I wanted to get everybody's reaction to some weird results that have to do with stress production and comprehension in English, as well as the use of stress cues for learning word segmentation. From a fairly atheoretic perspective, a syllable is stressed if it is perceived as more prominent than others. There are lots of things that can go into this perception, depending on the language:
- Suprasegmental acoustic cues: stressed syllables can be longer, louder, and have shallower spectral tilt than other syllables.
- Correlated prosodic events: stressed syllables can attract pitch accents, and the associated f0 events.
- Phonological cues: heavy syllables (those with a long vowel or coda) often attract stress.
- Lexical knowledge: individual words or classes of words may have predictable stress. For example, nouns in English tend to have initial stress while verbs have final stress (e.g. RECord v. reCORD).
Of course, these are all going to be massively correlated with each other, and the details vary from language to language.
I've come across some results recently that suggest that the acoustic cues are not important for English stress perception. First, adult English speakers appear to ignore suprasegmental cues.
Cutler (2014) compared how adult speakers of English, Spanish and Dutch perceived unstressed and stressed syllables that had been spliced to create a stress mismatch. For example, she took the first syllable of muSEum, and spliced it into replace the first syllable of MUsic, so the manipulated material did not have the suprasegmental stress cues. She found that Spanish- and Dutch-speaking listeners performed worse on a phoneme monitoring task with materials that had been corrupted in this way, but English-speaking listeners performed about the same. Her explanation appeals to the structure of the lexicon -- in English, there are very few pairs of words with identical prefixes except for the stress, but there are apparently many in Dutch and Spanish.
So this result suggests that adult English listeners ignore suprasegmental cues because they are not useful, given the English lexicon. We might think that the role of suprasegmental cues instead is to help with the acquisition of words. To learn words, infants must break a continuous sequence of phonemes into potential lexical forms.
Cutler and Carter (1988) famously pointed out that most words in English are stress-initial, which should help learning: when you hear a stressed syllable, you can be reasonably confident that it is at the beginning of a word. Subsequent computational work has found that stress is useful for statistical learners by annotating the input with stress cues from a dictionary.
However, there are reasons to think that the important components of stress for word segmentation in English are
phonological, not
suprasegmental. This past summer, I had a
computational modeling paper with Mark Johnson that compared the utility of observing stress from a dictionary with the utility of looking for syllable weight regularities. In short, the result is that looking for syllable weight regularities provides all the information about word boundaries that observing dictionary-annotated stress provides.
Moreover, there is some indication that suprasegmental cues to stress in Infant Directed Speech (IDS) are actually
weaker than in Adult Directed Speech (ADS).
Lee et al (2014) compared suprasegmental cues to stress in Australian corpora of English IDS and ADS, and found that no cue was stronger, and some were weaker (the extra swoopiness of IDS served as noise, rather than to strengthen the signal).
I have some preliminary results that suggest a similar pattern holds in American English. The comparison here isn't perfect (which is why I say preliminary): I'm using the official phonetic alignment of the Switchboard corpus of telephone speech for my ADS corpus, and my own forced-alignment of the Alex portion of the Providence corpus for my IDS corpus. To see how useful one important suprasegmental cue (log vowel duration) is, I plotted the
Receiver Operating Characteristic (ROC) curve of a logistic classifier that predicts the stress status (primary stress or no stress) of a vowel based on the vowel identity and its log duration. The output of the logistic classifier is the log-odds that a vowel is stressed, and the ROC curve is created by gradually dropping the threshold for declaring a vowel stressed. As you drop the threshold, you accept more false positives, but an accurate classifier should provide more true positives more quickly. A perfect classifier provides all the true positives before providing any false positives, and a random classifier provides true positives at the same rate as it provides false positives. An ROC curve plots true positives against false positives, so a perfect ROC curve has area one, and a random ROC curve lies on the diagonal.
For the IDS data (the Alex transcripts of the Providence corpus), the ROC curve is:
But for the ADS data (the Switchboard corpus), the ROC curve is:
So we can see that classification of stress status on the basis of vowel duration and identity is much better on the ADS dataset.
So, in sum, adult English listeners ignore suprasegmental cues to stress, we have evidence that suprasegmental cues to stress are degraded in English IDS compared to English ADS, and the utility of dictionary stress for word segmentation in English is completely redundant with syllable weight. Why do adult talkers produce suprasegmental acoustic cues to stress if adult listeners don't use them and infant learners can get the same information elsewhere? I have some speculations, but I'll wait to hear from the hivemind first.
