AI Is Making Voice Chatbots More Realistic Than Ever

When customers call one of the many hotels and restaurants that now use voice tech startup PolyAI’s services, there’s a chance a digital host will answer. Far from the robotic monotone one might expect from an automated line, the voice will sound chatty and conversational. It will hesitate and inflect in all the right places and maybe even sport a regional accent. And it has the potential to grasp context and meaning beyond simple keywords. These vibrant personas are made possible by a confluence of machine learning technologies, which some experts say has the potential to shift how conversational artificial intelligence interacts with humans. As AI-generated synthetic voices make robots sound more natural, companies are also experimenting with how to tap language generators trained on massive amounts of text, such as research group OpenAI’s GPT-3, to make chatbots more versatile and conversational. “You start with a big model and then you do some customization, you can get quite a bit of functionality as a result,” Gartner Research analyst Bern Elliot said. “The difference between these models and prior neural networks is stark.” This could have implications for customer service departments, many of which are still short-staffed due to the pandemic and labor shortages. Small talk machines There is a high demand for new technology in the chatbot realm, according to a Deloitte report last year. The consultancy expects the conversational AI market to grow 22% per year, reaching $14 billion in 2025. The trouble with training an AI algorithm on virtually the entirety of all text on the internet—as models like GPT-3 have been—is it can make for a powerful ability to generate realistic-sounding sentences, but is also extremely unpredictable. Left to their own devices, these algorithms are prone to meandering into non-sequitur chunks of text, which AI researchers term “hallucinations,” or mixing up basic facts. According to PolyAI co-founder and CEO Nikola Mrkšić, his company has found a way to tame the more random tendencies of the large language models that power PolyAI’s voice assistants. But it ultimately took a labor shortage for some businesses to take a risk on the technology. “People who would have had concerns around uncontrollable neural networks are now implementing it and are very happy with it, because they just had no choice,” Mrkšić said. Language generation AI has a variety of potential uses when it comes to customer service and other business applications, according to Elliot. One is as a routing mechanism for incoming calls or voice queries. Its greater understanding of the context and meaning behind words allows it to more accurately cluster calls by topic. It can also power more comprehensive language translation services than the current norm. When it comes to chatbots, Elliot said, it can be especially useful for areas where adherence to specific facts doesn’t matter as much, such as generating small talk dialogue for games “because it's creating interesting, original chitchat based on a specific conversation that you're having," he added. Early days According to Elliot, some technologists are exploring how to switch chatbots between the free-ranging banter of a program like GPT-3, which can make them more personable, and more structured dialogue when accuracy is important. One such technologist is Bob Summers, CEO and founder of Goodcall, which offers a roster of AI personas to handle customer service on behalf of small businesses. He said his company is currently looking into whether GPT-3 could be useful for generating bespoke greetings or dialogue that would then be pre-vetted for the AI to use. “We tell GPT-3, ‘Hey, you're the host of a small restaurant in Boston. You need to greet someone on the phone,’ and it can generate some sense of content around that,” Summers said. “We know that for the product to get more conversational and a little more friendly, those models will be helpful.” Companies like PolyAI and Nuance—which Microsoft acquired in June—are among the first to use large language models for voice chatbots, and the trend is still in its earliest days, according to a Gartner report on the topic. As the capabilities around generating both realistic-sounding voices and dialogue with AI continue to advance, startup founders imagine generating unique automated characters for individual brands that can best serve the needs of their respective audiences. These personas would operate across customer service lines and voice-enabled devices. Mrkšić said he expected the dialogue to improve even more as consumers grow comfortable with this type of AI, which will in turn create more training data to shape the machine learning program's understanding. "We will get more data, and that will be like a virtuous cycle of getting it to impact our deep learning architectures," he said. "And then we'll get even more accurate."

AI News

How AI Is Making Voice Chatbots More Realistic Than Ever

Tech advances make bots more conversational, but controlling what they say remains a challenge

Patrick Kulp

About

Subscriptions

Events

Publications

WORK SMARTER - LEARN, GROW AND BE INSPIRED.

Subscribe today!

Patrick Kulp