This algorithm enables to convert one’s voice to someone elses voice. Developped as a side project, my former company, Xebia, gave me the opportunity to present it at our yearly conference.
The algorithm works by using a vocal feature extractor (pitch + formants).
At learning time, the features are aligned with those of the target voice thanks to a Dynamic Time Warping algorithm. Then the mapping between the features is learned by the neural network. At synthesis time, my voice features are extracted and converted by the neural network without alignment. Finally I use my vocal synthesizer to generate the target voice.
A project in which I invested some time. The code contains
Python wrappers for API calls to get informations about the values of previous trades.
Python wrappers to download historical data
Recursive exponential smoothing functions for nonhomogeneous time sampling to extract smoothed features such as averages, variances at with parametrable smoothing, and their mathematical derivations.
Some strategies that I tried (arbitrage, noise scalping and more)
Fascinated by this audio effect I decided to code a vocoder by myself to see how it works.
Here is an example of how it sounds on my own voice
Modulator input sound
Carrier input sound
Result / output sound
Beat Box to Real Percussions converter (2016)
Here is an audio utility that converts “beatbox” sounds one make with its mouth to real percussion sounds. No Neural Networks were used here, only a research of a closest neighbor in the space of the features.
This repository contains code and documentation for some audio tools I made. The goal of the project is to group the work also done in my other sound-related projects, such as
Pierre Sendorek. “Amélioration de l’intégrité d’un système de positionnement GNSS/IMU” (Improvement of the integrity of a GNSS/IMU positioning system). PhD Thesis, French. CIFRE with Thales Avionics and Télécom ParisTech, june 2015. *Industrial Confidentiality.