案例:AceSpeller 點樣用 AI 將學校默書表相片變成練習
簡單答:家長用 AceSpeller 影一張學校默書表,App 會先裁切相片、讀出上面嘅文字,再自動分開「生字」同「句子」,家長確認或者修改之後,就可以用粵語或普通話讀出嚟俾小朋友聽寫。AceSpeller 默書100分係 MuduoPro 自己開發同營運嘅 App,以下係我哋做呢個功能嘅真實經過。
要解決嘅問題
香港小朋友每星期都有默書,學校會派一張默書表。家長要逐個字讀俾小朋友聽,又要讀得準,好花時間。我哋想做到:影一張相,就有一份可以重複聽寫嘅練習。
成個流程點做
- 影相同裁切:家長用相機或者相簿揀相,先裁走默書表以外嘅部分,可以一次揀幾張。
- 讀出文字:App 有兩種模式:標準模式喺手機本機讀取文字;高準確度模式用雲端 AI 讀取,適合比較難讀嘅相。
- 分開生字同句子:按規則自動分類,例如夠長或者有句號、問號嘅當句子,其餘當生字。
- 人手確認:家長喺開始前睇一次,改錯字或者刪走唔需要嘅內容。
- 讀出嚟練習:用粵語或普通話讀出,小朋友跟住寫。
遇過嘅問題同解決方法
1. 處理相片時 App 凍結:原本圖片前處理(例如縮細、調整)喺 App 主線程做,喺 iPhone 12 等機型處理大相時,畫面可以冇反應長達 15 秒。我哋將前處理搬去背景執行,畫面唔再凍結。
2. 相片太大反而慢:相機相片解像度好高,但讀文字唔需要咁大。我哋先將相片縮細到適合讀取文字嘅尺寸,處理快咗,準確度亦唔受影響。
3. AI 唔係百分百準:手寫字、反光、摺痕都會令讀取出錯。所以我哋冇假設結果一定啱,而係設計咗一個確認步驟,俾家長開始練習之前睇一次、改一次。
4. 唔同嘅分隔方式:有啲默書表每行一個詞,有啲用逗號或頓號分隔。我哋先按換行分,只有一行嘅時候先用逗號、頓號、分號分,減少分錯。
呢個案例教咗我哋咩
- 先處理資料,再諗 AI:裁切、縮細、分隔規則呢啲「唔係 AI」嘅步驟,對結果影響好大。
- 一定要有人確認:AI 讀錯係正常,設計好確認步驟,比追求百分百準實際。
- 留意用家感受:結果幾準都好,App 凍結 15 秒,用家已經會走。
- 按情況揀方法:本機處理快、唔使上網;雲端 AI 準確度高,處理難啲嘅相。兩種都提供,由用家揀。
呢幾點同我哋幫公司做嘅項目一模一樣,例如《收據影相自動入數》,就係將「影相、讀出、分類、確認」用落公司單據。
自己做最常卡喺邊?
- 相片質素參差,讀出嚟嘅文字錯得多,唔知點樣改善
- 讀出文字之後唔知點樣自動分類,要人手逐行執
- 處理大量相片時程式好慢,甚至凍結
- 唔知幾時應該用本機處理、幾時應該用雲端 AI
以上任何一樣卡住咗,都可以搵我哋。簽約前免費諮詢,我哋會直接話你知要做乜、大概幾錢。
我哋可以點幫你
AceSpeller 由設計、開發、上架到持續更新都係我哋自己做。同樣嘅經驗,我哋用嚟幫公司做 AI 流程自動化(單一流程 HK$5,000 起,最快 2 日交付)、AI 工具/程式開發(HK$8,000 起),同手機 App 開發(可撳原型 HK$5,000 起、標準 App HK$30,000 起)。想了解 AceSpeller,可以去我哋嘅產品頁。所有套餐可以睇套餐同價錢,或者直接同我哋傾一次,簽約前免費諮詢。
Case study: how AceSpeller turns a photo of a school dictation list into practice
The short answer: a parent photographs the school's dictation list in AceSpeller. The app crops the photo, reads the text, and automatically separates vocabulary from sentences. Once the parent has checked or corrected it, the app reads the list aloud in Cantonese or Mandarin for the child to write down. AceSpeller is an app MuduoPro builds and runs itself; this is the real story of how we built the feature.
The problem
Hong Kong primary pupils have dictation every week, and the school hands out a list. Parents have to read each word aloud, accurately, which takes time. We wanted one photo to become practice a child can repeat as often as needed.
How the workflow runs
- Photograph and crop: the parent picks a photo from the camera or album and crops away everything outside the list; several photos can be chosen at once.
- Read the text: the app has two modes. Standard mode reads the text on the phone itself; high-accuracy mode uses cloud AI, for photos that are harder to read.
- Separate vocabulary and sentences: rules sort the items automatically; for example, longer items or those with a full stop or question mark are treated as sentences, the rest as vocabulary.
- Check by hand: the parent reviews the list before starting, fixing mistakes or removing what is not needed.
- Practise aloud: the app reads each item in Cantonese or Mandarin and the child writes it down.
Problems we hit, and how we fixed them
1. The app froze while processing photos: image pre-processing (such as resizing and adjustments) originally ran on the app's main thread. On phones like the iPhone 12, processing a large photo could leave the screen unresponsive for up to 15 seconds. We moved pre-processing to the background, and the freezing stopped.
2. Large photos were slower, not better: phone photos have very high resolution, but reading text does not need it. We resize photos to a size suited to text recognition first, which speeds things up without hurting accuracy.
3. AI is not 100% accurate: handwriting, glare and folds all cause mistakes. Rather than assume the result is always right, we designed a check step so parents can review and correct the list before practice starts.
4. Different separators: some lists have one word per line; others use commas or enumeration commas. We split by line first, and only use commas and semicolons when there is a single line, which reduces mis-splitting.
What this case taught us
- Prepare the data before thinking about AI: cropping, resizing and splitting rules are not AI, but they matter enormously to the result.
- Always have a person check: AI mistakes are normal; a good check step is more practical than chasing 100% accuracy.
- Mind the user experience: however accurate the result, users leave if the app freezes for 15 seconds.
- Choose the method to fit: on-device processing is fast and works offline; cloud AI is more accurate on harder photos. We offer both and let users choose.
These are exactly the same lessons as in the projects we build for companies, such as Turning receipt photos into accounts automatically, which applies "photograph, read, sort, check" to business receipts.
Where do people usually get stuck?
- Photo quality varies, the extracted text has many errors, and you do not know how to improve it
- Not knowing how to sort extracted text automatically, so every line is tidied by hand
- The program slows down or freezes when processing many photos
- Not knowing when to process on the device and when to use cloud AI
If any of these is where you are stuck, talk to us. Consultation before signing is free, and we will tell you plainly what needs doing and roughly what it costs.
How we can help
We design, build, publish and update AceSpeller ourselves. We bring the same experience to companies through AI workflow automation (single workflows from HK$5,000, delivered in as little as 2 days), AI tools and software (from HK$8,000) and mobile app development (clickable prototypes from HK$5,000, standard apps from HK$30,000). To learn more about AceSpeller, see our products page. See all packages and pricing, or talk to us: consultation before signing is free.