Purpose

This project offers a privacy-focused solution for transcribing and summarizing audio recordings through entirely local processing. Using OpenAI's Whisper for transcription and local LLMs via Ollama for summarization, it processes audio files (MP3/WAV) entirely on your machine, ensuring sensitive content never leaves your environment.

The tool automatically generates structured summaries including:

• Executive overview
• Detailed content breakdown
• Action items
• Meeting metadata

The project uses a configuration-based approach (config.yaml) for easy customization of output formats, model parameters, and summary structures, making it adaptable for various meeting types and organizational needs

---

Setup

Select Python Interpreter Version Between 3.8-3.11

• I am using Python 3.10.11

Install ffmpeg Globally as PowerShell Administrator

NOTE: ffmpeg DOES NOT work in virutal environment and is required for Whisper to work. A "[Win2]File not found error" populates when attempting to use within a virtual environment, although it is not best practice, utilize your global environment instead!

• Follow the instructions HERE and install choclatey install via PowerShell Administration to install ffmpeg.

• Install FFmpeg:

    choco install ffmpeg

Requirements Installation

python3.X -m pip install -r requirements.txt --no-warn-script-location

Enable Long Paths

• From PowerShell Administrator run the following:

New-ItemProperty -Path "HKLM:\SYSTEM\CurrentControlSet\Control\FileSystem" -Name "LongPathsEnabled" -Value 1 -PropertyType DWORD -Force

Download PyTorch with CUDA Support for GPU Acceleration

• If you have NVIDIA GPUs, determine what compute platform you have present:

  nvidia-smi.exe

• Identify "CUDA Version"
• Navigate to: https://pytorch.org/get-started/locally/
• Select options specific to your environment and install the command specified!
• Once installation is complete run:

  python3.X pytorch_verify.py

Example Successfull Output:

• True
• NVIDIA GeForce RTX 3080 Ti Laptop GPU

---

Usage

1. transcribe.py (user input) or transcribe-args.py (commandline arguments)

Simple User Input

python3.X transcribe.py

Commandline Input

Single File

python3.X transcribe-args.py --mode single --input-file ./path/to/audio.mp3 --output-dir ./path/to/output --model large

Multiple Files

python3.X transcribe-args.py --mode multiple --input-dir ./path/to/audio/files --output-dir ./path/to/output --model base

• -h or --help: To see all available options and examples

• --mode: Choose between 'single' or 'multiple' REQUIRED

  • single: Transcribe one audio file
  • multiple: Transcribe all audio files in a directory

• --input-file: Path to the audio file REQUIRED

• --input-dir: Path to directory containing audio files REQUIRED

• --output-dir: Directory where transcription files will be saved REQUIRED

• --model: Choose Whisper model size (If not specified, the script uses the 'base' model by default.) OPTIONAL

  • tiny: Fastest, lowest accuracy
  • base: Good balance of speed and accuracy
  • small: Better accuracy, slower than base
  • medium: High accuracy, slower
  • large: Highest accuracy, slowest

---

2. summarize.py

Configurations

1. Download Ollama and run your chosen model, follow the process HERE.
2. Configure and update the REQUIRED settings in the config.yaml

python3.X summarize.py

Example config.yaml

llm: 
  model_name: "YOUR_MODEL_NAME" # REQUIRED
  max_retries: 3 # OPTIONAL
  retry_delay: 2 # OPTIONAL
  api_url: "http://localhost:11434/api/generate"
  options:
    temperature: 0.7 # OPTIONAL
    top_p: 0.9 # OPTIONAL
    max_tokens: 8000 # OPTIONAL

output:
  format: "md" # OPTIONAL
  log_file: "./data/transcript_processor.log" # REQUIRED

paths: 
  input_transcript: "./data/transcriptions/" # REQUIRED
  output_directory: "./data/meeting summaries" # REQUIRED
  audio_file: "./data/converted-audio/" # REQUIRED
  
Shortened for Space...

LLM Config Settings

• model_name: Choose your Ollama model
• max_retries: Number of API call attempts
• retry_delay: Delay between retries
• temperature: Controls response creativity, higher = more creative, lower = more concise (0.0-1.0)
• top_p: controls similarity sampling (accuracy) when generating a response (0.1-1)
• max_tokens: Maximum response length

Path Configuration

• input_transcript: Path to your transcript file
• output_directory: Directory for generated summaries

Output Settings

• format: Output format ('md' or 'txt')
• log_file: Location of log file

Prompt Templates

• executive_summary
• detailed_summary
• action_items

---

Troubleshooting

Ollama Connection

• Navigate to http://localhost:11434 to ensure Ollama is running.

• From terminal input the following to identify and ensure your model is downloaded successfully:

  ollama list

  ollama run YOUR_MODEL

---

TO-DO

Updating README

• Include the use of audio-converter.py at the beginning to ensure that audio conversion is in either .mp3 or .wav format.
• Separate audio-converter.py
• Make audio-converter.py it's own project and just have the current project reference it to be used...

Command Line Overrides

Incorporation of commandline overrides should they become a repetitive occurence.

# Future implementation in summarize.py
import argparse

def parse_args():
    """Parse command line arguments for path overrides."""
    parser = argparse.ArgumentParser(description='Transcript Summarization Tool')
    parser.add_argument(
        '--input-file',
        help='Override input transcript file path from config.yaml'
    )
    parser.add_argument(
        '--output-dir',
        help='Override output directory path from config.yaml'
    )
    return parser.parse_args()

Override Input File and Output Directory

python3.X summarize.py --input-file ./new/transcript.txt --output-dir ./new/output

Config.yaml Command Line Arguments

• config.yaml should take command-line arguments rather than directy file manipulation
• User should be able to specify the title of the summary based off of the original audio file name when placed inside of "meeting summaries" so that there is clarity.

Stream-line the Entire Process

• GOAL: Start with audio conversion (if needed) > transciption > summarization all with one command...