Sociophonetics: An Introduction

Special Feature: Tips on measuring vowels in poor-quality recordings.

7 wav files, 7 TextGrids - zip

Chapter 2 gives you some information about how to measure the formants with LPC, FFT spectra, and other methods, and chapter 5 shows you how to mark the onset and offset of vowels in common situations and where to take measurements within vowels. The examples given in the book are all from recordings that were made in quiet situations. The field recordings used for sociolinguistic studies and forensic linguistic applications often have a lot of noise in them, though. In addition, you may wish to analyze archival recordings that either have degraded over time or were made with equipment that didn't meet today's standards. Here are some tips on how to handle recordings like that.

One tactic that can be used for a wide range of recordings is helpful when you can't get a formant to track without getting a false formant at some other frequency. This often happens when two formants are close together, making them hard to separate in LPC. In a case like that, the easiest solution is to set the number of LPC coefficients to a great enough number that the problem formant does appear on the formant track. Then compare the formant track against the spectrogram and just ignore the false formant(s).

Sample 1 shows an example where the background noise is relatively constant. This kind of situation may occur when a fan or other machine, in this case a vending machine, is running in the room where the recording was made. Another problem in this recording is that the microphone wasn't close enough to the speaker's mouth. In spite of these problems, the vowel formants are often clear enough to read. The louder syllables certainly have readable vowel formants. As a rule, if you can understand the speech auditorily, you should be able to get usable readings from the recording. The softer syllables, particularly in function words, don't have clear formants. There's also a band of especially loud machine noise around 500 Hz. Formants that are close to this band of noise have to be scrutinized carefully. Note how LPC gives you an irregular formant track that follows the band of noise except where there's a sufficiently loud vowel. Anywhere that you see that irregular track, don't trust the readings for it. Thus, in the words if and enough, F₁ readings are unreliable. However, in the word she'll, F₁ looks straighter and can be trusted (for the most part), and in the words I and fast, F₁ is clearly away from the band of noise and readings of it are somewhat usable (though the formant track for F₁ is probably pulled down a little by the noise). Of course, phonetic phenomena of lower amplitudes, such as fricative spectra, and those that are highly sensitive to noise, such as voice quality, cannot be measured reliably in this recording at all.

Sample 2 shows an example in which the background noise, here vehicular traffic, comes and goes. As with sample 1, the microphone in sample 2 wasn't close enough to the subject's mouth. Even so, vowel formants in parts of the recording are quite readable. When the vehicular noise gets really loud, as when the subject says I have, the formants become hard or impossible to read. In a case like this, it's best to exclude vowels spoken during the loud noises.

Sample 3 exhibits a couple of problems. First, there's some sort of buzz, probably caused by a bad connection between the microphone and the recorder. Second, a telephone rings. These two problems affect the LPC formant track in different ways. The buzzing gives the formant track a saw-edge pattern where it occurs. Formants are clearly visible during the buzzing, but you should remember that you can't get much precision in the formant readings because the buzzing affects sound frequencies. Some such problems may distort the frequencies even more seriously, rendering any measurements unreliable. The telephone, which consists of a series of tones at steady frequencies, dominates the formant track in places. For the most part, the speech has a greater amplitude than the telephone, so you can get formant readings, but the ring tones probably pull the formant track toward them and thus they aren't completely reliable.

Sample 4 has two different species of birds in the background. The grackle, in the earlier part of the sample, has great enough amplitude that it obliterates the subject's higher formants. The lower formants can be measured reliably, though. The dove has much lower amplitude and doesn't dominate the formant track when the subject is speaking. The dove call may exert some pull on the formant track, but not much because it's so soft. For the most part, you can trust the subject's formants while the dove calls.

Sample 5 has other kinds of background noise. Here, people are talking and a radio is playing during an interview. You have to be especially careful with this sort of background noise. You can often distinguish bird vocalizations or mechanical noises from a human voice by their patterning, both on the spectrogram and in the formant track. Extra human voices, though, show formant tracks just like those of the interviewee, so you don't always know for sure whether a particular reading is from your subject or from somebody in the background. It's important to double-check, both by examining the spectrogram and by listening to the recording, so that you're absolutely certain that a measurement is from the right person. If in doubt, don't use the measurement.

