Humans are social animals, we love talking with one other. Typically, human conversational styles are informed by our personalities, expressed in tone of voice, level of knowledge, sense of humor, and so on. These qualities are now being applied to virtual assistants, the AI-powered human-machine conversational interfaces in our smartphones, homes, and cars.
The personality of a virtual assistant is defined by the words, tone, style, dialect, and behavior it uses. The goal of any virtual assistant is to establish user trust, engagement, and satisfaction.
Speaking at the recent RE-WORK AI Assistant summit, Dr. Ann Thymé-Gobbel, Voice UI/UX Design Leader at Sound United, said that content accuracy is most important virtual assistant trait.
Amazon Alexa’s recent attention-grabbing Super Bowl commercial features a host of Hollywood stars filling in for Alexa, who has “lost her voice.” The substitute virtual assistants mess around with users, for example a man who requests a grilled cheese sandwich recipe is mocked by celebrity chef Gordon Ramsay: “It’s name is the recipe, you #&*$%!”
Content accuracy metrics include accurate responses, correct info, dependability, reliability, low error rate, high comprehension, correctness, and minimal repeating. To meet these demands researchers require robust and reliable technologies in knowledge base, speech synthesis, speech recognition, and natural language understanding.
Dr. Thymé-Gobbel says voice quality ranks second, referencing a study comparing text-to-speech (TTS) synthesis with actual human voices. Although both deliver a similar level of trust, real human voices scored much higher in user enjoyment.
Given the same level of content accuracy and voice quality, users are more likely to engage with a virtual assistant that is not too chatty, admits when it is uncertain about answers, and is friendly yet professional.
Dr. Thymé-Gobbel says an overly crafted virtual assistant does not necessarily deliver a positive user experience, which is a surprising discovery given that major virtual assistant makers employ storytelling advisors to help compose answers. Google has a team tasked with imagining likely questions and creating responses for developers to code. Microsoft’s Cortana team meanwhile has eight full-time writers brainstorming what Cortana will say and how it will say it.
The recipe is simple. “Light humor or a quirky sense of humor is great. No jokes unless I ask. Don’t be too eager to please, and avoid being fake friendly,” says Dr. Thymé-Gobbel.
However, even a sense or humor or a strong personality might be inappropriate for virtual assistants in domains such as business, healthcare, education, etc. Virtual assistant makers must understand and stay true to different underlying interaction environments. For example, a healthcare virtual assistant should speak and behave as a trusted doctor or nurse would. Its user expectations and goals are very different from a home virtual assistant whose role is to be an advisor or entertainment content provider grounded in interaction and enjoyment.
Dr. Thymé-Gobbel suggests virtual assistant makers should enable further personalization. She believes users will respond favourably for example if a virtual assistant remembers what they previously said and delivers tailored services using personalized voice characteristics such as tone, accent, rate of speech, access preferences, etc.
While personalization can boost user engagement, Dr. Thymé-Gobbel also warns that “users don’t want machines to automatically personalize their behaviors.” An inappropriate personalization choice might lower user trust and satisfaction, and negatively affect performance. Also, a user communicating for example with an AT&T customer service bot may not require or desire personalized interaction to get a satisfactory user experience.
Dr. Thymé-Gobbel told Synced that studying the personality and personalization of virtual assistants has helped her dig into the technology behind them. “As a voice designer person, I have to decide on flows and systems, what to include and what not to include. People care most about having good content and being recognized. You know who I am, so you can be more helpful, like a real person.”
Historically, most innovative products have evolved from uniform to personalized. Ten years after the Model T Ford was introduced in 1908, the first world’s first mass produced car was available in one colour: black. The automotive industry subsequently spawned a multitude of models with different colours, designs and options to target specific market sectors. It’s a safe bet that a similar broadening of options is about to occur with virtual assistants.
Journalist: Tony Peng | Editor: Michael Sarazen