TranscriberApp Svelte Themes

Transcriberapp

Local app for private transcription & analysis of audio with Whisper, Pyannote & Ollama.

#audio #diarization #flask #local-llm #ollama #pyannote #svelte #transcription #whisper
Demo Download

TranscriberApp

A local web application for transcribing and analyzing audio files using local AI models (FasterWhisper, Pyannote, Ollama). Focuses on privacy and performance, with optimizations for Apple Silicon.

Key Features

  • Transcription: Audio-to-text using FasterWhisper.
  • Speaker Diarization: Identifies speakers using Pyannote.audio.
  • Speaker Name Detection: (Optional) LLM-based name proposal.
  • User Review: UI step to confirm/edit speaker names before analysis.
  • LLM Analysis: 'Fast' (summary) or 'Advanced' (multi-task) modes via local Ollama.
  • HTML Output: Formatted transcript generation.
  • Database Logging: Saves job results to SQLite.
  • Local & Private: All core processing runs locally.
  • Apple Silicon Optimized: Uses MPS acceleration.

Technology Stack

  • Backend: Python (3.11/3.12 recommended), Flask
  • Transcription: FasterWhisper
  • Diarization: Pyannote.audio
  • LLM Interaction: Ollama
  • Database: SQLAlchemy Core, SQLite
  • Audio Handling: Pydub
  • Frontend: Svelte, Vite, Tailwind CSS
  • Code Quality: Ruff

Demo / Screenshot

Requirements

  • OS: macOS (Apple Silicon recommended) or Linux. Windows might require adjustments.

  • Python: Python 3.11 or 3.12 strongly recommended. (Version 3.13 currently has known compatibility issues with dependencies like pydub).

  • Node.js: Required for the frontend (includes npm). Version 18+ recommended.

  • Ollama: Installed and running (ollama.com). Ensure required models are pulled (see Installation step 7).

  • Hugging Face Account & Token:

    • Account needed to accept model terms on Hugging Face (especially for Pyannote).
    • Read Access Token required (store in .env).

  • System Dependencies:

    • macOS (via Homebrew): Ensure Homebrew is installed. Then run:
      brew install ffmpeg cmake pkg-config protobuf
    • Debian/Ubuntu Linux:
      sudo apt update && sudo apt install -y ffmpeg cmake pkg-config libprotobuf-dev protobuf-compiler
    • (Other Linux/Windows: Install equivalent packages for ffmpeg, cmake, pkg-config, and protobuf development libraries).

  • Python Packages: Listed in requirements.txt.

Installation

  1. Clone the repository:
    git clone [https://github.com/GPTSam/TranscriberApp.git](https://github.com/GPTSam/TranscriberApp.git)
cd TranscriberApp
  2. Create Python Virtual Environment: (Use Python 3.11 or 3.12)
    # Replace python3.11 with your specific command if needed
python3.11 -m venv venv
  3. Activate Virtual Environment:
    source venv/bin/activate
# (Windows: venv\Scripts\activate)
    (Your terminal prompt should now start with (venv))
  4. Install System Dependencies: (If not already installed - see Requirements) (macOS example: brew install ffmpeg cmake pkg-config protobuf)
  5. Install Python Dependencies:
    # Use the correct Python interpreter for make if needed
make PYTHON_INTERPRETER=python3.11 install
# Or directly with pip:
# pip install --upgrade pip
# pip install -r requirements.txt
  6. Set up Hugging Face Token:
    • Accept model terms on HF (e.g., pyannote/speaker-diarization-3.1).
    • Create a read Access Token in your HF settings.
    • Create .env file: cp .env.example .env
    • Edit .env and add your token: HUGGING_FACE_TOKEN=hf_YOUR_ACTUAL_TOKEN_HERE
  7. Download Ollama Models: Pull the LLM models specified in config.yaml (or the defaults if you haven't edited it yet). At minimum for testing:
    ollama pull llama3:8b
ollama pull mistral:7b # Often used for name detection/summary fallback
# Pull others listed in config.yaml -> llm_models as needed
  8. Generate/Check Configuration:
    • Ensure config_schema.yaml exists and has content.
    • Generate the default config.yaml if it doesn't exist:
      # Run from project root, with venv active
python -m src.utils.generate_config_from_schema
    • Edit config.yaml: Open config.yaml and set input_audio to a valid relative path for an audio file in the audio/ directory (e.g., audio/sample.mp3). Review other settings like whisper_model and llm_models for your local setup.

Usage

Important: Both the Backend and Frontend servers need to be running simultaneously for the Web UI.

