Real-time Fallacy Detection

Overview

This project aims to perform real-time fallacy detection during events like presidential debates. It uses the Whisper for audio transcription. For natural language understanding and fallacy classification you have the option to use the OpenAI ChatGPT API or a local LLM through the text-generation-webui. I was able to run both whisper with the IconicAI_NeuralHermes-2.5-Mistral-7B-exl2-5bpw on a laptop with 3080TI 16GB of VRAM. Watch Video

Features

Real-time Audio Transcription: Uses OpenAI's Whisper ASR model for accurate real-time transcription locally or through API access.
Fallacy Detection: Utilizes OpenAI's ChatGPT to classify and identify fallacies in real-time.
Overlay Display: Provides a transparent overlay display to show both the transcription and fallacy detection results.
Text analysis: using GPT-3/4 or local LLM.

Dependencies

Anaconda
PyQt5 - Overlay
PyAudio - Audio processing
openai-whisper - ASR
openai - ChatGPT API
torch with cuda, for real time transtriction
Have the text-generation-webui running with the API flag

Installation

I build the application using Anaconda with python 3.9 on windows

Clone the repository:

git clone https://github.com/latent-variable/Real_time_fallacy_detection.git

Navigate to the project directory:
```
cd real-time-fallacy-detection
```
Create a conda environment:
```
conda create --name rtfd python=3.9
```
Activate the created environment:
```
conda activate rtfd
```
Install the required packages:
```
pip install -r requirements.txt
```
(Optional) Install the required packages to run whisper locally:
```
pip install -r requirements_local_whisper.txt
```
Installing VB-Audio, to forward audio ouput as an input device (*Optional, but I don't know how to redirect audio otherwise)

Usage

Run the main script to start the application:

python main.py 
optional arguments:
  -h, --help     show this help message and exit
  --auto         Automatically get commentary
  --api_whisper  Run whisper through api instead of locally
  --api_gpt      Will use use gpt api otherwise, by defualt will use a local LLM through the text-generation-webui API

If you plan to use your local LLM, please have the text-generation-webui running with the --api flag

Note: The application routes the audio to VB-AUDIO for processing and then redirects it back to the user for playback.

Arguments

--use_gpt(3/4): Use this flag to toggle between ChatGPT with or without GPT4. Default is to use local LLM.

Display

The application will display an overlay with two sections:

Top Box: Displays the fallacy classification from ChatGPT.
Bottom Box: Displays the real-time transcription from the Whisper API.

Press the Esc key to close the application.

Configuration

[Audio Settings] You can configure the audio input and outsource in the settings.ini.

device_input_name = VB-Audio <- must have
device_output_name = Headphones (2- Razer <- replace with your own *Note: when the application loads if will redirect the audio back to the output device

[Local LLM Settings]

instruction_template = mistral-openorca <- replace with the model specific template *Note: this is a custom template, which you will likely note have in your text-generation-webui

Rename the api_key.txt.template to api_key.txt and add your OpenAI API key to it.

License

This project is licensed under the MIT License.

sagi/Real_time_fallacy_detection