transcript_extraction_dev

Extract a conversation in text from an audio file. This project attempts to label speakers and evaluates two different approaches for performance comparison.

Repository Structure

• src/: Core Python scripts for transcription and annotation.
  • transcribe.py: Approach 1 (Whisper + simple-diarizer + LLM post-processing).
  • transcribe_v2.py: Approach 2 (WhisperX).
  • annotate_interview.py: LLM-based speaker naming utility.
• tests/: Test scripts (e.g., Langfuse connection test).
• data/: (Gitignored) All audio files, intermediate data, and final transcripts.
  • source/: Input audio files (.aac).
  • approach_1/:
    • working/: Intermediate files and checkpoints for Approach 1.
    • output/: Final annotated results for Approach 1.
  • approach_2/:
    • output/: Final results for Approach 2.
• docker-compose.yml: Local Langfuse setup.

Setup

1. Install dependencies:

   poetry install

2. Configuration: Copy .env.example (if provided) to .env and set your API keys:
   • ANTHROPIC_API_KEY: For LLM processing.
   • LF_SKEY, LF_PKEY, LF_HOST: For Langfuse tracing.
3. Ffmpeg: Ensure ffmpeg is installed on your system.

Usage

Place your audio files in data/source/ and run the scripts from the project root:

# Run Approach 1
python src/transcribe.py

# Run Approach 2
python src/transcribe_v2.py

Results will be generated in their respective subdirectories under data/.