Praat Tutorial: Measure Vowel Formants (F1 & F2) in Acoustic Analysis

This comprehensive section explains the Lobanov method of vowel normalization developed by Boris Lobanoff in 1971. Students learn to normalize vowel measurements by converting raw scores to z-scores specific to each speaker, calculating means and standard deviations separately for F1 and F2. The process involves extracting individual speaker data, computing z-scores, storing normalized values in new columns using cbind(), and combining multiple speakers' data using rbind(). Students then compare raw versus normalized vowel plots to understand how normalization removes speaker-specific variability while preserving relative vowel positions.

A new vowel normalization method for sociophonetics replaces the mean with the centroid of the convex hull enclosing vowels and uses the standard deviation of convex hull points instead of all points, improving vowel space matching and overlap compared to traditional Lobanov normalization.

Articulatory data collection employs three primary methods: X-ray microbeam photography (gold pellets on vocal tract points), Electromagnetic Articulography (EMA, electromagnetic sensors tracking pellet positions/velocities), and real-time MRI for simultaneous speech-production data. Absolute X&Y positions vary with speaker anatomy, so they must be converted to relative constriction measures using geometric transformations computing distances between articulatory points. This normalization creates speaker-independent features essential for cross-speaker analysis and practical speech recognition applications.

This section explains individual variation in vowel formants and the need for normalization. The instructor explains that different speakers have different absolute formant frequencies because they have different vocal tract sizes, with women and children typically having higher formant frequencies than men because their vocal tracts are smaller. To compare vowel formants across different speakers, researchers normalize the formant frequencies by subtracting each speaker's mean F1 and F2 values and dividing by the standard deviation. This creates a normalized space where each speaker's vowel system is centered at (0,0), removing the effects of individual vocal tract size. The section explains perceptual normalization in speech perception: listeners do not use absolute formant frequencies to identify vowels; instead, they use the relative relationships between vowels within a speaker's system. The brain learns the typical vowel system of each speaker and identifies vowels based on their position relative to other vowels in that speaker's system.

Alexey Lobanov (1978-2020) was a talented Russian theater actor who spent 20 years at the Irkutsk Academic Drama Theater, creating diverse and nuanced roles by finding the unique atmosphere and psychological depth of each character; he was remembered by colleagues as a professional, sensitive, and soulful artist who tragically died at age 41 in 2000 while fishing on Lake Baikal.

This section covers creating vowel space plots in R by plotting F2 on the x-axis and F1 on the y-axis with reversed axes. Students learn to distinguish male and female vowel patterns, where males cluster lower due to longer vocal tracts. The process includes adding multiple groups to scatter plots using points() with different colors, creating legends, and understanding how vowel formant frequencies determine vowel identity. This foundational skill enables visual analysis of acoustic vowel data.

Vowels exist in a continuous acoustic space rather than as discrete categories, similar to how colors exist in a continuous spectrum; this vowel space is defined by formant frequencies (F1 and F2) which correspond to the resonating cavities in the vocal tract, with the three corner vowels representing maximally distinct mouth shapes (spread open and rounded), and modern technology allows for objective measurement and synthesis of vowels by measuring these formant frequencies rather than relying on subjective articulatory descriptions.

Plotting F1 and F2 values on a coordinate system reveals systematic vowel patterns. After flipping both axes vertically and horizontally, back vowels cluster on one side while front vowels occupy the opposite region, mirroring traditional vowel chart layouts. The horizontal axis represents F1 (corresponding to vowel height/tongue elevation), while the vertical axis represents F2 (corresponding to vowel advancement/tongue position). This visualization demonstrates that vowel charts reflect actual acoustic measurements rather than arbitrary conventions. The systematic arrangement of vowels in this space provides an objective, quantifiable representation of vowel quality based on physical properties of the vocal tract.

The vowel space is represented visually as a trapezoid because the jaw moves like a hinge, creating more space at the front and top than at the back and bottom. This two-dimensional representation allows linguists to plot all possible vowel sounds based on their tongue height and frontness/backness features. Since the vowel space is continuous, this diagram shows how different vowel qualities relate to each other in a systematic way.

Vowels are primarily distinguished by their first two formant frequencies (F1 and F2). F1 corresponds to vowel height—lower F1 indicates higher vowels, while higher F1 indicates lower vowels. F2 corresponds to front-back position—higher F2 indicates front vowels, lower F2 indicates back vowels. This creates an inverted coordinate system where high vowels plot at the top (low F1) and front vowels plot at the right (high F2). For a typical speaker with ~17.5 cm vocal tract length, schwa has F1 ≈ 500 Hz and F2 ≈ 1500 Hz, serving as a central reference point. This mapping allows prediction of vowel quality from formant measurements and vice versa.

When measuring formants for monopthongs (vowels that don't change with time position), you should not measure at the start or end of the vowel because these areas are affected by co-articulation effects. The initial consonant (like the 't' sound) affects the beginning of the vowel, and any following word affects the end. Instead, measure in the middle section where the formants are steady and appear as horizontal lines on the spectrogram.

This section demonstrates how to extract spectral slices from spectrograms to measure formant frequencies and explains coarticulation effects. The instructor shows how to use the spectral slice function in Praat to extract the spectrum at a specific point in time, explaining that this is useful for obtaining precise formant measurements. The instructor demonstrates extracting the spectrum from the middle of a vowel to avoid coarticulatory effects from neighboring sounds, explaining that coarticulation refers to the influence of neighboring sounds on the articulation of a sound. The instructor explains that consonants affect the formant frequencies of adjacent vowels, and that when measuring vowel formants, it is important to sample from the middle of the vowel to minimize coarticulatory effects. The section also discusses how different speakers have different absolute formant frequencies because they have different vocal tract sizes, with women and children typically having higher formant frequencies than men.

Vowels are not static in time and cannot be thoroughly described by spectral measurements taken at a single point. Even though vowel charts appear to show static formant positions, vowel dynamics are especially important in English and form an entire research area. Researchers can create vowel charts with static formant positions, but tracking the movement of those formants over time provides a clearer picture of vowel distinctions and helps distinguish between dialects.

A formant track is a newer and better way to do voice print analysis on a computer. It looks at the movement of frequencies and can say something about the articulatory flexibility during the sound recording. The analysis showed that the creature did not have the ability to make the 'e' sound, which requires moving the tongue forward in the mouth. Since gorillas have their neck at a different angle and cannot move their tongue forward to make the 'e' sound, this suggests the creature was more man than ape.

The FACE diphthong in Received Pronunciation does not move in a close front direction at all, according to Gimson's chart. It starts just below cardinal vowel 2 (which sounds like the DRESS vowel) and then the tongue raises and retracts, moving from e to ɪ (eɪ). This trajectory is quite strange compared to what one might expect for a front vowel diphthong. This demonstrates how diphthong trajectories can vary significantly across languages and accents.