/Castena

LLM chatbot to interact with Spanish podcasts

Primary LanguageJupyter Notebook

Castena: LLM Chatbot for Interacting with Spanish Podcasts

Table of Contents

Introduction

The LLM Chatbot for Interacting with Spanish Podcasts is an AI-powered chatbot designed to enhance your podcast listening experience. It is specifically tailored for Spanish-speaking podcast enthusiasts who want to engage with and learn from their favorite Spanish podcasts in a whole new way.

This chatbot utilizes advanced natural language processing techniques to provide real-time text-based interactions with Spanish podcasts. Whether you want to generate transcripts, summarize episodes, or seek information about specific topics within a podcast, this chatbot is here to assist you.

Objectives

Worldcast Wild Project Wall Street Wolverine Marc Vidal

Video Name Channel Topic Status
Worldcast #45 - Roberto Vaquero Worldcast Politics Available
Marc Vidal - entrevista a Jose Luis Cava Marc Vidal Economy and Politics Available soon
Wall Street Wolverine - entrevista a Samuel Vázquez Wall Street Wolverine Politics and Society  Available soon
The Wild Project - entrevista a Operador Nuclear The Wild Project Society  Available soon

Features

  • Transcription: Get accurate text transcriptions of podcast episodes.
  • Summarization: Receive concise summaries of podcast episodes.
  • Search: Find specific information, keywords, or topics within podcasts.
  • Language Support: Currently designed only to interact in English, translating the outputs into Spanish (Spanish based LLM models will be available soon...).

Getting Started

Prerequisites

Before you begin, ensure you have the following requirements in place:

  • Python: You need Python 3.7 or later installed on your system.
  • Virtual Environment (Optional): It's recommended to create a virtual environment to manage dependencies.
  • GPU: not mandatory, but recommended (GPU used: Nvidia T4)

Installation

  1. Clone the repository:

    git clone https://github.com/AlbertoUAH/Castena.git
cd Castena

  2. Install the required Python packages:

    pip install -r requirements.txt

  3. Run app (via Streamlit):

    streamlit run app.py --logger.level=warning 2 > 'app_log.log'

Usage

  1. Launch the chatbot using the instructions in the "Installation" section.

  2. The chatbot will prompt you to enter the podcast name or URL you want to interact with.

  3. Choose one of the available interaction options, such as transcription, summarization, or search.

  4. Follow the chatbot's prompts to provide additional details or requests.

  5. Enjoy your enhanced podcast experience!

Evaluation

To evaluate our approach, labelled datasets are needed. To do so, we employ Label Studio library

Metrics

To evaluate LLM performance, QAEvalChain, cosine_similarity and a sentence similarity model are used

Results

Video Name Channel Topic QAEvalChain (% Corrected Answer) Mean Cosine Similarity Median Cosine Similarity Mean Sentence Similarity Median Sentence Similarity
Worldcast #45 - Roberto Vaquero Worldcast Politics & Opinion 20 out of 30 corrected answers (~67 %) 0.7943 0.8198 0.7225 0.7406

Contributing

We welcome contributions to improve the functionality and capabilities of this chatbot. If you'd like to contribute, please follow these guidelines:

  1. Fork the repository.

  2. Create a new branch for your feature or bug fix: git checkout -b feature/new-feature.

  3. Make your changes and test thoroughly.

  4. Commit your changes with a clear and concise message: git commit -m "Add new feature".

  5. Push your branch to your forked repository: git push origin feature/new-feature.

  6. Create a pull request against the main repository's main branch.

  7. Provide a detailed description of your changes and why they are valuable.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Tools

  • Colab notebooks (free tier)
  • TogetherAI (LLM support)

Buy me a beer! 🍺