Get AutoVC algorithm (open source) running through a simple python script on MacOS Monterey
Budget: $30 – $250 USD
Background:
Researchers have developed the AutoVC algorithm, an algorithm capable of cloning a voice. There are multiple python implementations of this AutoVC algorithm readily available on GitHub and most provide sufficient documentation. This AutoVC algorithm allows an audio file (wav) of a source speaker to be transformed into the voice of a target speaker using deep learning (torch).
Objectives:
Be able to run the AutoVC algorithm on a Mac (MacOS Monterey running on Apple Silicon) and perform voice transformations with the minimum possible effort. The quality should be comparable to the Zero-Shot section in the AutoVC Demo page: https://auspicious3000.github.io/autovc-demo/#zero-shot
The final deliverable should contain everything to be able to do voice conversion without needing to do local training.
Deliverables:
Write a python script that takes the following four parameters:
1) a path to a folder containing recordings of the source voice (if AutoVC expects a file rather than a folder, then it can be a path to a single .wav file)
2) a path to a folder containing recordings of the target voice (if AutoVC expects a file rather than a folder, then it can be a path to a single .wav file)
3) a path to an audio file (.wav) with a recording of the source voice
4) a path to an output audio file
The script is expected to take these four parameters and use the existing AutoVC algorithm to transform the source audio file (parameter 3) into the voice of the destination speaker and write that new generated audio to the output file (parameter 4). The final deliverable should require no training to be run, just the conversion.
The deliverable should contain all files necessary to run the script: pretrained models, libraries (for example in a venv), code, whatever is required. The structure is not relevant as long as the deliverable can be downloaded on a computer with MacOS Monterey running on Apple Silicon and run with little effort.
Risks:
All major components (e.g. torch) can be run on a mac with apple silicon. I got them to work in different projects. The difficulty probably is in simplifying the existing code to be able to run in a one-line script, find or generate the right pre-trained data to reproduce the quality shown in the AutoVC Demo page and package everything into a deliverable that is easy to run.
Researchers have developed the AutoVC algorithm, an algorithm capable of cloning a voice. There are multiple python implementations of this AutoVC algorithm readily available on GitHub and most provide sufficient documentation. This AutoVC algorithm allows an audio file (wav) of a source speaker to be transformed into the voice of a target speaker using deep learning (torch).
Objectives:
Be able to run the AutoVC algorithm on a Mac (MacOS Monterey running on Apple Silicon) and perform voice transformations with the minimum possible effort. The quality should be comparable to the Zero-Shot section in the AutoVC Demo page: https://auspicious3000.github.io/autovc-demo/#zero-shot
The final deliverable should contain everything to be able to do voice conversion without needing to do local training.
Deliverables:
Write a python script that takes the following four parameters:
1) a path to a folder containing recordings of the source voice (if AutoVC expects a file rather than a folder, then it can be a path to a single .wav file)
2) a path to a folder containing recordings of the target voice (if AutoVC expects a file rather than a folder, then it can be a path to a single .wav file)
3) a path to an audio file (.wav) with a recording of the source voice
4) a path to an output audio file
The script is expected to take these four parameters and use the existing AutoVC algorithm to transform the source audio file (parameter 3) into the voice of the destination speaker and write that new generated audio to the output file (parameter 4). The final deliverable should require no training to be run, just the conversion.
The deliverable should contain all files necessary to run the script: pretrained models, libraries (for example in a venv), code, whatever is required. The structure is not relevant as long as the deliverable can be downloaded on a computer with MacOS Monterey running on Apple Silicon and run with little effort.
Risks:
All major components (e.g. torch) can be run on a mac with apple silicon. I got them to work in different projects. The difficulty probably is in simplifying the existing code to be able to run in a one-line script, find or generate the right pre-trained data to reproduce the quality shown in the AutoVC Demo page and package everything into a deliverable that is easy to run.