Elerian AI Builds Human-Like
AI-Powered Voicebots With Deepgram

No caller enjoys the experience of trying to talk to an automated customer support line and being told again and again by an eerily mechanical voice “I’m sorry, I didn’t catch that, could you repeat it?”

Customers of call centers using IVR-based call routing or outdated AI are regularly frustrated by a bot that has gotten stuck in a loop, or a bot sending them through the same cycle of limited options over and over due to misunderstandings and an inability to ask clarifying questions. If an AI is going to respond in a human way, it has to hear, remember, and understand in the way a human would. This of course means building better speech recognition and language understanding AI.

AI can increase the productivity of human agents by automating simple request calls like “What is my balance?” or “How do I access the app on my laptop?” An added bonus is that an AI also does not lose its patience or act rudely. With sentiment detection, an AI-powered bot can detect the customer’s emotional state and respond in an appropriate and personalized manner.

So if Elerian AI could build a Digital Agent that sounds and feels totally natural, they could provide a consistently excellent customer experience on par (and in some circumstances ahead of) what a human agent can provide. This frees up human agents to focus on the more challenging aspects of support or sales.

A machine can’t be taught to understand a word it doesn’t know, right? To build a machine that truly understands humans, it first has to be able to hear and transcribe the words humans say. So the first step to building a machine that can understand and respond to humans is building an accurate automatic speech recognition (ASR) engine. This is in and of itself a monumental task. The ASR engine that powers a Natural Language Understanding AI has to be accurate and it has to be fast so the Digital Agent can respond to the caller in real-time. It also has to be able to handle a wide array of accents and have a domain-specific vocabulary and context. For example, when it hears “bite” in a computing company, it knows that word is “byte”.

Elerian AI recognized the complexity of building an ASR engine for their customers where callers have a variety of dialects and accents. This was particularly evident in South Africa, where there are 11 official spoken languages. General ASR engines available in the market simply could not handle this complexity, nor recognize the requisite domain-specific speech entities critical to the contact center use cases, such as social security numbers and other identifiers. According to Craig Akal, Co-founder of Elerian AI, a suitable ASR solution is critical for the Elerian voicebot.

Deepgram’s trained and tailored speech models along with Elerian AI’s entity recognition models allow for speech to text precision transcription – at greater than 90% accuracy.

For Elerian AI, a general ASR speech model could not handle the various dialects, accents, jargon, and slang that they heard on phone calls, and they needed a more custom domain-specific model. Deepgram was the only solution that could provide a trained domain-specific model that would meet their stringent requirements. And Deepgram’s real-time transcripts are produced with less than a 300-millisecond lag, which is another mission-critical factor for a human-like conversational AI bot. Lastly, Deepgram can deploy its ASR solution on-premises for better security. This means sensitive customer data stays out of the cloud. For Elerian’s banking and financial services customers, this is crucial.