Chapter Summary — Optimizing the Extraction of Vowel Formants

Skarnitzl, Vaňková & Bořil (2015)

@inbook{skarnitzl2015optimizing,
  author = {Skarnitzl, Radek and Vaňková, Jitka and Boril, Tomas},
  year = {2015},
  month = {06},
  pages = {165-182},
  title = {Optimizing the Extraction of Vowel Formants},
  booktitle = {Tackling the complexity in speech},
  editors = {Niebuhr, Oliver and Skarnitzl, Radek},
  isbn = {978-80-7308-558-2}
}

Overview

This chapter investigates how to improve automatic vowel formant extraction, focusing on whether the default settings in Praat and Snack/WaveSurfer are actually optimal. Although formants are widely used and easy to obtain with software defaults, the authors argue that accurate extraction is much more complex than it appears, especially when benchmarked against carefully established manual measurements.

The chapter has two main goals:

  1. Test multiple LPC/formant settings in Praat and Snack
  2. Compare Praat vs. Snack performance against manually labelled “ground truth” vowel formants

1) Building a “Ground Truth” Dataset

To evaluate software settings, the authors first created a manual reference dataset.

Dataset

  • 250 Czech vowel tokens
  • 5 short monophthongs: /ɪ ɛ a o u/

  • Produced by 10 native Czech speakers

    • 5 female
    • 5 male
  • F1–F3 measured at vowel midpoint
  • Two independent labellers measured every token

Labelling protocol

The authors used strict visual measurement guidelines:

  • Maximized Praat editor window

  • Frequency display:

    • 0–3 kHz for males
    • 0–3.5 kHz for females
  • 10 ms spectrogram window
  • Primary decision based on spectrogram darkness

  • Optional support:

    • Praat formant tracks
    • FFT spectrum view

Agreement results

Initial disagreement between labellers:

  • F1 MAD: 50.8 Hz
  • F2 MAD: 49.3 Hz
  • F3 MAD: 83.6 Hz

After collaborative correction of large disagreements:

  • F1: 29.0 Hz
  • F2: 29.1 Hz
  • F3: 32.9 Hz

This corrected average became the ground truth reference.

2) Why “Ground Truth” Is Still Uncertain

A major contribution of the chapter is showing that manual measurements themselves are noisy.

The authors identify four uncertainty sources:

A. Screen + mouse resolution

Because spectrogram display is pixel-limited, cursor values occur in discrete steps of ~24 Hz, even if the software shows 1 Hz precision.

B. Spectrogram resolution

With a 10 ms window, spectral resolution is inherently limited by the time-frequency tradeoff.

Estimated uncertainty contribution:

  • ~30 Hz

C. Estimating formant peaks from harmonics

Even when the vocal tract resonance is unchanged, small F0 changes can visually shift the apparent formant peak.

Observed effect:

  • Up to 80 Hz shift
  • Especially important for F1/F2 estimation

D. Between-labeller variability

Even after consensus corrections:

  • residual uncertainty ≈ 30 Hz

Combined uncertainty

The authors combine these into a 95% confidence uncertainty of ±54 Hz around any manually determined formant value.

This is one of the most important methodological insights in the chapter.

3) Praat Performance

The authors systematically tested Praat LPC settings.

Tested parameters

They varied:

  • LPC order: 8–12

  • Maximum formant frequency:

    • 5 kHz
    • 5.5 kHz

Main result

The default Praat setting already performs very well.

Slightly better alternatives were found:

  • Females: LPC9, 5 kHz
  • Males: LPC11, 5.5 kHz

However, the improvement over defaults was:

  • tiny for F1/F2
  • only somewhat larger for F3
  • still smaller than the ±54 Hz uncertainty band

So the authors conclude:

The practical improvement over Praat defaults is negligible.

4) Praat vs Snack

A major motivation was the widespread belief that Snack/WaveSurfer is better than Praat.

Result

The study found the opposite:

  • Praat consistently outperformed Snack
  • Especially strong advantage for F2
  • Snack showed much larger RMSD errors across F1–F3

This was a surprising result because informal phonetics practice often assumes Snack is superior.

Best Snack settings

Best Snack configurations differed by gender:

  • Females: autocorrelation
  • Males: covariance
  • 5 kHz range generally better than 5.5 kHz

Still, even the best Snack settings remained worse than Praat.

5) Praat Tracker vs Extractor

The authors also tested Praat’s formant tracker, which uses temporal continuity constraints (tracking trajectories rather than framewise extraction).

Result

  • Generally similar to Praat extraction
  • No consistent improvement
  • Worse for male F3

So tracking does not clearly outperform extraction.

6) Final Recommendations

Practical takeaway

For high-quality studio vowel recordings, the authors recommend:

  • Female speakers:

    • LPC order 9
    • max frequency 5 kHz

  • Male speakers:

    • LPC order 11
    • max frequency 5.5 kHz

But the bigger conclusion

Because manual uncertainty is already ±54 Hz:

Praat’s default settings are usually good enough in practice.

The gains from “optimized” settings are often not statistically meaningful.

Key Takeaways (Short Version)

  • Formant extraction is far less trivial than software defaults suggest
  • Manual “ground truth” itself carries ±54 Hz uncertainty
  • Praat defaults are robust

  • Slight LPC tuning can help:

    • females → LPC9 / 5 kHz
    • males → LPC11 / 5.5 kHz
  • Snack is worse than Praat on clean studio data
  • Praat tracker offers little benefit over extractor
  • F3 remains the most difficult formant to estimate reliably