Triple vowel formant tracker that measures formants using three different systems and computes the median of them. Additional scripts to visualize and identify outliers and mis-tracked formants. Also contains all scripts and notebooks to replicate results from Chapter 2 of dissertation.
- Notebooks to track formants and compute vocal tract length
- Scripts used to analyze vowel variability and generate results
The following Python dependencies are required to run the scripts in this repo: pandas, fnmatch, espspy, phonlab, audiolabel, (the latter three available here), numpy, seaborn, and matplotlib, in addition to other generic libraries that are pretty standard with any Python installation. Additional R libraries are needed and are listed in the R scripts/notebooks. In addition, this repo is configured for Python 3.7.0 or older. That doesn't mean it's unworkable with newer versions, but you will likely have to create some work arounds. One solution is to configure a Python 3.7.0 environment within Anaconda:
conda create --name py37 python=3.7.0 pip
conda activate py37
Throughout these notebooks, the output from the cells has been cleared as it contains deanonymized participant information
1_triple_formant_tracker.ipynb - track formants in three ways: 1) Inverse Filter Control Formant (Watanabe, 2001), 2) ESPS covariance, and 3) ESPS autocorrelation. Requires a .wav file for each speaker. Then, each tracker generates a directory containing the files with various acoustic measurements for each speaker (f0, formants, etc.). LPC order of 10 specified for the children and 12 for the adults.
2_get_median_formant.ipynb - compute the median measurement of each formant from the three trackers at 25%, 50%, and 75% of the vowel. Generates the file med_formants.csv which can be used in 3_check_for_outliers.ipynb, get_vt_length.ipynb, and to generate the manuscript results
2_get_median_formant_lpc8.ipynb - same as above, but attempting to track formants using an LPC filter order of 8 for children aged 4-6 years; this performed markedly worse than an LPC order of 10 (lots of mistracked F2) so LPC order of 10 was used for all children
3_check_for_outliers.ipynb - histograms that visualize overlap in formant distribution by age group and individual speaker to identify mis-tracked formants
get_vt_length.ipynb - compute child and adult vocal track length and DeltaF ratio using F1-F3 (for children) and F1-F4 (for adults); this script is sewn into the R script 3_vtl_results.Rmd that generates the results section
1_cleanvowels_eo.Rmd - select relevant lexical items and removes all formant measurements outside of 3 MADs from the median; not included in the public repo as it contains personal identifying information; generates the file cleaned_vowels_eo.csv which is used to run 2_rename_participants.R
2_rename_participants.R - map participant names to anonymized IDs; not included in the public repo as it contains personal identifying information; generates the file cleaned_vowels_eo.anonymized.csv which is used to run 3_vtl_results.Rmd
3_vtl_results.Rmd - generates results section of manuscript, including 1) calculation of vocal tract length/deltaF, 2) calculating vowel dispersion via coefficient of variation, 3) vowel normalization via Lobanov and DeltaF techniques, 4) modeling differences by normalization technique, 5) calculating formant ratios between supraglottal cavities, 6) modeling differences by supraglottal cavity formant ratios
validate_ratio.Rmd - script to validate the method for computing supraglottal cavity sizes using formant measurements (F1:F2 and F2:F3 ratios)
speaker_metadata_anonymized.csv - dataframe containing speaker metadata
calculate_CoV.py - short script to compute the coefficient of variation of formant measurements; vowel-extrinsic and intrinsic options
Watanabe, A. (2001). Formant estimation method using inverse-filter control. IEEE Transactions on Speech and Audio Processing, 9(4), 317-326.