FixGram 🤖 - Speech-to-Text Grammar Correction

A Streamlit application that combines Whisper AI for speech-to-text conversion and Llama 2 for grammar correction. Users can either type text directly or record audio, which will be transcribed and then grammatically corrected.

Features

• Speech-to-text conversion using OpenAI's Whisper
• Grammar correction using Llama 2
• Real-time audio recording capability
• User-friendly interface built with Streamlit
• Support for both text input and audio input

Prerequisites

Before running this application, make sure you have the following:

• Python 3.8 or higher
• Required models:
  • Llama 2 model file (llama-2-7b-chat.ggmlv3.q8_0.bin) in the models directory
  • Whisper base English model (will be downloaded automatically on first run)

Installation

1. Clone the repository:

git clone [your-repository-url]
cd [repository-name]

2. Install the required packages:

pip install streamlit whisper langchain ctransformers audiorecorder

3. Download the Llama 2 model:
   • Download llama-2-7b-chat.ggmlv3.q8_0.bin
   • Place it in the models directory of your project

Usage

1. Run the Streamlit application:

streamlit run app.py

2. Use the application:
   • Type text directly into the input field and click "Submit" for grammar correction
   • OR
   • Click "Click to record" to record audio
   • Click "Click to stop recording" when finished
   • Click "Transcribe Audio" to convert speech to text and correct grammar

Project Structure

project/
│
├── app.py              # Main application file
├── models/            # Directory for model files
│   └── llama-2-7b-chat.ggmlv3.q8_0.bin
│
└── README.md          # Project documentation

Key Components

• load_llama_model(): Initializes the Llama 2 model for grammar correction
• load_whisper_model(): Loads the Whisper model for speech-to-text conversion
• getLLammaResponse(): Processes text through the Llama 2 model for grammar correction

Cache Implementation

The application uses Streamlit's caching mechanism (@st.cache_resource) to optimize performance:

• Model loading is cached to prevent repeated loading
• Both Whisper and Llama models are loaded only once per session

Error Handling

The application includes error handling for:

• Model loading failures
• Audio processing issues
• Text input validation