.png)
.png){kind=small}
SUMMER 2021
The first portion of my research focused on speech study itself. Everything from lexicons, to the differences between tone and intonation, to the influence of vocal tract and vowel lengths. All of this information is crucial in order to understand how to modify virtual voice systems to produce the exact quality of voice to match a voice assistant persona.The next step was to audit the different voice technologies available on the market. Voice technologies allow written scripts to be read out loud by the system in a way that mimics human speech. This is known as Text-to-Speech, or TTS, technology. Most TTS systems allow you to mix a combination of voice traits, such as vocal tracts, prosody, and volume, to tweak the robotic voice as desired. RedRoute had built their initial system off of Amazon’s TTS system, Polly. After auditing the different technologies for the sophistication of customizability, we opted instead to use Microsoft’s vocal agent. Now that we had the tools to create the voice, I needed to create a persona off of which to model that voice.
Creating a voice persona is not unlike creating a user persona, which I was familiar with through my experience in product design. To kick off the process, I sent out a survey to all of the employees at RedRoute. In the survey, I asked participants questions such as:
• What is an example of a voice system you really disliked? Why?
• What is an example of a voice system that you really enjoyed? Why?
• If you felt like you could trust this system, what was it about the voice that made you feel like that was the case?
• Think of a positive customer service experience you had in real life. What was that interaction like?
• When you envision the RedRoute voice assistant as a real person, what are they like? What kind of job title might they have?
Some of the most important characteristics respondents valued were expertise, kindness, and lenience. Some of the professions they thought exhibited these traits most were teachers and airline attendants. I used this information to come up with a series of personas. After meeting with product, we landed on a persona we nicknamed “Ronnie” - a high school teacher with a stern but fair temperament. I used the Ronnie persona to tweak the Microsoft TTS voice system and create a set of sample voices that could potentially be used as the new voice of RedRoute.
In order to start the next portion of the project, I needed help from conversation design experts in the field. I conducted a series of informal interviews over several weeks with a wide range of conversation designers, from podcasters to consultants. During this time I met the expert who would become my mentor, Becca Evanhoe. Becca is a conversation design expert, course leader at Pratt University, and was in the process of literally writing the book on conversation design (now since released, called “Conversation with Things”). Becca taught me how to quantitatively evaluate scripts, run conversation workshops, and methodically design conversational flows. With this guidance, I was able to audit the entirety of RedRoute’s existing script system, categorizing the scripts into Intents, Utterances, and Slots. With all of these in mind, I was able to create a Customer Journey Map for each of RedRoute's core verticals to get a sense for how callers might branch in their paths through the system.
High-level requests users have, or tasks they want to accomplish through the system.
Example: “Book a new ride” or “Exchange an item”
What a user says to make that request - either a specific word, phrase, or sentence.
Example: “I need a ride to the airport”
A blank portion of an utterance where any value can be inserted.
Example: “I would like the [DRESS] in [COLOR]”
This was a lengthy process - even a simple conversation can have many branching paths, and in call center automation, the number of branches is quite extensive. I primarily used Excel to transcribe and annotate sample scripts from RedRoute calls, and LucidChart to diagram conversational paths.
The customer journey for a taxi client with RedRoute:
Having audited the current state of RedRoute scripts for strengths and opportunities, it was time to start writing new scripts. I primarily used a tool called VoiceFlow to chart out scripts along each of the defined Intents and their branches. I was able to use the Ronnie voice I had designed earlier to test how scripts sounded through the persona. The culmination of this process was creating a set of 10 new intent scripts to use for RedRoute’s e-commerce clients.
This was a lengthy process - even a simple conversation can have many branching paths, and in call center automation, the number of branches is quite extensive. I primarily used Excel to transcribe and annotate sample scripts from RedRoute calls, and LucidChart to diagram conversational paths.
A demonstration of the process of designing scripts in VoiceFlow:
In addition to the e-commerce scripts, I package up all of my learnings and research into the first ever RedRoute Conversation Design Guidelines. In the guidelines I included tools, learning resources, and a glossary of terms. The Guidelines were created to allow future script writers at RedRoute to create scripts with consistent themes and tones. I also included my work with voice design in the even the Voice Assistant persona needed to be tweaked in the future. Since the conclusion of this project, the new scripts have been implemented into the RedRoute system, along with the voice characteristics of the Ronnie persona.
This project was unlike any other form of design I had attempted before, and yet I was surprised by the overlap with UX and product design. Many of the fundamental principles are the same: base your practice in research, create a methodology and be consistent with it, and create an experience that is simple yet effective. While it was certainly challenging having to learn an entirely new medium, there is a rich community of voice and conversation designers that made the process of doing so both educational and enjoyable.