🎙️ Whisper Audio/Video Transcription App

📝 Description

This project is a transcription app built using the Faster Whisper model, which transcribes audio and video files into text. It is powered by Gradio for a user-friendly web interface and supports audio or video file uploads for transcription.

✨ Features

• 🎧 Transcribe both audio and video files (e.g., MP3, MP4, AVI, etc.)
• ⚖️ Supports multiple model sizes for performance vs. accuracy balance
• 🚀 GPU support for faster transcription using CUDA
• 🎥 Extracts audio from video files automatically
• 🔍 High-precision transcription with options for beam search and other configurations
• 🖥️ Simple UI built with Gradio for easy access and use
• ⬇️ Download the transcript in .txt format
• 🎛️ Tuning the model parameters via the interface

Demo

💻 You can try the Colab version here (remember to select GPU in 'Runtime Type' for faster execution ⚡)

🛠️ To do/fix

• 🖥️ MAC/AMD Support (if you have one of those contact me 📩)
• You tell me! 🙂

📋 Requirements

• 🐍 Python 3.11+
• 🔥 PyTorch (CUDA version + CUDA Toolkit if using GPU)
• 🎬 FFmpeg (must be installed and added to your system's PATH)
• 🖼️ Gradio

📦 Installation

Step 1: Clone the repository

git clone https://github.com/CorsiDanilo/whisper-utility.git

Step 2: Set up a virtual environment (optional but recommended):

python3 -m venv .venv
source .venv/bin/activate  # On Windows, use `.venv\Scripts\activate`

Step 3: Install the required dependencies

pip install -r requirements.py

Step 4: Install FFmpeg (if not already installed):

• 🐧 Linux: Install via your package manager (e.g., sudo apt install ffmpeg)
• 🍎 macOS: Install via Homebrew (brew install ffmpeg)
• 🖥️ Windows: Download FFmpeg and add it to your system's PATH.
  • Follow this guide to ensure it's in your system's PATH.

(OPTIONAL) Step 5: Install Pythorch and CUDA Toolkit for NVIDIA GPU

• Download and install PyTorch, in Compute Platform select the last version.
• Download and install CUDA Toolkit.
  • 🐧 Linux: follow this guide.
  • 🖥️ Windows: follow this guide.

🚀 Usage

Run the application:

python whisper.py 

The Gradio interface will open in your default web browser. From there, you can upload an audio or video file, and the transcription will be displayed.

💡 REMEMBER: When you are done click Clear and Close if you want to clean up the temporary files folder.

🎛️ Interface Guide

• Upload an audio or video file: Accepts audio formats like MP3, WAV, and video formats like MP4, AVI.
• Transcribe: Click this button to start the transcription process.
• Close and Clear: This button clears the folder where the file was temporarily stored and closes the application.

⚙️ Model Configuration

• Language: Set the transcription language. Default is Italian 🇮🇹 (it), but you can change it to English 🇬🇧 (en) or other languages.
• Model Size: By default, the large version of the Whisper model is used (large-v3), but you can switch to small-v3 for smaller, faster models.
• Device: The model automatically selects the device based on GPU availability (cuda or cpu).
• Beam Size: Set beam size for decoding. Default is 4, but you can reduce it to 1 for faster inference.

🛠️ Troubleshooting

• If you get the following error:

  Could not locate cudnn_ops_infer64_8.dll. Please make sure it is in your library path!
  

  Download the missing dll from here and put it into the bin folder of your CUDA installation folder.
  • 🗂️ The usual path is: C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\vXX.X\bin.

📄 License

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

🙏 Acknowledgments

• Faster Whisper by Guillaume Klein
• Gradio for the UI interface