Bio

Ying is a master student of astrophysics in beautiful Hokkaido, Japan. Though she always wanted to become a detective and was also fascinated by psychoanalysis, the movies Martian and Interstellar intrigued her into digging into more about the universe she inhabits. Now, she is doing simulations to investigate the impact of triaxial dark matter halos on the morphology of gaseous disc galaxies. She loves being stunned and inspired by new ideas and overturning facts, being at the ocean as much as possible and randomly doing crazy refreshing stuff and never doing it again.

Computational Essay: Contour Plots of Triaxial Dark Matter Halos Potential

Project: Gender-from-voice Predictor

Goal

By utilizing machine learning functions in Mathematica, we could predict genders of the input voices. The dataset we used is the open audio source VoxCeleb1 dataset, which contains over 100,000 utterances for 1,251 celebrities from YouTube videos. The powerful built-in Classify function helps us to classify real voices into categories, to learn from examples and to predict values from data. A neural network is applied to tackle a dataset with camouflaged or frequency-shifted voices.

Summary of Results

Because the built-in Classify function gives decent predictions on gender from real voice data but not camouflaged ones, we decide to train a neural network to deal with frequency-shifted voice recordings. We observed that the neural network could identify the gender of fake voices with a rough accuracy of 90% when the data is tagged female and male. In cases where the data was tagged by both gender and fakeness, the neural network returns a classifier with an accuracy of 83%.

Future Work

We could explore different layers for the neural network training to improve our classifier in identifying the gender of heavily frequency-shifted, camouflaged voices. Moreover, we didn’t deal with pitch-shifted voices, which was outside of our parameters.