Ann Bradlow - abradlow@northwestern.edu
Department of Linguistics
Northwestern University
2016 Sheridan Road
Evanston, IL 60208

Popular version of paper 4aSC1
Presented Thursday Morning, May 1, 2003
145th ASA Meeting, Nashville, TN

Imagine that you are travelling in a country where you have only a passing knowledge of the language.  You slip on a special speech enhancement device and suddenly the foreign-language speech around you seems so much easier to understand.   The device that you slipped on works by picking up the speech signal of a person you are talking to, enhancing the signal, and then delivering it to your ears through headphones. This is not a simple amplification device or a translation machine, rather it enhances the speech signal such that you can now more easily decipher the language of your current environment. For example, it makes word boundaries more salient and clarifies the pronunciation of all of the individual speech sounds.

Would the same device work for all languages or would you need a language-specific device? And, would the same device work for listeners with hearing impairments and for listeners in particularly noisy environments? By comparing the responses of different listener groups to naturally produced clear speech, our work has shown that the language experience of the listener plays an important role in determining speech intelligibility, suggesting that the optimal speech enhancement device will have to be customized to the needs of the listener and listening situation.
 

What is naturally produced clear speech?
When a talker believes that the listener is likely to have speech perception difficulties due to a hearing loss, the presence of some background noise, or a different native language background, she or he will typically adopt a "clear" speaking style that is distinct from her or his typical conversational speaking style. From careful acoustic comparisons of English conversational and clear speech samples, we know that English clear speech is typically slower and more precisely articulated than English conversational speech. Furthermore, these acoustic studies have demonstrated that the English clear speech enhancements reflect specific features of the sound structure of the language. For example, in English some vowels are always longer than others (e.g. the vowels in the words "feet, boot, and late" are always longer than the vowels in "fit, put, and let".) In clear speech, all vowels are lengthened (because of the slower rate of speech) but long vowels are lengthened more than short vowels. In other words, clear speech enhances the length contrast between English long and short vowels rather than just lengthening all vowels by the same amount.

Who benefits from clear speech?
Previous research has established that adult English listeners with either impaired or normal hearing typically find English clear speech more intelligible than English conversational speech. We wondered whether listeners with less advanced skills in the target language (in this case, American English) would get as great a speech perception benefit from clear speech as listeners with more advanced target language skills such as the adults tested in the earlier studies. We reasoned that since less advanced listeners (e.g. non-native listeners) may not be well practiced in decoding English speech signals, they may not know how to take advantage of all of the clear speech enhancements that talkers put into effect when producing clear speech. For example, if a listener is not already sensitive to the vowel length difference described above in many languages vowels do not differ in terms of length as they do in English - then the enhanced vowel length contrast of English clear speech will not be particularly helpful.

We hoped that finding out exactly who benefits from naturally produced clear speech would help delineate the features of clear speech that are most responsible for its enhanced intelligibility for specific listener populations.

A test of the clear speech benefit across populations.
In our work, we have been testing the conversational and clear speech perception abilities of various listener populations who vary in their language development. We tested children (8-12 years old), adults who are native speakers of English and adults who are non-native speakers of English. (The children were also native speakers of English).  Heres how the test works: A participant listens to sentences over headphones and then simply writes down what she or he hears. The sentences are all simple English sentences produced by a talker in either conversational or clear speech. To make the task challenging, the sentence recordings are mixed with noise; this also reflects real-world listening conditions better than perfectly "clean" recordings. When we do this test with kids, instead of writing down what they hear, they repeat the sentence to the experimenter.

Click here to hear examples of the stimuli we use in the clear speech test.

As you can see in the figure below, our results show that the size of the clear speech benefit varies across listener groups: the native listener adults showed a large clear speech benefit while the kids and non-native listeners showed smaller benefits. We interpret this as reflecting a difference in language development: the adult native listeners were best able to take advantage of the clear speech enhancements because they have had the most experience with spoken English and are therefore most sensitive to all of the subtle enhancement cues that the talkers put into the clear speech productions.
 

Conclusions and future directions.
Based on this cross-population comparison of the clear speech benefit we are beginning to see how clear speech intelligibility depends not only on  talker characteristics (previous research has shown that some talkers are better clear speech producers than others) but also on listener characteristics. Talkers can enhance the intelligibility of their speech by going into a clear speech mode, but the effectiveness of the enhancements depends on the experience of the listener. By characterizing both the talker and listener factors that contribute to speech intelligibility we may eventually be able to develop automated speech intelligibility enhancement strategies that will be optimized for a wide range of speech communication situations.

In parallel work, we are also investigating clear speech characteristics in other languages. This should help us delineate which clear speech features are universal and which are language specific, thereby bringing us closer to the ultimate goal of applying these findings to the development of automated speech enhancement devices that could convert a conversational speech signal into a highly intelligible, clear speech signal in a wide range of languages as well as a wide range of listening situations.