Voicing your preferences

x Sponsored content on Research Live and in Impact magazine is editorially independent.
Find out more about advertising and sponsorship.

Attracting consumers to a new concept is no mean feat, especially in our highly competitive markets. Those that work appeal to people rationally and emotionally – and, while traditional concept testing has been good at identifying the rational, it was to tap into the emotional side that GfK developed MarketBuilder Voice.

Aimed at early and late-stage screening for consumer goods, it combines quant-size panels with qual insight, and uses technology combined with classification and consumer behaviour expertise.

Miriam Comber, GfK’s strategic innovation director, market opportunity and innovation, says: “We were looking for something that would describe more discrimination. When going back to a client, you want to know if there is a winning option – particularly when screening ideas. You don’t want fuzzy results, with three or four that all look the same. 

“The problem is that what we’re often asking in concept testing is how much they like something, and then whether they’re going to buy it – and there isn’t necessarily a correlation between the two. We always assume there is, but it’s not 100%. 

“We have a measure of sentiment – overall liking – and then a measure called activation that looks at whether people spontaneously say ‘I’d like to buy that or try that’. You’re looking for people to spontaneously say yes, rather than you suggesting it to them.”

Voice, adds Comber, allows you to get to that spontaneity. The methodology involved creating classification models using 5,700 voice recordings. To get a baseline, respondents complete a voice-calibration interview online – they are asked to read a sentence and describe a room from a picture they are shown – and are then recorded while giving their views on the concepts they are shown. By combining natural speech with modelling, sentiment and activation, emotional levels generated by the concept can be identified.

“Whatever they say about the concept is our target reaction,” says Comber. “There’s a lot of natural language processing going on to measure different things in voice. We were homing in on one measure – passion. The voice trace goes through the model that looks at passion. Then the transcript goes through two other models – one for sentiment and one for activation [people saying they’d buy it or recommend it]. Identifiers of passion are modulation, pitch, volume and flutter. It was tested on 1,000 features and it’s measured on 10-30.”

Transcription is done by a person, so accents aren’t a problem, but it will be automated further down the line if accuracy can be achieved.

The model was built as a collaboration between GfK Verein, its market research think tank, and the University of Passau, Germany. There have been issues to overcome with this more unfamiliar form of research, as well as with how panellists’ behaviour affects the testing of the new technique. “People are used to doing surveys, but they’re not normally asked to speak. We get completion rates of about 30%. If you’re the kind of person who does surveys at work, or sitting in front of the TV with your family [speaking is more problematic]. So, it’s not just the technology, it’s the circumstance.”

However, Comber thinks the rise of voice-controlled tech will make this form of market research easier. 

“It will change because so many things are coming out now that are voice. People will get used to randomly talking to things in a way that a lot of younger people are; it’s spreading through society.” 

This article was first published in Issue 21 of Impact.