The advantages of deploying an NLU bot to Verizon 5G Edge with SignalWire and AWS Wavelength
By Evan McGee, CTO @ SignalWire
With their enhanced computing resources, the 5G networks being deployed across the world enable an entirely new class of application — one that can make the user experience more natural, more comfortable and dramatically better.
Even on fully deployed 5G networks, an application running on a mobile device can still incur latency of 100 ms or more due to the many hops data has to traverse through the public internet.
Here at SignalWire, we’re helping companies address the last frontier of latency by taking advantage of edge computing for video, audio and messaging applications. By locating computation for these use cases on the outermost edge, we’re able to remove myriad public internet hops and significantly improve performance.
Studies have shown that quick activation and response times are primary factors in providing the kinds of high-quality, natural communication experiences that directly translate to customer satisfaction and retention. On top of that, our infrastructure and edge-compute performance allows real-time communication apps to incorporate advanced features such as artificial intelligence-enabled natural language understanding (NLU) and other machine learning (ML) techniques at dramatically reduced latency compared to public internet implementations.
Features like ML and computer vision that are backed by server-side services — or any other use case where latency and quality of service are important — benefit from the new joint architectures being offered by players like SignalWire, AWS and Verizon. Innovative use cases such as augmented and virtual reality (AR/VR) gaming experiences, micro sportsbook rooms with live video feeds, live drone piloting and a truly pervasive Internet of Things (IoT) all require real-time feedback and gain significant improvement from the acceleration that edge computing affords.
In this post, we will explore an app built by SignalWire and deployed on Verizon 5G Edge with AWS Wavelength. Our goal was to build and deploy a voice chatbot that uses NLU sentiment analysis to listen to callers and provide reassuring responses.
5G networks and the “new cellular revolution”
Compared to previous technologies, 5G provides significantly higher bandwidth to make applications such as virtual reality or 4K streaming possible. This is coupled with incredibly low latency and a massive network capacity that can support up to 100 times the devices per unit area as 4G.
This leap in performance is achieved through two components. The first is the service-based architecture standard that allows companies to bring cloud services and open-source applications into the “classic” wireless networks. The second is the capillary distribution of base stations, antennas and edge computing resources to allow deployment of applications as close to the user as possible.
What is edge computing?
Edge computing is a distributed system paradigm that involves deploying applications closer to their final users to improve latency, increase quality and save bandwidth. The edge may refer to IoT devices inside homes, the base of a 5G tower or even data centers that are simply situated closer to end users.
While improved performance is one benefit, edge computing also has important privacy implications. As data travels shorter distances, even to the point of not leaving the 5G tower to which it was directed, the chances of interception by third parties drops.
In a 5G context, this means that cellular network users receive increased speed, performance and security. Instead of having to wait for applications to connect to faraway systems and services, they can receive immediate feedback.
What is AWS Wavelength?
Amazon Web Services (AWS) Wavelength is an AWS Zone optimized for mobile edge computing on the Verizon 5G network. Wavelength is unique in that it operates much like a standard EC2 zone, allowing customers to move services to Verizon’s edge without rearchitecting entire applications.
This makes it possible to run EC2 workloads closer to Verizon customers almost without modification. And because every Wavelength instance is connected to its parent AWS via a secure, high-bandwidth connection, developers are able to connect to the full range of AWS application programming interfaces (APIs) and tools in a seamless way.
Wavelength can even leverage Amazon virtual private clouds so that edge-deployed services on Wavelength can share resources with AWS internal servers and resources. Wavelength abstracts away all the complexity of an edge deployment and makes it a relatively simple exercise to set one up.
As there are multiple Wavelength zones available around the United States, you can choose zones that bring the edge closest to your clients and deliver the latency-sensitive applications that benefit most from that proximity advantage.
What is SignalWire?
SignalWire is a leader in server-side, real-time communications APIs, founded by leading open-source contributors including the FreeSWITCH project creators.
We provide innovative APIs for voice, video, audio, chat and messaging, plus a rich set of reporting endpoints and software development kits (SDKs) in multiple languages. Our services include audio/video conferencing, public switched telephone network and Voice over IP services, speech recognition, two-factor authentication, text-to-speech, messaging, and our audio/video conferencing SDK.
