Skip to content

Files

Latest commit

 

History

History

feature-computation

Folders and files

NameName
Last commit message
Last commit date

parent directory

..
feature-test-plots
feature-test-plots
 
 
images
images
 
 
mid-level-toolkit-modifications
mid-level-toolkit-modifications
 
 
README.md
README.md
 
 
baseFeatCodeToName.m
baseFeatCodeToName.m
 
 
computeFeatures.m
computeFeatures.m
 
 
computePrincipalComponentDistance.m
computePrincipalComponentDistance.m
 
 
featuresToTimeTables.m
featuresToTimeTables.m
 
 
getExpectedFeatureVecLen.m
getExpectedFeatureVecLen.m
 
 
getFragFeatureSet.m
getFragFeatureSet.m
 
 
makeDirIfNotExists.m
makeDirIfNotExists.m
 
 
makeMonoTrackSpec.m
makeMonoTrackSpec.m
 
 
plotFeatureDistributions.m
plotFeatureDistributions.m
 
 
plotFeatures.m
plotFeatures.m
 
 
plotFeaturesDual.m
plotFeaturesDual.m
 
 
plotFragFeatures.m
plotFragFeatures.m
 
 
readFragsMetadataToTable.m
readFragsMetadataToTable.m
 
 
reaper_on_dir.sh
reaper_on_dir.sh
 
 
runPCAfunctions.m
runPCAfunctions.m
 
 

README.md

feature-computation

This directory contains code for computing prosodic features of DRAL utterances ("short fragments") using the Mid-level Toolkit.

To estimate the prosody similarity of utterances, see Estimate the prosodic dissimilarity of utterances.

Compute prosodic features of DRAL utterances

  1. Gather the DRAL data.
  2. Setup MATLAB and Python (see Setup).
  3. Prepare the data for feature computation by running the Python script prep_for_feature_comp.py. This script concatenates utterances into new audios, estimates pitch with REAPER, and modifies the corpus metadata.
  4. Compute the features by running the MATLAB function computeFeatures.m. This script computes prosodic features using the Mid-level Toolkit and modified Mid-level Toolkit functions.

Plot prosodic features in an interactive figure

To plot the prosodic features of a DRAL utterance, run the MATLAB function plotFragFeatures.m.

Preview

plotFeatures.m example

Test audio

See ./test-audio/README.md.

Modifications to the Mid-level Toolkit

Modified pitch computation

The script lookUpOrComputePitchModified.m is modified from Midlevel Toolkit lookupOrComputePitch.m. The function lookUpOrComputePitchModified removes the fallback to VOICEBOX fxrapt enforcing REAPER for pitch tracking.

For more information, see documentation in lookUpOrComputePitchModified.m.

Modified CPPS computation

The script computeCPPSmodified.m is modified from Midlevel Toolkit computeCPPS.m. The function computeCPPS runs about 5 times faster by reducing the number of spectrogram windows computations (as tested by timeCPPSfunctions.m) and pads the return CPPS feature vector so that each 10 milliseconds of the input signal is assigned a CPPS value.

This modified function is unused; it did not have a strong correlation with the original function, as tested by plotCPPSfunctions.m.

Interactive feature plot

The function plotFeatures.m, called by plotFragFeatures.m is generalized and can be used with other features computed outside of DRAL.

Notes

I wanted to write all code in Python and use the MATLAB engine to call the Midlevel Toolkit functions I needed, but not all Midlevel Toolkit functions can be called from Python because they return unsupported types.