Vowel clustering using Hillenbrand et al. 1995 data

Vowel ID prediction and clustering using Hillenbrand et al. 1995 data

Clustering of the Hillenbrand vowel data

This repo contains Jupyter/IPython notebooks and Python and R scripts using the data from Hillenbrand et al. 1995 on the acoustic measurements on American English vowels. The vowels are either identified using a Support Vector Machine or separated into different clusters using k-means clustering and Gaussian mixture models. Learning algorithms are implemented in scikit-learn and visualization is performed in R using ggplot2 (clusters only). For the supervised learning, formant values are used as predictors; for the cluster analyses, formants and formant ratios are used as the features for the clustering.

Steps:

  1. Import data and remove rows with at least one missing observation
  2. Identify observation by word, vowel, and sex of speaker
  3. Map vowel characteristics to observation (e.g., front, open-mid, etc.)
  4. Create targets for speaker sex, word, and vowel (for supervised learning)
  5. Normalizing features (z-score)
  6. Create feature matrices
  7. Implement classification or clustering algorithms
  8. Visualize clusters and feature space

Directories and contents

data:

notebooks:

scripts: