Formants are easy to measure; resonances, not so much: Lessons from Klatt (1986)

Area of Science:

• Acoustics
• Speech Science
• Signal Processing

Background:

• Formants are typically identified as spectral maxima, but this definition differs from the actual vocal tract resonances of interest.
• Existing analysis methods, including linear predictive coding, are susceptible to 'harmonic attraction,' where formants are biased towards strong harmonics.
• This bias affects both automated algorithms and manual measurements, leading to systematic errors in speech analysis.

Purpose of the Study:

• To highlight the challenges in accurately measuring vocal tract resonances in speech signals.
• To discuss the phenomenon of harmonic attraction and its impact on formant estimation.
• To recommend best practices for speech analysis in the absence of fully automated, accurate techniques.

Main Methods:

• Review of established speech analysis techniques and their limitations.
• Discussion of the 'harmonic attraction' phenomenon as described by Klatt (1986).
• Comparison of reassigned spectrograms with linear predictive coding regarding error reduction.

Main Results:

• Formant estimates are often biased by harmonics, unlike human auditory perception.
• Harmonic attraction can lead to systematic errors in formant frequency measurements.
• While reassigned spectrograms show promise in reducing errors, they lack satisfactory automation for widespread use.

Conclusions:

• Current speech analysis methods have inherent limitations due to the influence of fundamental frequency (F0) and spectral granularity.
• Researchers should fully report and justify analysis settings used.
• Examination of the relationship between F0 and the first formant (F1) is recommended to identify potential harmonic bias.