medical talks; it gets sharper with use, and never makes anything up.
這是一個 Claude 的 skill（技能）。你把演講錄音跑「語音轉文字（ASR）」之後，常會得到一份藥名、試驗名、數字被聽錯的稿子；這個 skill 就是專門幫你把它小心地修好、重寫成乾淨又正確的繁體中文，最後給你一份乾淨的稿子＋一份「還要再確認什麼」的清單。
This is a Claude skill. After you run a talk's recording through voice-to-text (ASR), you often get a transcript full of misheard drug names, trial names, and numbers. This skill carefully repairs it and rewrites it into clean, correct Traditional Chinese — and hands you back a clean transcript plus a checklist of what still needs confirming.
先把類別搞對：它不做「轉錄」。 GitHub 上有數百個「醫學轉錄」工具，它們做的都是同一件事 —— 把聲音變成字。這個專案做的是下一步：把已經轉完、但被聽錯的字救回來。實際搜過 GitHub 上的同類專案後會發現：做這一步的，幾乎都是學術論文附的研究程式，或是純英文、只做簡單「找字換字」的小工具 —— 沒有一個是一般人拿來就能用的成品。
Get the category right first: this is not "transcription". There are hundreds of "medical transcription" tools on GitHub, and they all do the same job — turning sound into text. This project does the next step: rescuing the text after it has been misheard. Search GitHub for projects that do this step and you'll find almost only research code attached to academic papers, or English-only find-and-replace utilities — none of them is a finished tool ordinary people can just pick up and use.
在這個近乎無人的縫隙裡，它的核心優勢是：
Inside this almost-empty niche, its core strengths are:
① 絕不亂編，安全第一
① It never fabricates — safety first
- 它不會替任何「無法確認的資訊」硬填答案 —— 人名、藥名、數字、劑量都一樣，也不會捏造文獻引用。 It won't fill in anything it can't confirm — names, drugs, numbers, and doses alike — and it won't invent citations.
- 學術界的校正模型為了壓低「字錯率」，常會自由改寫、甚至腦補；這裡明確拒絕。不確定的地方會清楚標記，讓你一眼看到哪裡還要再確認。對醫學內容來說，這才是對的做法。 Academic correction models often rewrite freely — even hallucinate — to push down their error rate; this project explicitly refuses to. Anything uncertain is clearly flagged so you can see at a glance what still needs checking. For medical content, that's the right trade-off.
② 多稿交叉比對，準確度跳級
② Cross-checking multiple transcripts makes accuracy jump
- 把同一場演講餵給不同的轉錄軟體，會得到幾份「錯的地方不一樣」的稿子。把它們一起丟進來（2–4 份最佳），它會逐句交叉比對、用多數決收斂，真正的疑點自然浮出來。 Run the same talk through different transcription apps and you get several transcripts that are each wrong in different places. Hand them all in (2–4 is ideal) and it cross-checks them line by line, letting the majority decide, so the real trouble spots surface by themselves.
- GitHub 上目前沒有其他工具把這件事包成能直接用的功能。 No other tool on GitHub currently packages this as a ready-to-use feature.
③ 有一個會越用越聰明的詞庫
③ It has a word list that gets smarter the more you use it
- 你可以在自己的 Notion 建一份「這個字常被聽成那個」的對照表。每整理完一份稿，新確認的更正就會存回去，下一次同類主題的稿子就更好整理。 You can keep a "this-gets-heard-as-that" list in your own Notion. After each job, newly confirmed corrections are saved back, so the next transcript on a similar topic is even easier.
④ 專攻繁體中文＋中英夾雜的醫學內容
④ Built for Traditional Chinese medical content mixed with English
- 醫學演講常是中文講解、英文術語滿天飛，這正是一般轉錄軟體最容易翻車的情境，也是這個 skill 最擅長的主場。 Medical talks are often delivered in Chinese with English terms flying everywhere — exactly where ordinary transcription apps stumble hardest, and exactly this skill's home turf.
⑤ 完全不用寫程式、不用架伺服器
⑤ No coding, no server
- 它就是一份給 Claude 讀的說明檔。放進 skills 資料夾就能用，沒有安裝依賴、沒有 GPU、沒有部署。 It's just an instruction file for Claude. Drop it into your skills folder and it works — no dependencies, no GPU, no deployment.
2026-07（最近一次更新）
2026-07 (latest update)
- 📄 README 依新格式重寫：改為「簡介 → 為什麼選它 → 最新更新 → 快速開始 → 使用 → 目錄 → 貢獻 → 授權」8 段結構，並重新調查 GitHub 同類專案、更新賣點比較。 📄 README rewritten to the new format: reordered into 8 sections (About → Why → What's New → Getting Started → Usage → Structure → Contributing → License), with a fresh survey of similar GitHub projects backing the selling points.
- 🛡️ 詞庫防呆強化：只有「你親自點頭確認」或「查文獻證實」的更正，才會被存回詞庫；單純多份稿一致不算數，避免把「一起聽錯」的內容當成正確答案存起來。 🛡️ Safer glossary: only corrections you've personally confirmed (or verified against the literature) get saved back — multiple transcripts merely agreeing is not enough, so a shared mistake never gets saved as truth.
- 🔁 多稿比對：支援一次比對多份逐字稿（2–4 份最佳），用「多數決」找出真正的疑點。 🔁 Multi-transcript compare: compare several transcripts at once (2–4 is ideal) and let the majority decide the tricky spots.
