Long wait times. Confusing menu options. Insidious hold music. And after all that, you reach a customer service agent who can’t give you the help you need. No wonder we hate customer service.
A San Francisco startup founded by particle physics researchers is working to ease that agony, with help from AI and GPUs. The company, Deepgram, created technology that businesses can use to quickly assess customer calls to improve service.
“You really don’t want to be calling customer service, and you don’t want your time wasted,” said Scott Stephenson, Deepgram co-founder and CEO. “We want to help customers accomplish their goals more quickly.”
Why Call Centers are So Bad
There’s a lot at stake in customer service. Businesses that increase the number of customer calls that get resolved on the first try by just 1 percent save an average of $276,000 a year, according to a study by SQM Group, a customer service consulting firm. Customers whose problems aren’t resolved in one call are eight times more likely to leave, the study found.
“Companies really want to know what’s going on in their customer calls, but they haven’t had a cost-effective way to do this,” Stephenson said.
Deepgram uses GPU-accelerated deep learning to make speech searchable and detect patterns that show where calls went right or wrong. That now requires someone to listen to and evaluate recorded calls. Because it’s so labor-intensive, it only happens on only one call in a 100, Stephenson said.
Alternately, companies convert speech to text for review, but that’s plagued by problems with audio quality and errors, he said.
Keep the Customer Satisfied
Deepgram’s deep learning software allows businesses to check the quality of service on every call. By detecting patterns of sounds and phrases, it lets companies find out things like what upsets or pleases customers, whether call center agents use the right words, and what topics commonly emerge, Stephenson said.
Businesses can use this information to improve scripts and training for call center agents, identify sales leads, check compliance with regulations, and reduce the number of customers who defect to competitors, Stephenson said.
One Deepgram customer increased revenue by 3 percent by using the company’s technology, he said.
From Dark Matter to Big Data
Stephenson and co-founder Noah Shutty were in China researching dark matter for the University of Michigan when they came up with the idea for Deepgram. Shutty had recorded thousands of hours of his life, and he needed a way to find certain moments.
After solving Shutty’s problem, they realized their method could help businesses, too.
The pair trained their neural network on thousands of hours of audio largely made up of recorded customer calls. They created an open source deep learning framework called Kur, which encompasses the CUDA parallel computing platform and the TensorFlow deep learning framework. To accelerate training, they used GeForce GTX 1080 GPUs with cuDNN.
“We wouldn’t be able to do what we’re doing without GPUs,” said Stephenson.