Transcriber

A desktop app for transcribing Ghanaian audio files using Gemini AI.

---

Requirements

• Python 3.8 or newer
• Git (for automatic setup)
• An internet connection

Linux — install these once:

sudo apt install python3-tk xclip wl-clipboard git

Windows / Mac — Python and Git are all you need.

---

Setup

1. Get the app

git clone https://github.com/GhanaNLP/transcriber.git
cd transcriber

2. Run it

python transcriber.py

On first launch the app automatically installs everything it needs (huggingface_hub and the language validator). This takes about 30 seconds and only happens once.

A setup screen will appear asking for three things:

• Volunteer code — paste the code you were sent
• Audio folder — browse to the folder where you extracted your zip
• Output folder — choose where your transcript files will be saved

Click Start →. Your code is remembered for future sessions — you won't need to paste it again.

---

How to transcribe

1. The app shows one audio file at a time
2. Click ⎘ Copy audio file → open gemini.google.com → start a new chat → upload the file
3. Click ✦ Prompt 1 (News) or ✦ Prompt 2 (Conversational) to copy the right prompt → paste it into Gemini and send
4. Copy Gemini's response → click back on the app → paste into the text box
5. The app checks and saves automatically:
   • Repeated sentences are removed
   • Transcripts that are too short or too long are flagged
   • Transcripts that don't match your assigned language are flagged — this helps catch paste errors
6. Every 10 transcriptions your work is uploaded automatically

Which prompt?

• Prompt 1 (News) — formal news broadcast audio
• Prompt 2 (Conversational) — informal or mixed speech

If one gives a short or poor result, try the other. Still poor → click Skip ⇥.

---

Language checking

The app automatically checks that each transcript matches the language you were assigned. If it doesn't:

• Yellow warning — the match is low but could be fine (e.g. heavy English mixing). You can still save.
• Red block — the transcript looks completely wrong. Re-paste from Gemini or skip the file.

This catches accidental wrong pastes, not genuine mixed-language speech.

---

Skipping files

Click Skip ⇥ if you can't get a valid transcript after a few attempts. The file won't appear again. To undo a skip, remove its line from skipped.log in your transcripts folder.

---

Submitting

Your results upload automatically every 10 transcriptions. To upload immediately, click ⬆ Push Now. When you finish all your files the app pushes anything remaining automatically.

---

Troubleshooting

App won't open on Linux

sudo apt install python3-tk

"Assignment mismatch" on startup Make sure you selected the folder where you extracted your zip file, not a different one.

Push failed The app retries automatically next time you save. You can also click ⬆ Push Now once your connection is back.

Lost your code Contact the project coordinator to get it resent. Once entered it will be saved for future sessions.