August 07, 2006

Sex and speaking rate

After spending a couple of hours trying to track down a plausible source for Louann Brizendine's characterization of male and female speaking rates, I've come up empty. Worse than empty: the paper that she cites as the source for the claim contains no relevant information at all, and the rest of the literature on speaking rate not only fails to support her assertion, but also includes results that contradict it. If you can help me to do better, please let me know. Meanwhile, here's where I've gotten to.

On p. 36 of her new book The Female Brain, Prof. Brizendine writes: "Girls speak faster on average -- 250 words per minute versus 125 for typical males." In support of this assertion, the end-note in her book cites "Ryan 2000", which her biblography lists as Bruce P. Ryan, "Speaking rate, conversational speech acts, interruption, and linguistic complexity of 20 pre-school stuttering and non-stuttering children and their mothers", Clinical Linguistics & Phonetics, 14(1), pp. 25-51 (2000). Its abstract:

This is the second in a series of reports concerning stuttering pre-school children enrolled in a longitudinal study; the first was Ryan (1992). Conversational samples of 20 stuttering and 20 non-stuttering pre-school children and their mothers were analysed for speaking rate, conversational speech acts, interruption, and linguistic complexity. Between-group analyses revealed few differences between either the two children or two mother groups. Within-group analyses indicated differences that involved conversational speech acts and linguistic complexity. Most stuttering occurred on statements (M = 32.3% stuttered) and questions (M = 20.9% stuttered). Stuttered and disfluent sentences had higher Developmental Sentence Scoring (DSS) (Lee, 1974) scores (M = 10.9, 12.9, respectively) than fluent sentences (M = 7.6). Multiple correlation analyses indicated that speaking rate of mothers (0.561) and normal disfluency of children (0.396) were major predictor variables.

The only (new) quantitative information about speaking rates to be found in this paper is in Table 1:

A few numbers from other studies are quoted in the literature review, but none of them breaks speech rates down by sex, and none of the cited numbers are either 250 or 125. I'm at a loss to see how Prof. B. can interpret anything in this paper as support for the view that "girls speak ... 250 words per minute versus 125 for typical males". Perhaps her research assistant pulled the wrong index card for that talking point? In fact, the whole idea is so odd that I wonder if it's all a misunderstanding of some sort from the start. Perhaps someone read from a column of average pitch measurements -- where female values might be roughly twice (post-puberty) male values -- instead of a column of average speech rate measurements. On the other hand, there are no pitch measurements in Ryan 2000, so we'd have to combine this with a mistake in the reference. A less charitable interpretation would be that the 250-vs.-125 numbers arose through the same sort of intellectual "telephone game" that operated in the infamous Eskimo snow-vocabulary case, and a harried assistant assigned to track down a reference for the numbers just fixed on Ryan 2000 because "speaking rate" occurs in its title.

As far as I can tell, none of the literature that actually addresses the question of sex differences in speaking rates finds anything remotely resembling the disparity that Brizendine claims. For example, Michael P. Robb, Margaret A. Maclagan, and Yang Chen, "Speaking rates of American and New Zealand varieties of English", Clinical Linguistics & Phonetics, 18(1) pp. 1-15 (2004), says:

Various acoustic measures of speaking rate were calculated for 40 adult speakers of New Zealand English (NZE). These measures were then compared to a group of 40 adult speakers of American English (AE). Results of the analysis identified significantly faster overall speaking rate and articulation rate for the NZE group compared to the AE group. No gender differences in speaking rate or articulation rate were found for either variety of English.

There's a brief mention of some sex differences in speaking rate in a paper that Jiahong Yuan, Chris Cieri and I will be giving at ICSLP 2006 in September: Jiahong Yuan, Mark Liberman and Chris Cieri, "Towards an Integrated Understanding of Speaking Rate in Conversation". The link is to a four-page "extended abstract" that will go in the conference proceedings; given the four page limit, we cut our remarks on sex to

Males tend to speak faster than females, as shown in Figure 6. The difference between them is, however, very small, only about 4 to 5 words or characters per minute (2%), though it is statistically significant. It might be due to things that we would not normally think of as speech-rate parameters, such as differences in word-frequency distributions. The opposite patterns of segment-length difference between males and females in Chinese and English are interesting, and need more study.

If you're curious, this is something that you can easily look into yourself -- just snag some of the speech floating around on the internet, transcribe it (or find a transcript made by someone else), count the words, and divide by the elapsed time. This can be a great topic for a research paper, because the data is easy to get, and the measure can be relevant to all sorts of things: rhetoric (how do speech-rate changes track the arc of an argument?), mood (how do actors use speech rate to convey emotion?), gender studies (how do male and female speech rates vary across circumstances?), or comparisons across age, social class, language or interactional context. There are a couple of tricks, though, which mean that you have to be careful about comparing rates across sources and contexts.

One is the question of whether and how to count silences. This came up a couple of years ago on Language Log, in a discussion of one of the 2004 presidential debates ("The Rhetoric of Silence", 10/3/2004):

I've pointed out that in Thursday's debate, John Kerry's sentences were 17.7% longer than George Bush's. Since the two men had the same amount of time to speak, you expect this to mean that Kerry used fewer sentences. And he did, 468 to 476. However, that's only a 1.7% difference. Kerry accounted for most of his greater sentence length not by using fewer sentences, but by packing more words into the same amount of time. 15.8% more, 7,136 to 6,165.

