Can you hear a smile?
Of course you can!
Can speech analytics hear a smile?
Now that’s a more complex question, and requires a blog post to answer.
How Emotion Detection in Speech Analytics works
Speech analytics is still very much a new technology, it is learning some new things and being integrated in interesting ways. Emotions in Speech Analytics so far has been somewhat problematic, however we can talk about three distinct methods currently used.
There are API’s out there that will process speech an determine the emotive state that the customer is in. This technology exists, and it uses fluctuation in tone and pace and some sampling to get really excellent outcomes for demonstrations. Is it accurate? We don’t think so.
This is problematic for a couple of reasons. Think about determining emotion based only off tone and pace. How different is angry and serene for some people? Or perhaps fear and exhilaration? Sad and relaxed?
Needless to say, we don’t use this method.
Speech to text
The second method uses speech to text. Using API’s like HP Idols sentiment analysis, it is possible to convert speech to text, then have that text analysed for words matching positive or negative sentiment and the API will tell you how accurate it thinks it is in a percentage format.
The problem with this option is it takes even more work to aggregate all this sentiment into something useful. At the end of the day, it’s the insights that allow for change that we’re interested in.
This method is the most simple, but I would argue that it is the most affective. Using existing speech analytics technology and methods for ‘deep diving’ and mapping out why customers call we can organise complaints and praise into buckets and get very specific and accurate with regards to where negative sentiment is coming from. By really scaling up analytic libraries it is possible to listen to what customers are saying on a large scale!