Vowel AI — Research Guide

Technical methods and formulas. Copy-ready text and CSV templates included.

Vowel AI — Research Guide

الدليل البحثي — Vowel AI

Exact behavior per vowel_ai.py, viz.py, transforms.py.

1) Preprocessing & transforms

  • Numeric enforcement for F1,F2; invalid rows dropped.
  • Unit conversion: Mel & Bark formulas; Hz is default.
  • Normalization (per speaker): Lobanov z-scores; Nearey1 log-centering.
  • Arabic context: compute emphatic_env if contexts exist; else use given column.

2) Modeling & evaluation

AspectDetails جانبالتفاصيل
Features[F2,F1] default; optional F3,duration_ms,F0. السمات[F2,F1] افتراضيًا؛ وإضافيًا F3, duration_ms, F0.
Logistic RegressionStandardScaler; max_iter=200. الانحدار اللوجستيمع StandardScaler وmax_iter=200.
Random Forestn_estimators=200; defaults otherwise. الغابة العشوائيةn_estimators=200 والإعدادات الافتراضية.
Evaluationauto/holdout/CV; group-aware via GroupKFold/GroupShuffleSplit. التقويمتلقائي/احتفاظ/طيّ متقاطع؛ واعٍ للمجموعات عبر GroupKFold/GroupShuffleSplit.
OutputsConfusion matrices; classification_report.csv; predictions with probabilities. المخرجاتمصفوفات الالتباس؛ classification_report.csv؛ وتنبؤات باحتمالات.

3) CSV schema (complete)

speaker,sex,dialect,vowel,F1,F2,F3,duration_ms,F0,context_left,context_right,emphatic_env,notes
spk1,f,Cairene,i,300,2400,3100,95,220,س,,false,token near /s/
spk1,f,Cairene,e,500,2000,2800,110,210,ر,,false,
spk1,f,Cairene,a,780,1450,2500,120,200,ص,,true,adjacent emphatic /ص/
spk1,f,Cairene,u,350,900,2200,100,210,,ت,false,
spk1,f,Cairene,a,800,1400,2450,115,205,ق,,true,adjacent uvular /ق/
spk2,m,Najdi,i,310,2380,3050,90,120,خ,,true,adjacent /خ/
spk2,m,Najdi,a,760,1500,2400,100,115,,ط,true,following /ط/
spk2,m,Najdi,u,360,950,2150,95,118,ص,,true,
spk2,m,Najdi,e,520,1950,2750,105,122,,س,false,

4) Methods (ready to paste)

English

We analyzed vowel acoustics with the Vowel AI tool (Flask/Python; Matplotlib; scikit-learn). Tokens were annotated for vowel category and speaker, and measured for F1 and F2 in Hz. Preprocessing included optional unit conversion (Hz, Mel, Bark) and per-speaker normalization (Lobanov z-scores; Nearey1 log-centering). For Arabic, an emphatic_env flag was provided or automatically derived from adjacent segments (context_left/right) using a conservative set of emphatic/uvular/guttural consonants. Plots (F1×F2) used conventional axis orientation (F2 inverted) and optional ellipses (σ-dispersion, χ²-coverage, or mean-CI). Supervised models (Logistic Regression; Random Forest) were trained on F2,F1 (optionally F3, duration, F0) with evaluation via stratified holdout or k-fold cross-validation (speaker-held-out when specified). Outputs include confusion matrices, a classification report, and per-token predictions with probabilities. — created and developed by ظ

5) Result templates

A. Figure/description

Figure 1 displays the F1×F2 vowel space (Hz; F2 inverted), with IPA-labeled centroids and 1σ ellipses. Axes follow research-standard limits (F1: 200–900 Hz; F2: 600–2600 Hz). An emphatic contrast is operationalized via the emphatic_env flag, derived from adjacent segments when not present in the CSV. — created and developed by ظ

B. Interpretation

The inventory forms the expected triangle with front /i/ and back /u/. Emphatic environments shift /a/ toward higher F1 and lower F2, consistent with retraction under emphasis. Cross-validation shows robust classification for high-support vowels; confusions concentrate between mid front categories, suggesting partial overlap. — created and developed by ظ