Used to be a DIME, now a DOLLAR
April 2024
Alexina Whitley
Does the marksman always take “careful AIM” or, in recent times, do they sometimes take “careful NOTES” instead? In re-examining statements from the 80s with listeners from today, this is just one of the differences we found from the Revised – Speech in Noise (R-SPiN) corpus (Bilger et al., 1984).
What is the Revised - Speech in Noise corpus and how is it used?
In typical conversations, people often make predictions about what someone is going to say and when they’ll say it, which helps them to follow conversation and prepare responses in good time. One type of information that people can get from a speech signal to help make predictions is semantic context – the meaning of the words and sentences and the way they relate to one another. To test whether semantic context is used to predict upcoming words, studies typically use a sentence-final key word paradigm. This paradigm involves presenting participants with sentences where the final ‘target’ word is either predictable or unpredictable based on the amount of context provided by the preceding words.
The R-SPiN corpus has been widely used for decades across the field of hearing research. It contains 200 pairs of high and low context sentences matched on final target words. For instance, the word “SPOON” appears both in the high-context sentence “Stir the coffee with a SPOON” as well as the low-context sentence “Bob could have known about the SPOON”. The beginning of the high-context sentence provides contextual information that narrows the range of possible upcoming words, whereas the beginning of the low-context sentence does not. The structure of these sentences therefore allows researchers to examine the influence of semantic context in predicting and understanding upcoming speech.
Shortcomings of the R-SPiN
Whilst the R-SPiN is a useful tool for evaluating context use it has its limitations. If we want to keep using these R-SPiN sentences, it is important to assess how relevant they are in the modern day.
First, R-SPiN sentences were developed nearly 40 years ago, with the final target words being selected from the 30, 000 most frequently used words before 1952 (Kalikow et al., 1977). It is likely that words that were once frequently used are no longer so popular – are listeners today really going to predict “CROOK”?
Second, the original stimuli lists were developed with a specific population. Participants were American listeners, and, as you can imagine, there are likely to be differences between American and English word usage. Similarly, the stimuli were developed with young adults, yet expected final words may differ depending on age. This is particularly problematic if researchers intend to use these stimuli to study age differences.
Finally, and most critically, some sentences may simply no longer be sufficiently predictable. What do we mean by this? For some high context sentences participants will all respond with the same final word, yet for other sentences participants may give multiple responses without a clear preference. If no one response emerges as the most common, the semantic context may still not ensure ‘predictability’.
Testing R-SPiN sentences
We presented R-SPiN stimuli as text with final word removed to 240 older and younger participants from the UK and US. Participants were asked to complete each sentence using only one word. As a measure of predictability, we calculated the number of unique final word responses for each sentence and the number of participants who gave each response. For a final word to be regarded as predictable we required 67% of responses to be the same final word.
Our results showed that many R-SPiN sentences were not predictable (i.e., less than 67% of participants completed the sentence with the same word). Moreover, more of the 200 sentences were assessed as highly predictable for older adults (United Kingdom: 109; United States: 107) than for younger adults (United Kingdom: 75; United States: 81).
Importantly, the most common words chosen to complete each sentence also varied between cultural groups. For example, for the incomplete sentence “Ruth poured herself a cup of”, most Brits responded with “TEA”, however Americans were pretty evenly split between “TEA” and “COFFEE” (though a select few of both nationalities had more alcoholic beverages in mind).
Future directions
This work is crucial as it provides a better idea of the words people anticipate when listening to R-SPiN sentences. Based on these findings we can update the stimuli with final targets that are more appropriate for the population we are studying. As researchers, making sure we use appropriate sentences for our specific participants is essential to ensure that the differences we find are meaningful. We are now beginning a series of studies using these R-SPiN sentences, specifically those most appropriate for our British older adult participants, to assess how prediction is affected by hearing loss. And we won’t be asking about how much they can buy for a DIME!