Deepfake Radio

Code for using Elevenlabs voice cloning for fun with friends

Discord Bot

/speak
- Choose a voice, provide text and receive an audio clip in a message (embeds natively on Desktop but on mobile may need an app like VLC to open)
/conversation
- Provide a script in the format ---Voice Name: hi how are you? --Voice Name: Pretty good! and have the bot generate audio and splice it together into a conversation

Tools for processing and uploading audio samples for voice cloning

Config for specifying
- Source (local or YouTube URL)
- Audio start and end time
- Voice isolation using an open source ML model (Demucs)
- Set clip sizes to automatically clip into smaller parts

Bot Setup

Clone or fork this repo

git clone https://github.com/safurrier/deepfake-radio/

Set up the bot

Visit the Discord Developer portal to creata new bot application
In the top right, select New Application and give the app a name
If you want to update the bot avatar, description etc do so here
On the left, click Bot
Scroll down to find the section MESSAGE CONTENT INTENT and turn this on. Save Changes.
Scroll back up and click Reset Token. Make sure to copy this information down as you'll need it to run your bot
Create a file in the top level directory called .env. You can do this by making a copy of the file example.env
In the .env file, set BOT_TOKEN=$YOUR_TOKEN on a line. E.g. BOT_TOKEN=WWKDEQX8JOoxW61WPb6dXIzklNjaMHJf2zuHGk

Join ElevenLabs and sign up for an account with API access. Try promo code for BETA11 which may still be active for a free month of Creator tier access.
Copy your ElevenLabs API token (Top Right -> Profile -> API Key Section) to the .env file under ELEVEN_API_KEY like ELEVEN_API_KEY=asdfaskjd123498sasw
Create some voice clones using ElevenLabs.

See section Add Custom Voices for how to do this programtically, or clone them on the ElevenLabs platform
Branch voice_catalog contains some custom voices ready to go that you can copy into the voices directory if you so desire.
This is a good short guide on some tips to creating good voice clones
In general, quality > quantitity. (It's possible Elevenlabs doesn't use anything longer than 2 minutes for samples)
Diversity of voice samples and short clips can help
Consider adding many short clips and on the web UI editing the voice clones to select only the best ones
The best samples are of a clear voice with no interruptions or background noise (like a monologue or audio book).

Get a python environment

This was written with a python 3.11 environment but may work with other major versions
I would suggest using Conda as an environment manager as it's fairly simple to use
Installing conda is outside of the scope of this tutorial (some instructions here), but if it's already installed you can run the make command to get a python env setup

conda create --name deepfake_radio python=3.11 -y

or using the make command make create-conda-env

And then activate the env

conda activate deepfake_radio

Install the dependencies

make requirements or manually run:

  pip install -r requirements.txt
  chmod +x install_dependencies.sh
  sh install_dependencies.sh

Run the bot

python main.py

Invite the bot to your server

Return to the Discord Developer portal
Select OAuth 2 -> URL Generator
Under Scopes, select bot
Under Bot Permissions select at least Send Messages, Send Messages in Threads, Embed Links, Attach Links, and Use Slash Commands
Copy the generated URL underneath (you may want to save this link somewhere if you plan on inviting the bot to multiple servers)
Select the server that you're a modmin on and add it to server. Authorize the bot.

Use the bot! Use the /speak command to select a voice add text and generate audio. If no voices are present in the options, try running /update-voices first. Have fun!

Deploy as a Railway App

Go to Railway
Sign up using Github
Verify your account
Create a new project
Deploy from a Github Repo -> select your copy of this repo.
Add Variables -> Add the env variables for BOT_TOKEN and ELEVEN_API_KEY (optionally set UPLOAD_VOICES if you want to upload voices before startup. Do not set PROCESS_VOICES as it breaks the Runway deployment as of now)

Add Custom Voices

You will need audio samples to create voice clones in ElevenLabs. You can do this through the platform, or alternatively there are automated tools to do so included here. By default, the bot will process and upload custom voices to the ElevenLabs API before startup. Previously added voices are skipped.

There may or may not be a voice catalog branch with many voice samples already set up. All you would have to do is copy over the voices you'd like to into the voices directory and turn on env variable UPLOAD_VOICES. You didn't hear it from me.

All voice clones belong in the voices directory and require a config.yaml file. The two accepted sources for samples are a local mp3 file and a Youtube video link
Each config file requires a name, an optional description and at minimum, a specified source and location setting.
There is an example config under voices/example:

name: Joe Biden
description: President Joe Biden

1:
  source:
    location: input/biden.mp3
    type: file
2:
  source:
    location: https://www.youtube.com/watch?v=KADpsS8fbg8
    type: youtube
    # Note: start_time and end_time are optional
    # but must be in format "HH:MM:SS"
    start_time: "00:00:01"
    end_time: "00:10:01"
  isolate: True
  clip_size: 60

Each number is a separate source, location is a URL for Youtube and filepath for file (relative to the config file). In general this should be input/your_sample.mp3

There are other optional configurations to facilitate creating good samples including start and end time, voice isolation and clip size.

The voice isolation uses an open source model demucs for voice isolation. This can be helpful if there is background noise, but does not always work well. It will also increase the processing time substantially.

Make sure in the .env file PROCESS_VOICES and UPLOAD_VOICES is set to a value. Remove this env var if you don't want custom voices to be processed, and uploaded on bot startup.
NOTE: Remove the example voice otherwise it will be uploaded as a voice clone to your ElevenLabs account
Run the bot with env vars set for PROCESS_VOICES (if adding new custom voices that need processing) and/or UPLOAD_VOICES (if processed files present and no processing needed, just set this to upload to ElevenLabs)
If you're looking for voice clone samples, there may or may not be a branch on this repo with available files.

Future Improvements and TODOs

[] Add CLI tool for processing + uploading voices

[] Add tag support for voice config

[] Improve voice selection option to sort by tags

[] Have the bot set the API key using a command

[] Add tests

TODO: Tests

Where we're going, we don't need tests. If the bot breaks on your server and your friends complain tell them to kick rocks or fix it themselves.

Original inspiration of the speak command from ElevenLabsBot

safurrier / deepfake-radio