TheFloatingString / AGORA

Automated Generation and Omission Recurrent Architecture (AGORA). This model inputs speech (audio recording) and replaces hate speech and profanity with generated textual content. (Speech to text model.) McGill's submission to Project X, 2022-23.
0 stars 0 forks source link
generative-model hate-speech-detection speech-to-text

AGORA - Automated Generation and Omission Recurrent Architecture

Given a speech input (audio recording), this model replaces harmful speech with generated textual content. (Speech to text model.)

Installation and Setup

Configure environment variables:

set OPENAI_API_KEY_AGORA=<API_KEY>

Setup

git clone https://github.com/TheFloatingString/agora.git
cd agora
pip install -r requirements.txt

In a Python file:

from src.agora import Agora

agora_model = Agora()
response = agora_model.transcribe_audio("filepath_to_speech_audio.wav")
print(response["outputText"])

Quickstart examples

python -m quickstart.run_sample

Analyze AGORA's ability to Recognize Offensive Content in the Jigsaw Dataset

Note: move train.csv into data/jigsaw-data from the Jigsaw dataset on Kaggle (https://www.kaggle.com/competitions/jigsaw-toxic-comment-classification-challenge/data)

python -m src.run_jigsaw_data
python -m src.analyze_results

Filter and Paraphrase the Speech-to-Text Functionality for Offensive Content

Run the folowing, while making sure to change the filename from 1 to 10 at each new run.

Warning: the audio files contain explicit content.

python -m src.run_audio_files