The output of the Speech Transcription Farm is JSON for an easy export of the data to existing BI systems. With this data, calls can be selected based on the subject or other call data.
More and more often we get the question: can you also say something about “emotion” of the caller? Are our customers satisfied after the call or not? And can you indicate whether things are going well during the conversation between the customer and the contact center employee or whether someone needs to intervene?
With language and speech technology, we can increasingly detect emotions in speech. The basic emotions of fear, anger and happiness are more or less the same in all cultures and are therefore easy to detect. It is more difficult to detect things like irony and sarcasm because they are more subtle and differ from culture to culture. But silences in a conversation, crosstalk and raising the voice are ’emotion’ markers that we can capture very well.
In addition to analysing the sound signal, speech recognition also allows us to recognise what has been said. In this way we know what is being said and how it is being said. In this way we can also see how long people pause before they pronounce the next word.
So we look at “what” someone says and “how” they say it. When things get out of hand, we see that people no longer allow each other to speak and thus start to speak through each other. They also start to talk louder and use certain words.
With the combination of all these measurement data, we can indicate quite accurately that the conversation may be at risk of derailment. We use the same approach to determine whether callers are “satisfied” or not.
What is the sentiment in the conversation? This is not very simple, but what if it turns out that the average that the computer calculates and that what people calculate using KTO scores and NPS measurements is close to each other? Are we going to question customers less often with a KTO or NPS measurement after a customer contact?
By combining the sentiment score of the conversation with the meta data, underlying customer contact processes and information about products, it is possible to train in a much more targeted way and to optimise specific processes.