Linguistics Seminar Series presents do we need formants in speech? by Dr. Michael Keifte, Dalhousie University.

Throughout the history of phonetics and phonology, formants have been used to describe a wide range of phenomena in both speech production and perception. There are many advantages of focusing on formants: they arise naturally from resonances of the human vocal tract; they are manipulated directly in speech through the movement of speech articulators; they are easily analyzed from spectrograms; they have been used to unify theories of production and perception; and they can be used as parameters in speech synthesis. However, we know very little regarding the relationship between formants and how listeners actually understand naturally produced (i.e., spontaneous) speech. Although formant measurement may appear to be a simple process, models of speech perception based on formant frequencies fail to account for a large amount of perception data. The reality is that what we see on spectrograms may not be a good representation of what listeners actually hear, and there is very little evidence that there is any kind of formant-based representation of speech sounds in the brain. Further, virtually no speech-technology application (e.g. speech recognition systems such as Siri) measures formants at all. In this talk, I present the results of several experiments that test the limits of listeners’ ability to hear formants or resonances in vowel-like contexts. This research ranges from psychoacoustic studies in which formant-like sounds are almost completely decontextualized from speech to the modelling of formant transitions in recorded continuous (i.e., not in the laboratory) speech. These studies present a strong case that listeners do rely on formants in speech perception but that traditional ways of determining formant frequencies may not correspond well to what is happening in human speech perception.