Nebula: Introduce Nebula Transcription Service (Local Whisper + GPT Vision Integration)

[Anishyou]

Overview:

• Adds the initial version of the Nebula Transcription backend service.
• Provides full local transcription with Whisper and slide number extraction using GPT-4o Vision.

Main Features:

1. Download lecture videos from a given URL using ffmpeg.
2. Extract audio from video and convert it into .wav format.
3. Transcribe the extracted audio locally using openai-whisper (default model: base).
4. Automatically uses GPU (CUDA) if available, otherwise CPU.
5. Capture frames at transcript segment timestamps.
6. Crop the bottom 5% of frames where slide numbers are typically shown.
7. Detect slide numbers from frames using OpenAI GPT-4o Vision model.
8. Return a structured JSON response with:
   • Start time, end time, transcribed text, and detected slide number for each segment.
9. Automatically clean up temporary files after processing.

API Endpoint:
POST /start-transcribe

Input: {
  "videoUrl": "<video-link>",
  "lectureId": 123,
  "lectureUnitId": 456
 }

Output: JSON containing a list of transcribed segments with slide numbers.
example output

{
  "language": "en",
  "segments": [
    {
      "startTime": 0.0,
      "endTime": 5.2,
      "text": "Welcome to today's lecture...",
      "slideNumber": 1
    },
    ...
  ]
}

How to start the Nebula Transcriber:
Read - README.MD

[alexjoham]

The first version looks good in my opinion. I added some comments. Also consider adding some doc comments and also maybe if you could have more files to not have the whole functionality in one file. Also please have a look at the linting test as there are some failing tests

[alexjoham]

Another small suggestion: Consider working here with DTOs as you do in Artemis

[Anishyou]

needs to look into it.