UXPA2019: Voice User Interface Design for Customer Support: Design guidance based on user research and lessons learned

Editor's Notes

• #3: Challenges, lessons learned, and best practices in designing Voice User Interfaces (VRUs) and Interactive Voice Response Systems (IVRs) for customer support using examples from actual user studies conducted. How many people here have designed a voice interface before?? Agenda: Intro on VUIs an IVRs IVRs and UX Challenges in designing IVRs and VRUs Design exercise How to design IVRs and VRUs Best practices for IVR and VRU design supported by recent user research
• #4: VUIs allow users to interact with computers and devices through voice recognition technology. Invisible, hands-free and eyes-free interfaces. Allows for more conversational and human-like interaction. Surge in popularity in recent years.
• #5: An IVR is an automated telephone answering system that allows a caller to navigate through a series of prerecorded messages and use menu options through touch tone buttons or voice recognition. They were created to automate tasks so callers would not always need to speak to a live agent.
• #6: Bell Telephone Labs first began synthesizing speech known as “Voder”, a voice operating demonstrator only. Primitive IVR solution developed using touch tone dialing with a limited vocabulary(DSP) system became first available for consumers rather than having many agents route the calls. System could store digitized speech, play it back and understand human speech using Dual Tone Multi Frequency (DTMF) decoding. Computers and telephone data could be integrated on the IVR system. Leon Ferber’s Perception Technology first mainstream market competitor. In businesses all around the world, automatic call routing and call queuing was standard practice. Systems could use speaker-independent voice recognition of a limited vocabulary instead of requiring the person to use DTMF signaling. Voice response becomes more common and cheaper to implement through increased CPU power and migration of speech applications to the Voice XML standard.
• #7: Help with tasks such as getting stock quotes, banking, and booking flights IVRs were poorly designed however, end users tried to find direct ways of getting transferred immediately to an operator. Around the early to mid 2000s, realize importance of IVR usability https://www.personalics.com/2015/12/03/multi-channel-customer-experience-still-failing/
• #8: Designing for the ears and not for the eyes. Reliance on short term memory for audio prompts. Risk of disruption from environmental noise. Issues faced with real conversations. Issues face with real conversations such as accents, speed of and volume of speech, limitations in language.
• #12: User research Collect demographics (job role, geography) Assess experience and use of IVRs Assess reasons for calling support Assess use of other means of getting the service or support (chatbots, websites, in person support) Call or work environment (noise levels, mobility, use of hands, accessibility) Ensure fluency in English
• #13: Lucid Chart is a great tool for sketching out IVR design flows Once a basic menu structure for your IVR has been diagrammed can begin creating the audio prompts for the introduction and menu prompts Should consider response modality, typical responses, edge cases and error conditions for a few scenarios Once the audio prompts have been written and finalized on their content and style then can record them into audio files in your voice of choice using recording software. Twilio software allows IVR diagram sketching, menu routing and IVR testing by dialing into a phone number (allows both numeric or voice response)
• #14: Task completion rates and task errors (incorrect menu selections and utterances). Task completion times through the IVR and then in getting assistance. Satisfaction ratings for each scenario. Ease of completion ratings for each scenario. When user testing Important to test realistic call scenarios (user interviews, call data) Should plan to test minimum of 6 participants per round (can be in person or remote) Can use a ”Wizard of Oz” technique of playing recordings based on their selections/responses Twilio allows participants to dial a phone number into your call flow and allows voice or key press response and can be used to test the quality of voice responses for menu selections Important to create design recommendations and re-test them with users
• #16: Consider using speech: If callers will use the application in a hands-busy or eyes-busy environment such as when they are mobile.] If it saves time and money If users are visually impaired or have limited use of their hands Consider using touchtone when: Users will operate the application in an extremely noisy environment When speech does not work well because of a caller’s accent, speech disability, high noise environment, or sensitive info must be entered such as PIN numbers. Users are hearing impaired or have difficulty speaking or in an environment where speech is prohibited
• #17: Initial prompts should not attempt to mention both speech and touchtone and assume that speech will work. Touchtone can be provided as a backup if caller has problems with speech input. Press or say prompts tend to be wordier then better designed speech prompts Press and say prompts require callers to remember two things about the choices including the content of the choice and the number associated with the choice Callers prefer touchtone for entering numeric data strings For semi numeric data most callers used speech for entering data such as date of birth Menu item should come before action “For Y press X” instead of “Press X for Y”
• #18: As a general rule of thumb it is recommendation to design an IVR that minimizes demand on working memory and attention required because of their reliance on the auditory system. Keep prompts short so callers do not have to interrupt and less reliance on working memory Ensure caller responses only two or three syllables Use brief pauses liberally so callers can being speaking without interrupting Ask only necessary questions Obtain information one piece at a time in consumable pieces
• #19: Better to use flatter menu structure that has fewer levels but more options per level The number of options per menu should be limited to five or fewer Menu items should be placed into unambiguous groups of four or five (card sorting could be useful) Specific menu items should appear before more general ones and frequent menu options should appear before less frequent ones Use recognition rather then recall: IVR should understand many variations of a word or speech, think about different ways users might verbalize an action or word
• #20: 1) Here the menu options go from 1 to 3 and skip 2 designated for an agent 2) Voice response more appropriate rather then remembering numbers maybe
• #24: Implicit confirmation used for high confidence recognition and for which misrecognition does not have serious consequences to the caller but still requires some confirmation, i.e. you selected Win7. The system repeats the caller’s choice but does not require the caller to say yes or no to confirm it Should be used sparingly. Explicit confirmation -If the system is about to perform an action it cannot undo or if business rules require a confirmation before execution. i.e. was the PIN number you entered 123?
