News Quiz from TRT World has recently won awards in two separate categories in Google Actions Challenge. Our quiz tests user knowledge about the current events with new questions every day and exist both on Google Home and Amazon Alexa.
It took a lot of iterations to get there. We were faced with unexpected challenges as developing an app for a voice interface requires a different mindset. This article will detail our process of getting to the end product and the lessons we learned. As one of the first comers to the voice assistant platform in the media industry, we’d like to share some of the practices we developed along the way.
Conception of the idea
After receiving our first Google Home around the end of 2016, we started thinking of ideas unique to this new platform. Our first step was to investigate what our competitors were doing. Interestingly, we only saw 30 apps on the platforms excluding the news briefs. One category that caught our interest was the quizzes. As an age-old concept, quizzes managed to stay relevant and it seemed one of the simplest and most engaging experiences we can provide on the voice-activated devices.
Another advantage of the quiz was the ability it would give us to repackage and distribute the same content in other domains like smart TVs, website, mobile apps. The quiz could also be used to redirect the users to the relevant content on our site. This thinking had a great influence in the way we planned for our news quiz as we had to make the interactions smooth enough to keep the user retention high.
Throughout this planning process, we tested all of the apps made by news organizations for Google Home and Amazon Alexa. Our takeaway was that at the start, almost none of the apps made my media companies seemed to grasp the new type of interactions required for the voice-assisted platforms. We will go into the detail of our analysis in another article.
Designing the MVP
For the MVP, we decided to go with a fixed number of questions. The number of questions in the quizzes made for voice-assisted platforms ranged from 4 to 10. However, voice is unique in terms of the user attention and we decided to keep it short and decided to keep the question number at 3.
Since our focus was to make this app a weekly news quiz, we had to make sure that the architecture would support the weekly updates to the product. The other design consideration we had was the support for linking each question to a news story.
In order to quickly get a prototype out, we used API.ai or also known as Dialogflow. The system allowed us to map certain words to certain actions, but there was a problem. We wanted the quiz to progress as the user answers. So we had to call a webhook each time the user started the quiz and responded as to what the user sent to the server. The interface API.AI provided helped us to match what user said to the corresponding action in the webhook.
2. 4 DO’s and 3 DON’Ts for Training a Chatbot NLP Model
3. Concierge Bot: Handle Multiple Chatbots from One Chat Screen
Technical and UX Challenges
The initial challenge
While building the prototype, we encountered a problem regarding the answers. As the questions were to change every week and the answers could be in any form, we quickly realized having the user utter every answer would be very difficult. There were cases where we couldn’t pronounce our own answers correctly or where we did pronounce correctly, Google Assistant wouldn’t understand and throw an error.
The solution we came up for this problem was having the user say the number of the answer instead of the text. However, we made sure that if the user says both the number and the answer itself, we’d still pick it up.
Before launching the product, we compiled a list of speakers with various accents to understand whether our number approach is ideal. One immediate problem we noticed was the user onboarding. Without a proper onboarding, most of the users did not speak the number of the answer and got confused. Also, some people had a problem saying two and three. So we mapped “do” to two and “tree” to three. Thankfully, Google Assistant allowed us to see every word user said in order to invoke an action. This showed us where the users had the most problem. For instance, we realized that some users were asking the question to be repeated or just generally asking for help. Upon seeing this, we quickly implemented these functions.
Although we were one of the first-comers to the voice assistant space in our industry, our app lacked a special ingredient that would make the users talk to their friends about it. One thing that most apps failed to do was change the default voice of the device and make it feel more human. Some even argued that the TTS is better as the users’ expectations are lowered due to the robotic voice. We decided to change that entirely and make the app completely human narrated.
Another critical thing we implemented to help the game feel more playful was custom responses to right or wrong answers to a question. We wrote and did the voice-over for over 40 responses. The response would come randomly from the pool. These responses took almost all the users who tested it by surprise. If you got the question wrong, our app would say congratulations in a jovial tone only to be followed by that’s wrong. In fact, one of the user testing participants wanted to take the quiz again so that she’d hear the other responses we had. This style later on was awarded by Google as the best persona in apps on Google Assistant among thousands of applicants.
Despite our foray into the weekly human narrated quiz, we believed we could do better and increased the frequency from weekly to daily. This forced us to rethink our ad-hoc approach to publishing new questions every week to Google Home and Alexa. We sat down with the product manager of our CMS in order to automate the process as the actions require a specific file format for the audio files and the text should match the audio exactly for it to be on Google Assistant. We managed to build a quiz publishing platform inside our CMS. This platform allowed us to publish the quiz with full editorial workflow integrated and the audio and image option presented. The quiz can now be sent to the website, smart TVs, Google Home, Amazon Alexa and the other platforms our quiz is present through our CMS.
On the content side, we faced a backlash from the content team due the lack of resources on their side. Producing questions and voicing them over on a daily basis caused heated discussions. Firstly, they proposed preparing a batch of questions beforehand and do the voiceover daily. This approach worked for a while, but then we noticed the change in voice-over person every day didn’t sound right as the intro, outro and responses had our original entertaining voice while the questions were voiced by someone else. We changed tactics and got all the questions rewritten to fit our persona and goal of having a proper news quiz. All the questions had been voiced-over at once to be uploaded on a daily basis.
Coinsmart. Beste Bitcoin-Börse in Europa
Source: https://chatbotslife.com/how-we-made-an-award-winning-google-assistant-app-f33a0f795d09?source=rss—-a49517e4c30b—4
