○

The future is here.

MARS utilizes state-of-the-art Language Model capabilities to seamlessly integrate Large Language Model (LLM), Text-to-Speech (TTS), and Speech-to-Text (STT) technologies. Our mission is to provide an exceptional voice AI assistant experience with fast inference speed, delivering natural and intelligent interactions akin to a real person.

---

Table of Contents

• Project Overview
• Project Structure
• Key Components
• Features
• Usage
• Future Development
• License

Project Overview

MARS is an advanced voice AI assistant developed by a collective of computer science students from EMA EMITS College Philippines. It leverages cutting-edge technologies including Large Language Models (LLMs), Text-to-Speech (TTS), and Speech-to-Text (STT) to provide a seamless and intelligent conversational experience.

Project Structure

The project is organized as follows:

• App: Main Flask application for processing user requests and responses.
• Fine-Tune: Training data and scripts for LLM fine-tuning.
• Knowledge Base: Text files containing relevant information for the AI assistant.
• Module: System prompts and instructions for the AI assistant.
• Static: Static assets (images, SVGs).
• Web: HTML files for the user interface.

Key Components

1. App.py

• Inputs: User requests (text or audio)
• Processing:
  • Audio transcription (OpenAI Whisper API)
  • Context retrieval (Langchain for RAG)
  • Response generation (Fine-tuned OpenAI GPT-4o model)
  • Text-to-speech synthesis (OpenAI TTS API)
• Storage: Supabase bucket for audio file management
• Outputs: Text and audio responses

2. Fine-Tune Files

• fine_tuning_dataset.jsonl: General and MARS-specific training data
• main_training_data.jsonl: General training data
• mars-v1.jsonl: MARS-specific training data
• token_counter.py: Token counting utility

3. Knowledge Base

• eecp.txt: Information about EMA EMITS College Philippines
• mispronoucation.txt: Instructions for handling user mispronunciations

4. Module

• prompt.py: General AI assistant instructions
• system_prompt.py: System-level prompts and capabilities
• system_prompt.txt: System prompt text file

Features

• Text-to-Speech (TTS): Natural speech synthesis using OpenAI's TTS API
• Speech-to-Text (STT): Accurate transcription via OpenAI's Whisper API
• Large Language Model (LLM): Custom fine-tuned OpenAI GPT-4o model
• Retrieval-Augmented Generation (RAG): Context-aware responses using Langchain
• Audio File Management: Efficient serving through Supabase bucket storage

Usage

MARS offers a web interface for user interaction, with potential for expansion to other platforms. It can understand, process, and respond to a wide range of prompts, including:

• Factual questions
• Information requests
• Commands
• Natural conversational interactions

Future Development

MARS is in active development, with planned enhancements including:

• Expanded knowledge base and domain expertise
• Advanced language understanding and generation
• Integration with additional AI technologies
• Enhanced user interface and interaction mechanisms

---

Note: MARS is currently in its developmental phase, with ongoing refinement and enhancement efforts.

View Conceptual Model

MARS is licensed under the terms of the MIT License.

© 2024 MARS. All rights reserved.