How'd he do that? New LexicoTardis® technology from the Rockridge Institute? Well, there are four obvious possibilities. First, Kerry might have talked faster. Second, he might have used shorter pauses. Third, he might have paused less often. Fourth, he might have used intrinsically shorter words.

A few quick and simple measurements suggest that the second of these four was the key factor. In the section of the debate that I examined, Bush used about the same number and frequency of pauses as Kerry did, but Bush's pauses were much longer. In between the pauses, Bush actually talked faster, but the pauses were so much longer that his overall speech rate was slower.

[...] Bush's overall speech rate was slower (155 words per minute vs. 167 words per minute), but while the two men were actually talking, Bush talked considerably faster (220 words per minute vs. 202 words per minute).

If you're going to leave out the pauses, you have to decide what a pause is. Are you going to leave out all the little silences, no matter how small? The littlest ones aren't really pauses, they're aspects of speech like "stop gaps", like the tiny silence while your mouth is closed during the [p] of apple -- but these are variably prolonged at phrase boundaries, and you'd need to decide when a prolonged stop gap turns into a silent pause. What about "filled pauses", hesitations that the speaker fills with sounds like "uh"?

Luckily, these decisions don't make a big difference in speech rate numbers. The big effect is leaving out pauses or not leaving them out. But if you're going to take them out, then different practices about what to remove will still create a small effect, and so you'd better do it the same way across all the conditions you're interested in.

Another issue comes up in dialogue and multi-party conversations. What are you going to do about overlapping speech? This can create a significant effect as well. Here's a quote from the previously-cited Yuan, Liberman and Cieri paper:

One indication of the nature of this problem can be seen by exploring various definitions of speaking rate applied to a modest-sized corpus that has been carefully aligned at the word level. This is the version of English Switchboard corrected and aligned at ICSI, comprising 2,438 conversations. If we calculate the overall speaking rate for this corpus, by simply adding up the number of words spoken by both participants, and dividing by the total elapsed time of the conversations, we find an overall average rate of 196 words per minute (WPM). For individual conversations, the rate measured in this way ranges from 111 WPM to 291 WPM. If we use the word alignments to exclude silences, non-speech noises and so on, we find an average “net” speaking rate of 236 WPM, with a minimum of 158 WPM and a maximum of 312 WPM. However, if we calculate the rate by adding up all of the turns for each speaker in order to get the total time, we find that there is so much overlap in the turn boundaries that the average turn-wise rate (total word count divided by the sum of segment times) is only 164 WPM, or 14% less than the rate calculated using the total conversational time.

For the "Switchboard" conversations, we thus found average rates of 196 WPM by one method (total words divided by total conversation time), 236 WPM by a second (total words divided by speech time excluding silent pauses), and 164 WPM by a third (total words for each speaker divided by the time allotted to that speaker's "turns", which included some short silences and some speech from the other parties). In my earlier post, I reported on average rates for males and females in a different corpus of conversational English, using the third method, but relying on turn boundaries established by a different crew of transcriptionists. This method produced average rates of 174.3 WPM for men and 172.6 WPM for the women. I'm fairly confident that this reflects different turn-segmentation practices by the transcriptionists, not a 6% faster speech rate in the second corpus.

There's no "standard" way to do all this, as far as I know. Just decide what to do, based on the goals of your study and the data you're working with, try to keep the practices consistent across the variables of interest to you, and document your decisions when you write it up. The one thing you shouldn't do? Don't just make up a couple of numbers claiming a 2-to-1 difference for the groups of people that you're interested in characterizing. Somebody might check.

[Update -- Lameen Souag points us to Ann Cutler and Donia Scott, "Speaker Sex and Perceived Apportionment of Talk", Applied Psycholinguistics 11 (1990): 253-272. It's not available on line, alas, but the description at ERIC reads:

Investigates whether listener bias contributes to the mistaken notion that women talk more than men. Perceptual effects (misjudgments of rates of speech) and attitudes to social roles and perception of power relations are suggested to be among the factors contributing to the misjudgment.

Lameen's description, based on reading about the paper in Jay Ingram's Talk Talk Talk:

Apparently, they took various dialogs and, for each dialog, recorded it being read by two women, by two men, by a man and a woman, or by a woman and a man. When a man and a man, or a woman and a woman, are reading the dialog, listeners judged each of them to be taking about equal time; when a man and a woman are reading the same dialog, listeners (male or female) consistently judged the woman to be talking more.


[ Update #2 -- speech rate information for females and males in Dutch conversation can be found in Diana Binnenpoorte, Christophe Van Bael, Els den Os and Lou Boves, "Gender in Everyday Speech and Language: A Corpus-based Study", Interspeech 2005. Their data is from 50 male and 58 females who participated in face-to-face conversations, and 40 males and 61 females who participated in telephone dialogues. Speech rate measurements can be found in their Table 1:

Converting from words per second to words per minute, that's 223 WPM for the males vs. 220 WPM for the females with pauses included, and 274 WPM for males vs. 266 WPM for females if pauses are excluded. Again, a (meaninglessly small) advantage for the males, in the opposite direction from Brizendine's claim. (Hat tip to Theo Vosse.)]

[Update #3: Paul Foulkes reminds me that in "Relations of sex and dialect to reduction". Speech Communication 15: 39-54 (1994), Dani Byrd checked the 630 speakers who read the two "calibration sentences" for the TIMIT database, and found that the male speakers were 6.2% faster on average. Though small, the effect was found to be statistically significant -- and again in the opposite direction from Brizendine's claim. ]

Posted by Mark Liberman at August 7, 2006 10:05 AM