January 13, 2005

Phonetic gaydar again

In response to an earlier Language Log post, Bert Vaux sent some information about work by a former student of his: Michael Schuler, " That's the gayest thing I've ever heard! The phonetic and phonological content of the perceptions of gay- and straight-sounding speech in English." MA thesis, Harvard University, 2003. I haven't seen a copy of Schuler's thesis, so I've based this post on a summary that Bert sent me.

Twenty male speakers of various sexual orientations made four recordings each: reading a list of words, reading from an economics textbook, acting out dialogue from a Fierstein play, and talking without a script about their freshman dorm. The speakers reported their sexual orientation, perhaps on the Kinsey scale.

Thirty listeners were asked to rate (selected parts of?) these passages on five seven-point scales: plain/dramatic, feminine/masculine, straight/gay, reserved/emotional, affected/ordinary. The speakers were also asked to rate their own voices on the same scales.

For all the forms of speech except for the word lists, the listener ratings on the straight/gay dimension correlated in a statistically-significant way with speaker self-reports of sexual orientation. The correlations were better for the acted passages than for the textbook reading, and better still for the free narrative. I assume (though I don't know) that the free narrative portions were chosen so that the content did not provide information about sexual orientation.

  Correlation between listener straight/gay rating
and speaker self-identification of sexual orientation
Word list
r=0.265 (p=0.304)
r=0.486 (p=0.048)
r=0.545 (p=0.024)
r=0.690 (p=0.002)

Schuler found that for speakers who were perceived as sounding "gay", listener ratings moved towards the "gay" end of the scale as the material moved from wordlists to textbook passages to play-acting to free narrative, while for speakers who were perceived as sounding "straight", listener ratings moved towards the "straight" end of the scale over the corresponding sequence of passages. He interprets this as suggesting that "both gay and straight speech are deviations from a mid point, rather than straight being the baseline".

Schuler also found that speakers' evaluations of their own voices were "inaccurate", meaning (I think) that they did not correlate well with listeners' evaluations.

Finally, (Bert reports that) Schuler found that most of perceptual effects in his study -- at least in the free narrative -- appeared to be due to pitch range, as opposed to vowel quality, consonant place, other characteristics that have been cited in other studies. However, I don't know any of the details about how he determined this -- I'll say more about this when I've seen a full copy. Meanwhile, here's his abstract:

The hypothesis underlying this research is that listeners can judge speakers based on phonological and phonetic features of their voice. Building on the work of Gaudio 1994, Crist 1997, and Rogers and Smyth (in press), data was collected to allow analysis of many features and many possible listener judgments; the results given in this paper are largely restricted to pitch features and listener judgments of sexual orientation. My findings verify Gaudio's claim that listeners largely can identify the sexual orientation of the speakers based solely on voice recordings, and that pitch properties in speakers' reading voices do not give strong indications to listeners whereby they can make accurate judgments of sexual orientation. The most important new finding is that gay-sounding speakers sound gayer when speaking freely than when reading and straight-sounding speakers sound straighter when speaking freely than when reading. This is important because it suggests that features that allow listeners to judge sexual orientation are more pronounced when speaking freely than when reading, and that the features are not necessarily absent when reading, just less extreme. This explains why the results of this study show many measures where pitch properties correlate with listener judgments of free response recordings while judgments of read passages show reduced or no correlation. Through this lens earlier literature can be reevaluated to the extent that small correlations based on read passages could reflect larger correlations that would exist in freer speech registers, if those registers had been tested. This presentation of the material uses extensive statistical analysis to draw its conclusions.

Bert sent audio samples of representative gay-sounding and straight-sounding speech from this study, but I don't know whether the speakers gave permission for their recordings to be published, so I'll try to find out before posting them.

The most important lesson that I'd like to draw from this study has nothing to do with perception of gender stereotypes in speech.

Experimental phonetics has become easy!

I don't mean to take anything away from Mike Schuler, who obviously put a lot of work into his project. But just a few years ago, he would have had to splice reel-to-reel audio tapes to make his test materials -- several times, in order to get randomized presentation orders. He would have had to spend many hours with expensive and cranky machinery in order to make phonetic measurements to correlate with listener judgments. He would have had to spend more long hours with a calculator in order to do the statistical analysis.

Today, everything can be done with an ordinary laptop and some free software.

[Update 1/21/2005: I've temporarily linked to a copy of (the body of) Michael Schuler's MA thesis, kindly provided via email by Michael himself, who is traveling in China. He wrote that he will put the thesis on the web himself, along with supporting materials including his recordings, when he gets back from his trip. ]


Posted by Mark Liberman at January 13, 2005 02:22 PM