- 🌐 沒有 Notion 也能用：連不上詞庫時會自動略過、繼續整理，不會卡住。 🌐 Works without Notion: if the glossary can't be reached, it just skips that step and carries on instead of getting stuck.
- ⚖️ 加入 MIT 授權：正式成為可自由使用的開源專案。 ⚖️ MIT license added: now officially a free-to-use open-source project.
請照下面的步驟安裝。
Follow these steps to install.
# 1. 下載專案 / Download the project
git clone https://github.com/kau10082/transcript-fix.git
# 2. 把整個資料夾放到 Claude 的 skills 位置
# Move the whole folder into your Claude skills location
# 例如 / e.g. /skills/user/transcript-fix/
# 3. （可選）設定你自己的詞庫 / (Optional) set up your own glossary
cp config.example.json config.local.json
# 接著打開 config.local.json，填入你的 Notion ID
# then open config.local.json and fill in your Notion IDs關於 Notion（可選）：詞庫功能需要在 Claude 裡連上 Notion 連接器（在 Claude App 的「Connectors／連接器」設定裡按 Notion → 授權即可，不需要 API key）。不想用詞庫也沒關係，skill 照樣能運作。
About Notion (optional): the glossary needs the Notion connector in Claude (open Connectors in the Claude app, click Notion, and authorize — no API key needed). Don't want the glossary? That's fine, the skill still works without it.
安裝好之後，使用方法非常簡單：在 Claude 對話中丟給它一份醫學逐字稿，然後用白話講出你的需求就好。
Once installed, using it is very simple: in a Claude conversation, give it a medical transcript and just say what you want in plain words.
可以這樣說（任一句都會觸發）：
You can say things like (any of these will trigger it):
- 「幫我校稿這份逐字稿」／ "Help me proofread this transcript"
- 「把這份演講的逐字稿整理乾淨」／ "Clean up this lecture transcript"
- 「幫我做語意重建」／ "Do a semantic reconstruction for me"
- 「這是同一場的另一份逐字稿，幫我對照」／ "Here's another transcript of the same talk — compare them for me"
小訣竅：如果你有同一場演講的多份逐字稿（例如不同軟體各轉一次），請一起丟進來 —— 比對之後準確度會大幅提升，2–4 份效果最好。
Tip: if you have multiple transcripts of the same talk (e.g. from different transcription apps), hand them all in together — cross-comparing them boosts accuracy a lot, and 2–4 copies work best.
整理完你會拿到兩份檔案：一份是清乾淨的逐字稿，另一份是「還需要再確認的事項」清單。
When it's done you'll get two files: a cleaned-up transcript, and a checklist of things still worth double-checking.
transcript-fix/
├── SKILL.md # 核心：skill 的完整流程與規則 / The skill's full workflow and rules
├── README.md # 就是你正在讀的這份說明 / The document you're reading now
├── config.example.json # 設定範本，供你複製 / A config template for you to copy
├── config.local.json # 你的私人 Notion 設定（本機、不會上傳）／ Your private Notion settings (local, never uploaded)
├── .gitignore # 指定哪些檔案不上傳 / Tells git which files to keep out
└── LICENSE # 授權條款 / The license
各檔案的作用，白話說明如下：
Here's what each file does, in plain terms:
SKILL.md：整個 skill 的大腦，寫著它該怎麼一步步整理稿子。這是最重要的檔案。SKILL.md: the brain of the skill — it spells out exactly how to clean up a transcript step by step. This is the most important file.config.example.json：設定的「範本」。你把它複製成config.local.json再填入自己的資料。config.example.json: a settings template. You copy it toconfig.local.jsonand fill in your own details.config.local.json：你的私人設定（Notion 詞庫位置）。它只留在你電腦上，永遠不會被上傳，所以你的資料很安全。config.local.json: your private settings (where your Notion glossary is). It stays on your computer and is never uploaded, so your data stays safe.
非常歡迎你一起讓這個專案變得更好！不管是修正錯字、改善流程，還是回報問題，都很感謝。
Contributions are very welcome — thank you for helping make this project better! Whether it's fixing a typo, improving the workflow, or reporting a problem, it all helps.
想改程式碼（送出修改）：
Want to change something (submit an edit):
- Fork 這個專案到你自己的帳號。／ Fork this project to your own account.
- 建一個新的分支來做修改。／ Create a new branch for your changes.
- 改好之後，發一個 Pull Request（PR） 過來，簡單說明你改了什麼。／ When you're done, open a Pull Request (PR) and briefly describe what you changed.
只是想回報問題或提想法：
Just want to report a problem or share an idea:
直接到 Issues 頁面 開一則 Issue，用白話描述你遇到的狀況就可以了。
Just open an issue on the Issues page and describe what you ran into in plain words.
本專案採用 MIT 授權。簡單來說：你可以自由使用、修改、分享，甚至用在商業用途，唯一的條件是保留原本的授權與版權聲明（也就是標示出處）。
This project uses the MIT License. In short: you're free to use, modify, share, and even use it commercially — the only condition is that you keep the original license and copyright notice (i.e. credit the source).
完整條款請見 LICENSE 檔案。
See the LICENSE file for the full terms.
由 Claude 協作打造。 / Built with Claude.