What’s wrong with these formants?

Abstract

Speech is frequently modeled as an all-pole (autoregressive) linear filter applied to a voice source, commonly called Linear Predictive Coding. This model has been used for many years to provide measurements of the resonance frequencies (formants) characterizing speech sounds. That notwithstanding, it has also been well recognized that the use of this method to measure formants is very often fraught with problems, but just how serious are these problems, and can anything be done to improve upon LPC formant measurements? A few examples will demonstrate that the problems plaguing LPC formant measurements are too often intractable, and frequently yield formant values which are nowhere near the truth. For example, the voice bar will be shown to falsely lower the reported first formant value in many simple cases. It might even be suggested that reliance on this technique has impeded our understanding of exotic speech sounds which do not display a typical formant pattern. Precision formant measurements can be provided in most instances by reassigned spectrograms that have been specially processed to highlight signal components. Examples will be shown and used as benchmarks against which the LPC method can at last be properly evaluated.