SpeechFlowGuard

A machine learning web API that detects toxic language in user comments using classical ML models (TF-IDF + Logistic Regression). Built with FastAPI, trained on the Jigsaw Toxic Comment Classification Challenge dataset.

✅ Features

• Multi-label classification:
  • toxic, severe_toxic, obscene, threat, insult, identity_hate
• Real-time REST API (FastAPI)
• Modular codebase
• Dockerized for portability
• Preprocessed with custom regex cleaner

🧪 Model Details

• Vectorizer: TfidfVectorizer (max_features=4096, stop_words='english')
• Classifier: LogisticRegression (class_weight='balanced', max_iter=500, C=1.6)
• Trained on: Jigsaw Toxic Comment Dataset

🗂️ Project Structure

SpeechFlowGuard/
├── app/
│   ├── main.py
│   ├── api.py
│   ├── model.py
│   ├── schemas.py
│   ├── utils.py
│   └── config.py
├── data/
│   ├── data_processed.csv
│   └── trains.csv
├── models/
│   ├── tf-idf_vectorizer.pkl
│   └── classifier.pkl
├── notebooks/
│   ├── data_cleaning.ipynb
│   └── tf-idf_model_train.ipynb
├── .gitignore
├── docker-requirements.txt
├── Dockerfile
├── README.md
└── requirements.txt

🧰 Technical Stack

• Language: Python 3.12+
• Framework: FastAPI (ASGI-compatible)
• ML Model:
  • TfidfVectorizer for feature extraction
  • LogisticRegression (one classifier per label, binary relevance method)
• Serialization: dill for saving sklearn models
• Request Schema: Pydantic-based input validation
• Serving: Uvicorn for ASGI serving
• Containerization: Docker

📡 API Endpoints

The FastAPI server exposes the following endpoints:

GET /

Returns a welcome message to confirm the API is live.

Request:

curl http://localhost:8000/

Response:

{
  "message": "Hello and welcome to SpeechFlowGuard API"
}

POST /predict

Performs multi-label classification on the input text and returns the predicted probabilities for each toxicity label.

Request:

POST /predict
Content-Type: application/json

Request Body:

{
  "text": "You are a criminal person"
}

Response:

{
  "toxic": 0.6774,
  "severe_toxic": 0.039,
  "obscene": 0.0994,
  "threat": 0.1204,
  "insult": 0.5151,
  "identity_hate": 0.6681
}

🛠️ Git Setup & Repository Cloning

If you haven't installed Git:

🔨 Install Git

Windows:

Download from https://git-scm.com/download/win and install with default settings.

Ubuntu/Linux:

sudo apt update
sudo apt install git

macOS:

brew install git

📦 Clone the Repository

git clone https://github.com/RohanSardar/SpeechFlowGuard.git
cd SpeechFlowGuard

🔧 How to Train the Model

Ensure you have the following installed:

• Python (≥ 3.12)
• Conda (for Conda-based setup)
• Virtualenv (install via pip install virtualenv if not already available)

🐍 Using conda

Create a conda virtual environment

Run the following command to create a virtual environment in a specific directory:

conda create -p venv python=3.12 -y

Activate it

conda activate venv/

Install dependencies

pip install -r requirements.txt

💻 Using virtualenv

Run the following command to create a virtual environment in a specific directory:

python -m virtualenv venv

Activate it

• Windows

venv\Scripts\activate

• Linux/macOS

source venv/bin/activate

Install dependencies

pip install -r requirements.txt

Use the Jupyter notebooks in notebooks/ or create a script to:

1. Load and preprocess the dataset.
2. Train TF-IDF and LogisticRegression models.
3. Save them using dill.

🐳 Docker Setup

🔥 1. Build the Image

docker build -t speechflowguard .

🚀 2. Run the Container

docker run -p 8000:8000 speechflowguard

You can also access the interactive API docs at:

• Swagger UI: http://localhost:8000/docs
• ReDoc: http://localhost:8000/redoc

Example using cURL

curl -X POST http://localhost:8000/predict \
  -H "Content-Type: application/json" \
  -d '{"text": "You are a criminal person"}'

Response

{
  "toxic": 0.6774,
  "severe_toxic": 0.039,
  "obscene": 0.0994,
  "threat": 0.1204,
  "insult": 0.5151,
  "identity_hate": 0.6681
}