Vowel AI — Student Guide

Beginner-friendly walkthrough: upload a CSV, visualize vowels, train a small model, and predict new tokens.

Vowel AI — Student Guide

دليل Vowel AI للطلاب

Upload a CSV with F1 and F2, see the vowel space, train a simple model, and predict new tokens.

1) What happens to your CSV

  1. Checks F1 and F2 are numbers; bad rows are discarded.
  2. يتحقّق من كون F1 وF2 أرقامًا؛ تُزال الصفوف غير الصالحة.
  3. Adds default vowel/speaker if missing.
  4. يضيف أعمدة افتراضية vowel وspeaker إن غابت.
  5. If you lack emphatic_env but include context_left/right, the tool flags emphatic/guttural neighbors (e.g., ط ض ص ظ ق غ خ ح ع, T D S Z q g x H E).
  6. إذا لم يوجد emphatic_env وكان هناك context_left/right، ستضع الأداة علَماً عند وجود جارٍ مفخّم/حلقي (مثل ط ض ص ظ ق غ خ ح ع وT D S Z q g x H E).
  7. Applies your chosen units (Hz/Mel/Bark) and optional normalization (Lobanov/Nearey1).
  8. تطبّق الوحدات التي اخترتها (Hz/Mel/Bark) والتطبيع الاختياري (Lobanov/Nearey1).

2) What the vowel plot looks like

Front vowels have higher F2; low vowels have higher F1. By convention, F2 decreases to the right.

F1 (Hz) F2 (Hz, inverted) /i/ (n=12) /a/ (n=15) /u/ (n=10)

3) How to read a confusion matrix

Rows = true class; columns = predicted. Dark cells on the diagonal are good; off-diagonal cells show confusions.

Predicted → ← المتنبأ به True ↓ الحقيقي ↑ /i/ /a/ /u/ /i/ /a/ /u/ 18 3 1 4 17 3 2 3 19
  • Diagonal dominance ⇒ classes are separable.
  • If /e/ vs /i/ confusion is high, consider adding F3 or duration.
  • سيادة القطر ⇒ الفئات متمايزة.
  • إن كثر الالتباس بين /e/ و/i/ فجرّب إضافة F3 أو المدة.

4) Example CSV (copy & paste)

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,

5) Methods paragraph (ready to paste)

English

We analyzed vowel acoustics with the Vowel AI web tool. Tokens were labeled for vowel and speaker and measured for F1 and F2 in Hz. The tool optionally converted units (Hz, Mel, Bark) and applied per-speaker normalization (Lobanov z-scores or Nearey1) when pooling speakers. For Arabic data, an emphatic_env flag was provided in the CSV or automatically computed from adjacent segments (context_left/right) using a conservative set of emphatic/uvular/guttural consonants. Vowel spaces (F1×F2) were plotted with centroid labels and optional ellipses (dispersion, coverage, or mean CI). Supervised models (Logistic Regression or Random Forest) were trained on F2,F1 (optionally F3, duration, F0), evaluated with stratified holdout or k-fold cross-validation. The tool exports confusion matrices, a classification report, and token-level predictions with probabilities. — created and developed by ظ

6) Result templates

A. Figure description

Figure 1 shows the F1×F2 vowel space (Hz; F2 inverted), colored by vowel and faceted by dialect. Centroids are labeled with IPA and token counts; shaded ellipses display 1σ dispersion. Axes use standard limits (F1: 200–900 Hz; F2: 600–2600 Hz). An emphatic contrast is illustrated by comparing tokens flagged emphatic_env=true versus false. — created and developed by ظ

B. Interpretation

The vowel triangle is evident with /i/ fronted and /u/ backed. In emphatic environments, /a/ shifts toward higher F1 and lower F2, consistent with emphasis-induced retraction. Classifier performance is strongest on high-support vowels (e.g., /i/, /a/), with confusions concentrated in the mid-front region (/e/ vs /i/). Dialectal facets show greater /i/ fronting in Cairene relative to Najdi. — created and developed by ظ