Experiment 3: Vowel quality at F1 peak vs. midpoint
Experiment 3: Vowel quality at F1 peak vs. midpoint
Hypothesis
- Vowel is better characterized at F1 peak than at midpoint
- LAFF labelers seem to place Vowel landmarks at midpoint
Assumptions
- F1 has proper peak, not at endpoints
- F1 clear and measurable, without nasalization, glottalization or aspiration
- Landmark location in LAFF database is important
Data filtering
- Use only proper vowels (not schwas and syllabic sonorants)
- Use only vowels with proper peaks (not at endpoints)
- Do not use shibboleth utterances (sa1 and sa2)
- Result is 37008 tokens (27062 train, 9946 test)
Methodology
- Formant values taken at F1 amp peak and at midpoint
- K Nearest Neighbors (KNN) classification
- Euclidean distance metric (first three formants)
- K = sqrt(N)
Results
- Not much performance difference between F1 peak and midpoint
- McNemar test shows significant difference
- Ambiguous, but location does not appear critical