You’re headed to your favorite drive-thru to grab fries and a cheeseburger. It’s a simple order and as you pull in you notice there isn’t much of a line. What could possibly go wrong? Plenty.

The restaurant is near a busy freeway with roaring traffic noise and airplanes fly low overhead as they approach the nearby airport. It’s windy. The stereo is blasting in the car behind you and the customer in the next lane is trying to order at the same time as you. The cacophony would challenge even the most experienced human order taker.

With IBM watsonx™ Orders, we have created an AI-powered voice agent to take drive-thru orders without human intervention. The product uses bleeding edge technology to isolate and understand the human voice in noisy conditions while simultaneously supporting a natural, free-flowing conversation between the customer placing the order and the voice agent.

Watsonx Orders understands speech and delivers orders

IBM watsonx Orders begins the process when it detects a vehicle pulling up to the speaker post. It greets customers and asks what they’d like to order. It then listens to process incoming audio and isolates the human voice. From that, it detects the order and the items, then shows the customer what it heard on the digital menu board. If the customer says everything looks right, watsonx Orders sends the order to the point of sale and the kitchen. Finally, the kitchen prepares the food. The full ordering process is shown in the figure below:

http://www.w3.org/2000/svg" version="1.1"/>" />

There are three parts to understanding a customer order. The first part is isolating the human voice and ignoring conflicting environmental sounds. The second part is then understanding speech, including the complexity of accents, colloquialisms, emotions and misstatements. Finally, the third part is translating speech data into an action that reflects customer intent.

Isolating the human voice

When you call your bank or utilities company, a voice agent chatbot probably answers the call first to ask why you’re calling. That chatbot is expecting relatively quiet audio from a phone with little to no background noise.

In the drive-thru, there will always be background noise. No matter how good the audio hardware is, human voices can be drowned out by loud noises, such as a passing train horn.

As watsonx Orders captures audio in real time, it uses machine-learning techniques to perform digital noise and echo cancellation. It ignores noises from wind, rain, highway traffic and airports. Other noise challenges include unexpected background noise and cross-talk, where people are talking in the background during an order.  Watsonx Orders uses advanced techniques to minimize these disruptions.

Understanding speech

Most voice chatbots began as text chatbots. Traditional voice agents first turn spoken words into written text, then they analyze the written sentence to figure out what the speaker wants.

This is computationally slow and wasteful. Instead of first trying to transcribe sounds into words and sentences, watsonx Orders turns speech into phonemes (the smallest units of sound in speech that convey a distinct meaning). For example, when you say “shake,” watsonx Orders parses that word into “sh,” “ay” and hard “k.” Converting speech into phonemes, instead of full English text, also increases accuracy over different accents and actively supports a real-time conversation flow by reducing intra-dialog latency.

Translating understanding into action

Next, watsonx Orders identifies intent, such as “I want” or “cancel that.” It then identifies the items that pertain to the commands like “cheeseburger” or “apple pie.”

There are several machine learning techniques for intent recognition. The latest technique uses foundation and large language models, which theoretically can understand any question and respond with an appropriate answer. This is too slow and computationally expensive for hardware-restrained use cases. While it might be impressive for a drive-thru voice agent to answer, “Why is the sky blue?”, it would slow the drive thru, frustrating the people in line and decreasing revenue.

Watsonx Orders uses a highly specific model that is optimized to understand the hundreds of millions of ways that you can order a cheeseburger, such as “No onions, light on the special sauce, or extra tomatoes.” The model also allows customers to modify the menu mid-order: “Actually, no tomatoes on that burger.”

In production, watsonx Orders can complete more than 90% of orders by itself without any human intervention. It’s worth noting that other vendors in this space use contact centers with human operators to take over when the AI agent gets stuck and they count the interaction as “automated.” By our IBM watsonx Orders standards, “automated” means handling an order end-to-end without any humans involved.

Real-world implementation drives profits

During peak times, watsonx Orders can handle more than 150 cars per hour in a dual-lane restaurant, which is better than most human order takers. More cars per hour means more revenue and profit, so our engineering and modeling approaches are constantly optimizing for this metric.

Watsonx Orders has taken 60 million real-world orders in dozens of restaurants, even with challenging noise, cross-talk and order complexity. We built the platform to easily adapt to new menus, restaurant technology stacks and centralized menu management systems in hopes that we can work with every quick-serve restaurant chain across the globe.