An Outlier Analysis of Vowel Formants from a Corpus Phonetics Pipeline

Emily P. Ahn (University of Washington)

• Python 3.9.0
• Montreal Forced Aligner 2.0.0 (via conda)
• R 4.2.2
• Praat 6.0.16

• Paper PDF: ./Outliers_Interspeech_230601.pdf
• Poster PDF: ./poster_interspeech_emilyahn_2023.pdf

@inproceedings{ahn_outlier_2023,
  title={An Outlier Analysis of Vowel Formants from a Corpus Phonetics Pipeline},
  author={Ahn, Emily P and Levow, Gina-Anne and Wright, Richard A and Chodroff, Eleanor},
  year = {2023},
  booktitle = {Interspeech},
}

• Disclaimer: scripts that are 'quick-and-dirty' start with q_ and are not meant to reflect high quality code

• download data, follow data/cv8/{lang}/processCommonVoice_{lang}_v8.txt steps
  • (includes using praat scripts, Epitran for G2P, MFA for forced alignment)
• extract formants with src/getFormantsCommonVoice_highlow.praat (under high and low settings)
• assign each speaker to either high or low formant setting (will use only low for analysis) with python src/assign_formant_range.py > data/cv8/formants/settings_highlow.tsv
  • filter this file to just get low setting speakers: head -1 data/cv8/formants/settings_highlow.tsv > data/cv8/formants/settings_low.tsv; cat data/cv8/formants/settings_highlow.tsv | awk '$3 == "low" {print}' >> data/cv8/formants/settings_low.tsv
• quick script to just get formant data of F1/F2 midpoints and low setting speakers only: src/q_simplify_cv_formants.py
• additional script to get AVG F1/F2 per vowel per speaker: src/get_avg_form_cvlow.py

• download audio from (Faith Comes by Hearing)[https://www.faithcomesbyhearing.com/audio-bible-resources/mp3-downloads]
• segment audio: follow steps from CMU Wilderness (dataset)[https://github.com/festvox/datasets-CMU_Wilderness]
• download TextGrids from (VoxClamantis)[https://voxclamantisproject.github.io/]
• download formants from VoxClamantis (now stored in data/wild/formant_raw/); F1/F2 midpoints only
• convert formant phones from XSAMPA to IPA with python src/q_xs2ipa.py data/wild/formant_raw data/wild/formant_ipa formant_csv
• convert TextGrid phones from XSAMPA to IPA with python src/q_xs2ipa.py /Users/eahn/work/typ/data/voxclamantis/tg/{lang} data/wild/tg_ipa/{lang}/ tg

• From formant files, create distributions per vowel with src/mahal_{wild,cv}.py
  • uses SciKit Learn package MinCovDet
  • output to another csv file with column 'MD' (Mahalanobis distance)
  • src/mahal_cv.py can take a subset argument to only get n utterances (used in this case for Common Voice Swedish)
• To get outliers at 0.1% edge of distribution, use mahal_thresh = 13.82 (alpha = 0.001, 2 degrees of freedom)
• Subset errors (thresh = 13.82, num_samples = 100) with python src/q_random_errors.py data/wild/mahal/{lang}_vowels_all.csv data/wild/annotate/{lang}_vowels_100.csv
  • Subset at other points along the MD distribution (e.g. 'near-perfect' samples of Mahal dist < 1.0) python src/q_random_good.py data/cv8/mahal/kazakh_vowels_all.csv data/cv8/annotate/kazakh_good_40.csv 1 40
• Move only num_samples audio and textgrids into folders for annotation:

# wild errors
lang="SWESFV"; for utt_id in `cat data/wild/annotate/${lang}_errors_100.csv | tail -n +2 | cut -d"," -f1 | sort -u`; do source_file="/Users/eahn/work/typ/data/audio/wav_seg/${lang}/${utt_id}.wav"; cp $source_file data/wild/audio/${lang}_errors_100; done
lang="SWESFV"; for utt_id in `cat data/wild/annotate/${lang}_errors_100.csv | tail -n +2 | cut -d"," -f1 | sort -u`; do source_file="data/wild/tg_ipa/${lang}/${utt_id}.TextGrid"; cp $source_file data/wild/annotate/tg/${lang}_errors_100; done
# wild good
lang="SWESFV"; wav_dir="data/wild/audio/${lang}_good_40_wav"; mkdir $wav_dir; for utt_id in `cat data/wild/annotate/${lang}_good_40.csv | tail -n +2 | cut -d"," -f1 | sort -u`; do source_file="/Users/eahn/work/typ/data/audio/wav_seg/${lang}/${utt_id}.wav";  cp $source_file $wav_dir; done
lang="SWESFV"; tg_dir="data/wild/annotate/tg/${lang}_good_40_tg"; mkdir $tg_dir; for utt_id in `cat data/wild/annotate/${lang}_good_40.csv | tail -n +2 | cut -d"," -f1 | sort -u`; do source_file="data/wild/tg_ipa/${lang}/${utt_id}.TextGrid";  cp $source_file $tg_dir; done
# cv errors
lang="kazakh"; wav_dir="data/cv8/annotate/${lang}_errors_100_wav"; mkdir $wav_dir;for utt_id in `cat data/cv8/annotate/${lang}_errors_100.csv | tail -n +2 | cut -d"," -f1 | sort -u`; do source_file="data/cv8/${lang}/prep_validated/${utt_id}.wav";  cp $source_file $wav_dir; done
lang="hausa"; tg_dir="data/cv8/annotate/${lang}_errors_100_tg"; mkdir $tg_dir;for utt_id in `cat data/cv8/annotate/${lang}_errors_100.csv | tail -n +2 | cut -d"," -f1 | sort -u`; do source_file="data/cv8/${lang}/aligned_validated/${utt_id}.TextGrid";  cp $source_file $tg_dir; done
# cv good files
lang="kazakh"; wav_dir="data/cv8/annotate/${lang}_good_40_wav"; mkdir $wav_dir;for utt_id in `cat data/cv8/annotate/${lang}_good_40.csv | tail -n +2 | cut -d"," -f1 | sort -u`; do source_file="data/cv8/${lang}/prep_validated/${utt_id}.wav";  cp $source_file $wav_dir; done
lang="hausa"; tg_dir="data/cv8/annotate/${lang}_good_40_tg"; mkdir $tg_dir;for utt_id in `cat data/cv8/annotate/${lang}_good_40.csv | tail -n +2 | cut -d"," -f1 | sort -u`; do source_file="data/cv8/${lang}/aligned_validated/${utt_id}.TextGrid";  cp $source_file $tg_dir; done

• Guidelines: see ./AnnotationGuidelines.pdf
• Master spreadsheet including re-annotations with new Formant category: data/master_reannot_221114.csv

• logistic regression for Kazakh high vowel deletion case study:
  • prep data: src/prep_logregsib_data.py
  • run regression in R: src/logreg_deletion.R

We thank Anna Batra, Sam Briggs, Ivy Guo, and Emma Miller for their work on the annotations, processing of results, and exploratory data analyses. Author EPA was supported by NSF GRFP grant DGE-2140004, and Author EC by SNSF grant 208460.