Our goal as a company is to bring advanced, programmable telecommunications into the world of low-latency and pervasive computing. Since the company was founded, our mission statement has included the goal of achieving global latency below 50 ms, and we continue to meet that challenge.
For this demonstration, we built an application that utilizes the low-latency speech recognition capabilities of the SignalWire Cloud API to provide data to a local sentiment analysis service.
Leveraging our unique, elastic, cloud-based deployment allowed us to put application instances directly on a 5G Edge node, reducing latency and streamlining resources for efficiency — an approach we take in the development and testing of edge-deployed telephony services.
While our session border controllers and call control — including the speech recognition and application logic components — have been deployed to AWS Wavelength, the reporting, logging and billing stays within the main cloud infrastructure, helping to minimize changes and change management.
The demo application
From the start, we built our application with the goal of measuring the latency and round-trip time of a WebRTC service. To demonstrate the effect of edge deployment, we opted for a simple NLU-powered chatbot that’s written in Node.JS and consists of two main parts — a front end and a back end.
The front end is the actual web page you can see on the demo page, which utilizes SignalWire’s WebRTC client in its v2 version. When there is interaction with the bot, events are reported to the browser via a real-time WebSocket connection. Meanwhile, the network statistics used to measure the round trip, latency, jitter and packet loss are extracted from the browser API.
The back end is a SignalWire Relay consumer. Relay is a SignalWire innovation that controls a call via a WebSocket connection and that provides extremely fast interaction with next to no overhead. This is crucial to ensuring that measurements taken during this test are as accurate as possible since quick application response time does not affect network interactions.
Sentiment analysis is performed directly on the edge node using an open-source Node.JS library, node-nlp, developed for the insurance and security markets. The algorithm used is called Senticon and it combines frequency analysis with a list of common words to improve accuracy.
In our test, we took three main data points into consideration: connection time, call setup time and round-trip time.
Connection time — defined as how long it takes the client to log in before it can place a call — is important because the shorter the connection time, the more likely the user’s experience with the application will be quick, responsive and natural. Call setup time is the time it takes for a logged-in client to hear the first audio in the call. The shorter this time is, the more likely the application will be perceived as readily available by users, which has been shown to increase usage or reduce bounces.(1) Finally, round-trip time measures the conversation latency and is an important part of the perceived quality of service that’s typically measured in a mean opinion score.
Testing was performed using an iPhone® 12 device on the Verizon 5G network that connected to two instances of the same application running on similar servers. One instance was on the Wavelength network while the other was in the same geographical zone, Oregon, on standard AWS machines.
Results on all three data points were impressive. On average, connection time was three times faster when using the Wavelength instance, and subsecond times were achieved. Call setup times were consistently 1.4 times faster, exhibiting less of a difference because negotiating Interactive Connectivity Establishment candidates necessarily involved connections to IPs outside the 5G edge zone. The round-trip time, similarly subject to the need for connections outside the zone, registered an improvement that was 1.6 times faster.
5G Edge compute zones allow app providers to move workloads closer than ever to the user, and the combined impact of 5G speed, edge compute performance and SignalWire’s real-time APIs deliver significant and impressive benefits to responsiveness and quality of experience.
The ability to deploy our cloud-native infrastructure in Verizon 5G Edge with AWS Wavelength edge zones presents an incredible and unique opportunity to incorporate next-generation capabilities like ML, natural language understanding and metadata analysis into applications that would otherwise be rendered unusable by processing and transmission delays. Because it’s implemented as an AWS zone, deployment to Wavelength is a smooth and familiar process, making it an even easier choice for teams looking to grow adoption of their advanced applications.
When latency decreases across remote experiences, the result is typically a consistent increase in satisfaction, adoption and usage. The significant performance gains we’ve been able to demonstrate, even when external connections were involved, opens the door to an impressive range of uses, including automated interaction with NLU, real-time moderation and analysis, naturally flowing conversations, real-time sports micro-booking, live interactive AR, VR, gaming and more.
At SignalWire, we are excited to be a part of this evolution in cloud computing that will make possible a host of powerful integrations to real-time experiences.
The author is solely responsible for the content. Its inclusion does not imply endorsement by Verizon of the content, the third party or its products or services.
— — — — — — — — — — — — — — — — — — — — — — — — — — — — — —