Automatic Cot-Caught Merger Detection

This repository contains code for automatically detecting the cot-caught vowel merger in American English dialects using a machine learning approach.

Overview

This project uses a random forest classifier to identify speakers with the cot-caught merger, a phonological feature distinguishing some North American dialects. The model is trained using formant frequencies and vowel length from speech recordings. The cot-caught merger is a phonological shift where the vowels /ɑ/ (as in "cot") and /ɔ/ (as in "caught") are pronounced the same.

Data

The data comes from the DARPA TIMIT Acoustic-Phonetic Continuous Speech Corpus, available at https://catalog.ldc.upenn.edu/LDC93S1. Dialect region 6 is New York City (no merger) and dialect region 7 is the Western US (merger).

Usage

1. Clone the repository.

2. Install the required libraries using pip install -r requirements.txt.

3. Prepare your data:

   • TextGrid files must be named in the following format: dialectregion_gender+speakerID_sentence#.TextGrid. For example: DR7_MWRE0_SX337.TextGrid.
     • DR6 indicates the non-merger dialect region, and DR7 indicates the merger region.
   • Sound files (.wav) must have the same name as the TextGrid files, except with the .wav extension, for example DR6_m123_1.wav
     • The sound files should correspond to the TextGrid files in terms of the speech recording each contains.

4. Run the script process_and_train.py:

   python process_and_train.py --textgrid_dir <path to textgrid dir> --sound_dir <path to sound dir>

   Replace <path to textgrid dir> and <path to sound dir> with the actual paths to your directories containing the TextGrid and sound files, respectively.