1. Run the Backend (Flask API Server):

  • Open Terminal 1.
  • Navigate to the project root (TranscriberApp).
  • Activate the Python venv: source venv/bin/activate
  • Start the server:
    # Replace python3.11 if you used a different version for venv
make PYTHON_INTERPRETER=python3.11 run-web
# Or directly: python app.py
  • Leave this terminal running. It serves the API on http://localhost:5001.

2. Run the Frontend (Svelte Dev Server):

  • Open Terminal 2.
  • Navigate to the frontend directory: cd frontend
  • Install Node dependencies (only needed once): npm install
  • Start the dev server:
    npm run dev
  • Leave this terminal running. It serves the UI, usually on http://localhost:5173.

3. Use the Web Interface:

  • Open your web browser and navigate to the frontend URL (e.g., http://localhost:5173).
  • Workflow:
    1. Use "Upload Audio File" to select and upload your audio. Wait for success message.
    2. Adjust settings in the "Configure Pipeline" section if desired.
    3. Click "Start Pipeline".
    4. Monitor the "Job Status" section for progress updates and logs.
    5. When the status becomes WAITING_FOR_REVIEW, the Review Dialog will appear.
    6. Review the transcript context, assign/correct speaker names, and click "Confirm Names".
    7. The pipeline (Part 2) will continue. Monitor status until COMPLETED.
    8. Download links for results will appear in the status section.
    9. Use "Stop Pipeline" to attempt cancellation or "Reset" to clear the state for a new run.

4. Command Line Interface (Testing/Scripting):

  • Runs the full pipeline non-interactively. Review step is automatic.
  • Ensure venv is active in your terminal.
  • Examples (run from project root):
    # Use config.yaml settings (ensure input_audio is set)
make PYTHON_INTERPRETER=python3.11 run-cli

# Override input and run in advanced mode
make PYTHON_INTERPRETER=python3.11 run-cli ARGS="--input-audio audio/my_audio.m4a --mode advanced"

Output Files & Data

  • logs/: Daily log files (YYYY-MM-DD.log).
  • transcripts/intermediate_transcript.json: Raw diarized transcript.
  • transcripts/intermediate_proposed_map.json: (Optional) LLM proposed map.
  • transcripts/intermediate_context.json: (Optional) Context for name detection.
  • transcripts/final_transcript.json: Transcript with final names.
  • results/transcript.html: Formatted HTML transcript.
  • results/summary.txt: 'Fast' mode output.
  • results/advanced_analysis.json: 'Advanced' mode output.
  • llm_training_data.db: SQLite DB with job results.

Speaker Handling Process

  1. Diarization -> Speaker IDs (SPEAKER_00).
  2. Transcription -> Text segments.
  3. Merge -> Assign ID to text.
  4. (Optional) Name Detection -> Propose names.
  5. User Review -> Confirm/edit names.
  6. Final Mapping -> Apply names.

Database Logging

Job results logged to llm_training_data.db.

Troubleshooting

(Expanded based on our findings)

  • Pyannote/HF Token Errors: Accept model terms on HF, check .env token.
  • Ollama Errors: Ensure service running, models pulled. Check Ollama server logs.
  • Port Conflicts (5001): Stop process using port (check with lsof -i :5001), change FLASK_PORT.
  • Python ModuleNotFoundError: Ensure venv active, run make ... install or pip install -r requirements.txt.
  • Build Errors (sentencepiece): Ensure system dependencies (cmake, pkg-config, protobuf) installed via package manager (e.g., Homebrew). Use Python 3.11/3.12, as 3.13 has issues (audioop missing).
  • Frontend ERR_CONNECTION_REFUSED (Port 5173): Ensure frontend dev server (npm run dev) is running in the frontend directory.
  • Frontend API Errors (500 / SyntaxError): Ensure backend server (make ... run-web) is running. Check backend terminal logs for tracebacks. Verify API proxy in vite.config.js targets the correct backend port (5001).
  • NPM EACCES Cache Errors: Run the sudo chown -R ... command suggested by npm to fix cache permissions.
  • Performance: Check logs for MPS/CUDA usage. Use smaller models / int8.

Future Enhancements

  • Implement transcript text editing.
  • Add more analysis tasks.
  • Improve UI feedback/visualization.

License

MIT License - see LICENSE file.

Author / Contact

Samuel Willems / Legendaddy / willems.samuel@gmail.com

Top categories

tailwind daisyui admin template popup mdsvex portfolio blog form ecommerce ui carousel auth dark seo image routing
Loading Svelte Themes