AIVish

Generate voice cloned audio recordings using tortoise-tts for security awareness testing.

Disclaimer

As with any pentesting tool, this should only be used with proper authorization during security awareness training (phishing/vishing). There are a number of checks to prevent abusive use of the tool for anything related to extortion, threats, slurs and foul language, and anything sexual or content related to war and terrorism.

Installation

You can run this tool on any system, however, it's designed to work with NVIDIA GPUs and you may lose quality if you're not on a system with an NVIDIA GPU. In the most extreme case, I have gotten AIVish to run on a small Kali VM running on a laptop. Note that AIVish is VERY resource intensive and also requires about ~20 gigs of storage. If you don't have any system that can handle that, you can spin up a G4dn.xlarge (Cheapest NVIDIA GPU on AWS) to run this tool. I have tested this tool on a G4dn.xlarge (~$0.50 an hour) running the Deep Learning Proprietary Nvidia Driver AMI GPU PyTorch community AMIs and it works well!

pip install git+https://github.com/d-woosley/AIVish

NOTE: Make sure you system is updated before installing (sudo apt update && sudo apt upgrade -y)

Troubleshooting Install

I've noticed that the recent versions of Deep Learning Proprietary Nvidia Driver AMI GPU PyTorch don't come with Rust (error: can't find Rust compiler) with is needed for tokenizers. To fix this issue, I had to install rust and run as root.

Troubleshooting Install

sudo -s
apt update
apt upgrade -y
python3 -m pip install --upgrade pip
curl --proto '=https' --tlsv1.2 -sSf https://sh.rustup.rs | sh
source ~/.bashrc

# Install aivish
pip install git+https://github.com/d-woosley/AIVish

Run

You will need a zip file containing 3-5 audio clips (around 10 seconds each) of the voice you want to clone (.wav format).

NOTE: I've used Audacity to get these clips from YouTube in the past.

     e      888 Y88b      / ,e,        888
    d8b     888  Y88b    /   "   d88~\ 888-~88e
   /Y88b    888   Y88b  /   888 C888   888  888
  /  Y88b   888    Y888/    888  Y88b  888  888
 /____Y88b  888     Y8/     888   888D 888  888
/      Y88b 888      Y      888 \_88P  888  888

usage: aivish [-h] -d <DIRECOTRY> [-i <FILE>] [-q {ultra_fast,fast,standard,high_quality}] [-o <FILE_NAME>] [-v]

Generate voice cloned audio recordings using tortoise-tts for security awareness testing

options:
  -h, --help            show this help message and exit
  -d <DIRECOTRY>, --dir <DIRECOTRY>
                        The path to the directory that contains the audio samples (".wav") of the voice to clone.
                        (minimum=3, ideal=5 clips of 10 seconds)
  -i <FILE>, --pretext-file <FILE>
                        The text to be generated into speech using given voice (If not given, you will be prompted)
  -q {ultra_fast,fast,standard,high_quality}
                        The quality level of the output: ultra_fast, fast, standard, high_quality (Default: high_quality)
  -o <FILE_NAME>, --output <FILE_NAME>
                        The output file name of the cloned voice audio file (Default="voice_clone.wav")
  -v, --verbose         Make the screen output verbose (show progress bars when generating clips)

by: Duncan Woosley (github.com/d-woosley)

Quick Run

aivish -d audio_files -i prompt.txt

NOTE: If no text file is provided using the -i flag, you will be prompted on the screen for input

Run with Lower Quality

aivish -d audio_files -q fast

Define Output File Name

aivish -d audio_files -o jdoe_missedcall.wav

Example Run

~/AIVish/tests$ aivish -d jdoe_clips -i pretext.txt -q ultra_fast

     e      888 Y88b      / ,e,        888
    d8b     888  Y88b    /   "   d88~\ 888-~88e
   /Y88b    888   Y88b  /   888 C888   888  888
  /  Y88b   888    Y888/    888  Y88b  888  888
 /____Y88b  888     Y8/     888   888D 888  888
/      Y88b 888      Y      888 \_88P  888  888


  [-] Downloading models used by Tortoise from HuggingFace. This could take a minute on the first run...

  [-] Loading voice from /home/ubuntu/AIVish/tests/jdoe_clips

  [-] Generating 3 sentences

    [+] Sentence 1 generated in 28.46 seconds. (Estimated Completion Time: 21.34 seconds)
    [+] Sentence 2 generated in 16.24 seconds. (Estimated Completion Time: 12.19 seconds)
    [+] Sentence 3 generated in 13.98 seconds. Done!

 [+] Complete! voice_clone.wav created!

Someday-Maybe Additions:

• Web interface to upload, generate, and download voice clips via the local network
• Integrate RVC for live voice cloning
• Integrate with PBX to fully handle calls as web softphone
• Separate controller host from worker hosts and allow for dynamically spinning up instances in the cloud for compute