Is that an Interesting conversation?

Project Lead: Anmol Madan

Being able to measure interest and engagement using cell phones or PDAs in conversation would be an asset for focus groups, travel industry, call centers, online dating, marketing and sales training, etc. We think that this may be possible by measuring activity, stress, mirroring behavior and engagement in a person's speaking style.

In this study, we measured speaking style for 200 3-minute conversations, about randomly chosen topics. Each session consisted of 10 successive conversations between 2 participants (of the same sex). Participant's rated thier interest in each short conversation on a scale of 1-10.

Male Interest

For the men, using all features we can predict 44% of the variance (r=0.66, p=0.02). Using the stress and activity measures alone (for both men), we can predict about 35% of the variance (r=0.59, p=0.01). The activity features predict 27% of the variance (r=0.52, p =0.007). The influence feature and back-and-forth features by themselves did not produce statistically significant correlations.

In addition to the actual ratings, we also calculated the mean rating for each male over the 10 sessions. The mean rating for all 10 successive conversations, in a sense, represents factors other than the topic itself, and we call it the ‘mood + compatibility’ factor. For men, the correlation of speech features with this mean rating was slightly higher than with their actual ratings (r=0.7144, p=0.003) – implying that it is possible to predict their ‘mood + compatibility’ for these conversations slightly more accurately than their overall interest in that conversation. Stress and activity from the person himself showed a high correlation with this factor (r=0.56, p=0.0001).

The distribution of male ratings can be split into a two-class model – we labeled them as high interest (rating >=8) and low interest (rating < 8), marked in red and blue in the figure. For this 2-class model, the speech features explain 47% (r=0.69, p=0.01). Stress and activity features for both people also show a high correlation (r=0.595, p=0.01). Using the activity features (speaking time and voicing rate for both men) in a linear SVM classifier, it is possible to classify samples into these two classes (red and blue) with 74% accuracy.

Female Interest

The speech features are again highly correlated with the interest answers for women, and explain 45% of the variance (r=0.67, p=0.006). The stress and activity measures (for both women) predict just over 35% of the variance (r=0.6, p=0.004) of her interest answers. Unlike the male responses, the influence/engagement measures (r=0.46, p=0.05) and only her stress and activity measures (r=0.48, p=0.006) also played an important role.

Similar to the men, we also calculated the mean rating for each woman over the 10 sessions, and labeled it the ‘compatibility + mood’ factor. The speech features showed significant correlation with this mean (r=0.631, p=0.03). Stress and activity for both women (r=0.6, p=0.004), her stress and activity measures alone (r=0.53, p=0.001) and the engagement measures (r=0.36, p=0.06) also showed significant correlation.

Similar to men, the distribution can be divided into classes of low interest (blue, answer < 7) and high interest (red, answer >=7). All features together explain 42.5% of the variance (r=0.65,p=0.01) and the stress and activity measures explain about 35% of the variance (r=59,p=0.005).

Conclusion

The speech features explain about 45% of the variance in self-reported interest ratings on a scale of 1-10 for both men and women. Interestingly, the stress and activity measures alone explain about 30-35% of the variance in ratings. This shows that our speech features capture at least some of the verbal body language and social signaling that people use when they are interested in a short conversation.