Sample 6 shows a very different recording. This one was made in a quiet situation, but it's an old recording, originally made on an acetate disc in 1941. It's also been copied over a number of generations, and some of the copying may have involved amplification of certain frequencies. As you can see when you make a spectrogram of sample 6, frequencies below about 1000 Hz have relatively high amplitude, while those above 1000 Hz are faint. Compounding the problem is the speaker's high F₀. This recording is usable, but you may have to break a rule or two to get readings out of it. When you use LPC on this recording, you'll notice that you get too many formant readings below 1000 Hz and not enough above 1000 Hz. You could try to amplify the higher frequencies and damp the lower ones to equalize the amplitude. Another strategy is to use a fewer LPC coefficients for formants (mostly F1) below 1000 Hz and more coefficients for higher formants. This is usually a bad practice, but it's necessary for this recording. One other thing you can try is to change the upper limit of the LPC frequency range. It's crucially important to know where to expect the formants for each vowel so that you recognize when you have too many or too few formant readings.

Sample 7 shows part of the same interview as sample 6, but with filtering to correct the problems seen in sample 6: amplitudes below 1000 Hz have been damped and those above 1000 Hz have been amplified. LPC has an easier time with this version. Again, though, you have to know where to expect the formants.

Special Feature: Mergers

Determine how many phonological contrasts each of the following speakers makes from the lot, cloth, and thought classes and, if more than one contrasting class is involved, how the words are classified.

• Speaker 1: 27 wav files - zip
• Speaker 2: 83 wav files - zip
• Speaker 3: 57 wav files - zip
• Speaker 4: 42 wav files - zip
• Speaker 5: 69 wav files - zip
• Speaker 6: 76 wav files - zip

Additional thoughts

A fair amount of experimental study has examined factors that influence vowel duration and vowel reduction in recent years, and there are a number of studies that postdate those cited in the text. New factors have come to light in the process, including phonological neighborhood density (the number of words that are phonologically similar to the word being affected), the commonness of the word, and the number of times it is repeated in a situation. Smaller neighborhood densities, increased commonness, and more repetitions reduce the duration and increase the amount of reduction. Most of these studies have been conducted in laboratory situations and it's uncertain to what extent they play out in real-life conversations, though Smiljanic and Bradlow (2005) compared clear laboratory speech against conversational speech. Here are some of these studies:

There have also been experimental studies besides those cited in the book arguing for or against Exemplar Theory and for or against the related issue of underspecification. McLennan, Luce, and Charles-Luce (2003; see the reference in the book), argue against Exemplar models. Lahiri and Marslen-Wilson (1991) favor underspecification. Scott and Cutler (1984), Evans and Iverson (2004), Floccia et al. (2006), and Sumner and Samuel (2009; see the reference in the book), however, found evidence for improvement in perception of a dialect with greater exposure to a dialect. Similarly, Clopper and Pisoni (2004, 2006) found that listeners who had lived in more than one region performed better at dialect identification. Coupled with evidence from Munro et al. (1999; see the reference in the book) and Evans and Iverson (2007) that subtle changes occur in adults' speech production when they move to a new dialect area, this evidence would favor an Exemplar account. Some of these studies have employed the prime/probe, or prime/target, method in experiments. This method was briefly described in chapter 3. It involves playing two words, the first one called the prime and the second called the probe or target, and then asking subjects to react to the probe/target in some way, such as identifying whether it's a real word or not. The prime may be semantically related, phonologically similar, or unrelated and dissimilar to the probe/target. Note that the prime, though usually heard immediately before the probe/target in experiments, may be heard many minutes or even hours beforehand. In general, semantic relatedness shortens reaction times by listeners to the probe/target, while phonological similarity does not, and this information is used to examine cognitive processing.

Additional thoughts

Experimental research on how variation is cognitively processed has been increasing over the past few years. McLennan et al. (2003; see the reference in the book) and Pitt (2009; see the reference in the book) examined the tapping of intervocalic coronal stops in American English, and Gaskell and Marslen-Wilson (1996) and Gow (2002, 2003) examined assimilation in consonantal sequences. These studies, often measuring processing speeds and/or using a prime/target design, have tested whether lexical representations or phonological abstractions are involved in the cognition of speech, with mixed results.