A Computational Approach to Foreign Accent Classification

Emily Ahn

Wellesley College Undergraduate Senior Thesis // May 2016

Full write-up in pdf form can be found in the writeup/ folder or in the Wellesley College repositories.

1. Folders in this repository:

alignments/

• Contains forced alignents and transcriptions for the text-dependent classifier
• The 7 subfolders correspond to each of the 7 transcribed accents
• Simple transcriptions of speech files are organized by accent. The format is a .csv file compiled via releasing transcription tasks on Amazon Mechanical Turk, then personally cleaned up by the author. *Note: errors still exist in some transcriptions.

trans-results/ and untrans-results/

• Console print logs of results from the text-dependent "trans" (transcribed) classifier and the text-independent "untrans" (untranscribed) classifier

traintestsplit/

• Contains lists of filenames that were split into train and test data, via a randomized 75:25 split

formants/

• Script and data (in csv format) to test GMM classifcation based on 3 vowel formants for AR, CZ, and IN accents

2. Main scripts

• Text-independent (untranscribed) Classifier
  • gmmClassifier.py || full script; loads data, trains models, classify test data
  • gmmTrain.py || modularizes training only, stores models in directory
  • gmmTest.py || modularizes testing only
• Text-dependent (transcribed) Classifier
  • parseTextGrid.py || prepares data by converting forced alignments of speech into plp features (sorted by accent and phoneme)
  • phoneClassifier.py || full script; gmm Classification of transcribed phonemes

3. Miscellaneous scripts

• avgnpy_test.py || takes average of each dimension of PLP vector across all time windows from a given sound file
• miniClassifier.py || does univariate GMM classification of AR, HI, MA