Acoustic Phonetics

Chapters 2 to 5 described speech from the speaker's side: where the tongue is, what the larynx does. This chapter describes it from the sound's side. A listener, a microphone or a program has no access to the tongue; it has only the sound wave. Acoustic phonetics measures that wave and links what it finds back to articulation.

What Acoustic Phonetics Studies

Wikipedia describes acoustic phonetics as the subfield that studies the acoustic properties of speech sounds. It looks at waveforms in the time domain (such as duration, amplitude and fundamental frequency), in the frequency domain, or both together, and links these to articulatory and auditory phonetics and to abstract units such as phonemes.

The links in the third column are my own bridging to earlier chapters. The Wikipedia page for acoustic phonetics itself names duration, amplitude and fundamental frequency but explains little more, so the later sections rely on its pages for formants, voice onset time and Praat.

Formants

A formant is a broad peak in the spectrum of a sound, produced by acoustic resonance of the human vocal tract. Formants shape the sound of the voice but are not sound sources themselves. The first two, F1 and F2, mainly determine vowel quality:

• F1 is higher for open (low) vowels like [a] and lower for close (high) vowels like [i] and [u]. So F1 tracks height (Chapter 5).
• F2 is higher for front vowels like [i] and lower for back vowels like [u]. So F2 tracks backness.
• Lip rounding tends to lower F1 and F2 in back vowels.

Typical values

Wikipedia gives these averages for a male voice, in hertz:

Read the numbers against the rule. [i] and [u] are both close, so both have a low F1 (240 and 250). [a] is open, so its F1 is much higher (850). [i] is front, so its F2 is high (2400); [u] is back and rounded, so its F2 is low (595). Formants are estimated from a spectrogram or spectrum analyser, or by linear predictive coding (LPC), which estimates the vocal tract's resonances from a recording.

Voice Onset Time

Voice onset time (VOT) is "the length of time that passes between the release of a stop consonant and the onset of voicing." It is the acoustic measure behind the aspiration diacritic of Chapter 6.

• Longer VOT means stronger aspiration. Navajo [kʰ] is about 160 ms, about twice English [kʰ] at about 80 ms; Navajo unaspirated [k] is about 45 ms.
• For velar stops, about 20-30 ms is unaspirated, 50-60 ms weakly aspirated, 80-90 ms moderately aspirated, and over 100 ms strongly aspirated.
• English initial /b d ɡ/ are partially voiced; initial /p t k/ are aspirated with positive VOT. In the page's figure, /t/ in tie has a VOT of 95 ms and /d/ in die 25 ms.

Praat

Praat is free, open-source software for analysing and synthesising speech, created by Paul Boersma and David Weenink at the University of Amsterdam and first released in 1992. It is licensed under GPL-3.0-or-later and runs on Windows, Linux, macOS, FreeBSD and Solaris. It is downloaded from praat.org (the source is on GitHub at github.com/praat/praat). It can:

• extract vowel formants and prosodic details such as pitch and intonation;
• display spectrograms and waveforms, which show voicing and whether a given segment is present;
• manipulate, synthesise and annotate sounds;
• run scripts through a built-in GUI scripting language.

A first session (my suggested workflow)

Try It Yourself