ddurieux / whisper_translate

Real Time Whisper Translation

Adaptation

This script is an adapation of the script from https://github.com/davabase/whisper_real_time

Introduction

This script transcript english audio and translate in french, need chatGPT account from the translation part.

It works by constantly recording audio in a thread and concatenating the raw bytes over multiple recordings.

Requirements

The requirements are:

• an NVIDIA card with CUDA cores (like RTX...)
• python 3
• ffmpeg

Installation

You need install:

• install ffmpeg, on windows, add the bin path to the environment PATH
• install CUDA toolkit https://developer.nvidia.com/cuda-toolkit, YOU NEED INSTALL IT BEFORE PYTHON LIBS DEFINED AT NEXT BULLET
• install git
• Install python dependencies:

  pip install -r requirements.txt
  

• For chatGPT translation edit the file transcribe.py and set the key in the begining of the file.

Usage

List audio devices with command:

python transcribe.py --default_microphone list

Choose the device you want, for exemple "mic" and run:

python transcribe.py --default_microphone "mic" --model small

If you graphic card has 10 Go of memory and more, you can replace small by medium

Integrate translation into OBS

The script write on screen, but it write too in file subtitle.txt in current folder.

IN OBS, add text GDI+, check Chatlog Mode and the line limit to 3.

Special case: translate streaming video

Linux

Under linux, you can select the sound card device named pipewire.

Windows

• install VB-Audio in case you want translate streaming video and not only your microphone
• manage in "app volume and device preferences" in windows settings to put your browser video into Cable Input (VB-Audio Virtual Cable)