In polvanrijn/contouR: Compute dynamic features on pitch contours

base_dir = "/Users/pol/owncloud/HiWi_folder/02_running_projects/02_MA_thesis/01_Data/Preprocessed/"
pt_path = paste0(base_dir, "PitchTiers/08_manual_corrected/")
Fujisaki_path = paste0(base_dir, "Fujisaki/")
library(contouR)

Linear regression slope compared

Compute slopes by drawing a linear regression line. Let's first get an intuition by looking at the first 5 slopes.

files = list.files(pt_path)
files = files[grepl(".PitchTier", files)]
compute_regression_slopes(files[1:5], pt_path, plot_only=TRUE)

Now compute the slopes for all files:

features = compute_regression_slopes(files, pt_path)
head(features)

Let's check the correlations. The regression intercept correlates strongly with the naive intercept. There is also a smaller positive correlation between naive and regression slope. However, it is not correlated with the slope between global minimum and maximum.

library(corrplot)
interesting_columns = 2:9
interesting_columns = names(features)[interesting_columns]
cor_df = features[, interesting_columns]
print(corrplot(cor(na.omit(cor_df)), method="number", type = "upper"))

features$filename = features$name
features = add_meta_data(features)

interesting_columns = c("a", "naive_slope", "b", "naive_intercept")
for (col in interesting_columns){
  print(significance_test(na.omit(features), col))
}

Phrase command

From these optimal hyper parameter combinations we can compute features on the phrase command

top_scores = read.csv("top_scores.csv")
phrase_features = compute_phrase_features(top_scores, Fujisaki_path)
names(phrase_features) = c("name",  "num_phrases", "Ap", "phrase_T0", "phrase_reg_slope", "phrase_reg_RMSE", "phrase_naive_slope", "phrase_naive_RMSE")
head(phrase_features)
features = combine_features(features, phrase_features, key = 'name')

We see that both slope computations (and also the RMSE) are highly correlated. Also we find a negative correlation between the amplitude and the slope

corrplot(cor(na.omit(phrase_features[, 3:8])), method="circle", type = "upper")

An visualize the features in boxplots:

phrase_features$filename = phrase_features$name
interesting_columns = names(phrase_features)[c(2:5, 7)]
for (col in interesting_columns){
  print(significance_test(na.omit(phrase_features), col))
}

ICCs

We can also devide the contour in it's ICCs. Let's first give some examples for the ICC separation:

compute_EAC_ICCs(files[1:5], pt_path, plot_only=TRUE)

We can now extract the features from the ICC as Mirjana Rajkovic, Jovicic, Grozdic, Zdravkovic and Subotic (2018) propose

feature = compute_EAC_ICCs(files, pt_path)
head(feature)
features = combine_features(features, feature, key = 'name')

We can have a look at the correlations:

# Many
feature$filename = feature$name
cor_df = feature[, names(feature)[2:10]]
corrplot(cor(na.omit(cor_df)), method="circle", type = "upper")
cor_df = feature[, names(feature)[2:11]]
corrplot(cor(na.omit(cor_df)), method="circle", type = "upper")

An visualize the features in boxplots:

interesting_columns = names(feature)[2:11]
for (col in interesting_columns){
  print(significance_test(na.omit(feature), col))
}

Last but not least, we have to save the features in a CSV for later use:

write.csv(features, "slope_features.csv", row.names = FALSE)

polvanrijn/contouR documentation built on Feb. 10, 2020, 3:45 a.m.