• #25: Short and simple. Clarify speaking with a machine. Include language selection. Include purpose. Include commands (small set of global navigation functions that many speech enabled IVRs support (Repeat, Help, Go Back, Start Over, Transfer to Agent, Exit). Include interrupt at any time Preference to not play “calls may be monitored or recorded” messages in the introduction unless There is a legal obligation to play the message The IVR is recording
• #26: Natural speech: Natural speech is easier for listeners to process mentally. Natural speech: Confirmed by user research in which most participants preferred the more human sounding female voice compared to the more automated and choppy sounding voices Music selection: User research found for music selection, most preferred modern music that was calm and soft to listen to.
• #27: Transfer to agent only when evidence caller having trouble. Do not announce an agent is available upfront even if one is. A call back option can also be useful when agents are busy. Respond immediately to requests for an agent indicating recognition of the caller’s request. When to transfer to an agent If callers experiencing serious problems with the system, better to get them out of the system quickly and give them another means to access the desired information such as a web link or transfer to an agent.
• #28: Replacing hold times with call back option effective at solving some IVR issues. Best to use only when a traffic problem has been identified. Offer the call back option early and often. Most participants who received the call back information said they would use the call back feature. instead of waiting for an agent on the phone. Some wanted the call back message to tell them more information about the specific time they would receive the call back Solving IVR issues At smoothing out call volume, lowering call abandon rates and improve customer experience Best practices for call-back option: Informing callers well in advance of the average abandonment time. For example, if customers abandon at 60 seconds, you may want to offer them a call-back within 30-40 seconds, to coincide with their growing frustration level. Many companies offer a call-back early in the queue (i.e., right as a caller is placed on hold), however this is often ineffective unless the offer is repeated at regular intervals. In many cases callers tend to ignore informational messages (e.g. “the wait time is longer than usual”, “your call is important to us”), so offering the call-back message when they grow impatient is the most effective way to underscore its usefulness.
• #29: AIs offer many potential advantages over human agents in IVRs: Faster and more efficient interactions with callers than human agents. No wait times as with human agents. Saves money in the long run. Works well for routine and common caller issues and requests. Highly complex issues, edge cases or call anomalies might be better handled by human agents AI should be viewed as a means of enhancing and augmenting the highly skilled role of a call centre agent
• #30: Some participants noted they wanted to begin with Watson and not to have to proceed through the IVR at all as faster.
• #31: Consider including an option right at the beginning of Watson that if they have already tried to find a solution on Help@IBM to press 1, otherwise to press 2. Phone vs. email solution Majority overall preferred the email solution over the phone based solution because they could refer to the information later from the link contained within the email and did not need to do the search themselves and remember what to search on. Some participants noted that if a phone solution was used it should be followed up with an email solution for later reference Some participants preferred having both a phone and an email solution available in case they could not access the network. Some noted they wanted the email to have the steps right in the email without using a link. A few noted including a phone number to contact in the email solution would be valuable
• #32: AI did not always pick up the utterance if the exact words were not used i.e “Reset network password” or “AT&T” Consider making the AI NLP a bit more flexible to word detection i.e. ”password reset” “AT&T”, “network password” Consider shortening the AI prompts and keeping them concise to only include essential information to the user for the specific issue. URLs Understanding urls was challenging for four participants. For longer urls, consider spelling out the url and saying it slowly or sending the url in an email message. Many participants preferred the voice solutions over an email solution for simple problems as easier to listen then to toggle screens. More more complex solutions with a larger number of steps i.e. >6 then email might be better as mentioned by a few participants URLs also mentioned to be better offered through email, or to spell out the url in the voice solution and speak it more slowly. Many participants preferred AI over an advisor for a first point of contact unless a very complex solution.
• #33: When designing for voice, it’s important to test the voice-controlled interactions as two-way conversations first, temporarily putting the technology and function of the device on the back burner. Consider what a positive and successful interaction would be between a person and a service, device or assistant. You may find that a verbal response is not always necessary or preferable. Ensure the VRU takes into account and respects the caller’s expectations.  The design also needs to take into account a variety of usage environments (e.g... mobile, inside/outside, or accessibility issues) so that the VUI will work for a wide range of callers.  Do your user research early and often Ensure best practices in design are followed