Salesforce AI

Voice cloning App

Voice cloning AI

AI concept

Development

Swift UI

Development

Swift UI

How can we break the black box model of voice cloning and make the process more intuitive and accessible?

MY ROLE

Alex

Designer

Soyun

Designer

Gahui

Researcher

Mingjin

Researcher

Team

With amazing teammates, I led the end-to-end research-through-design process, including prototyping voice clonig AI and UI/UX design for the product.

The problem : Black box

text

Voice cloning systems are powerful but opaque. Even the most popular models from Hume AI and Elevenlab are completely unexplined. You record your voice, submit it to a model, and receive a clone, but what happens in between is a black box.

Background

In a world where social media and virtual content often overshadow live events, we aimed to rekindle the appeal of live concerts in Pittsburgh with a new service design.

Why salesforce?

text

Salesforce AI Research was interested in exploring a more interpretable, human-centered voice cloning system — one that gives users greater control and transparency throughout the process. This research also had potential internal applications, such as creating more brand-aligned voices for Sales and Customer Support functions.

Background

What we did

text

We started with an intentionally blank canvas to co-design with users. We recruited people with experience in voice cloning and TTS, and asked about their frustrations and pain points with existing tools. We then ran functional prototype walkthroughs to uncover user motivations, mental models, and where trust broke down.

Background

What we did

text

Finding 1:   Users believe that output quality depends on providing representative input data to capture the characteristics of their voice.

Background

Mid-fi Prototype

text

After our first round of testing, we received clear feedback: the input flow was too detailed, causing information overload, and unclear copywriting was adding to the confusion.
We revisited the input stage with three changes.

Background

Mid-fi Prototype

text

Background

03/ Ideation

Initial brainstorming

What are design opportunities?

Through competitive analysis, I discovered design opportunities for this AI concept.

AI for a group, not just individuals

I envisioned a recommender system that can interpret and reconcile inputs from multiple users, generating suggestions that reflect the "group" perspective.

Enable dynamic preference input

The new design should allow users to set filters collaboratively, and adjust individual priorities and preferences.

Lower the inconvenience, raise the fun

By leveraging criteria-based recommender, we can ease the burden of manual compromise and negotiation, unlike the current reliance on text-based comments.

Concept

Bubble-inspired, playful trip planning!

Inspired by Apple’s Image Playground, I created a moodboard for an interaction that combines a glass-like aesthetic with flexible, dynamic behavior.

Wireframes

Quickly sketching out the main flows.

Considering components, and the overall design system, I quickly sketched the main flows of (1) editing preferences, (2) group-AI curation process, and (3) voice/image-based AI interactions.

Flow 1

1. Edit preference for my proxy 2. Adjust weight of each preferences 3. Share my proxy with friends

Flow 2

1. Add trip details 2. AI analyzes each proxy's preferences. 3. AI suggests curated destinations, by prioritizing diffent preferences.

Flow 3

1. Chat with AI to adjust preference weight or receive recommended tags. 2.Upload images to get more personalized or related keywords (bubbles).

4/ Design Iterations

Detail 1

Aesthetic UI touch to increase the "AI" feel

Gradient UI

I tried to add gradient "just right amount" to increase the feel of interactiveness.

Liquid glass aesthetic

Although we were building our product earlier, one of our golas was to explore the glass UI. With the ios26 release, we could integrate new UI into ours super neatly.

Detail 2

Small tweaks to increase the fun!

Card- placing UI

As each member’s preference is highlighted, the results are displayed in a card deck format. This turns the stress of group decision-making into a moment of fun, like card games!

Detail 3

Novel interactions for profile editing

Engaging bubble interactions

Setting up different proxies for each trip can feel tedious. With simple touch-based weight adjustments and bubble-like physical traits, users can seamlessly engage with the bubble ecosystem.

05/ Takeaways

Dev hand-off

This was my first time working on a UI-heavy project directly with a front-end engineer from scratch. I learned how to clearly translate UI and interaction needs into developer-friendly terms while iterating through constant collaboration.

Dev hand-off

Explore, Learn, Enjoy

At first, I felt pressured to build an AI concept by strictly following UI “must-dos.” But I soon realized it was restricting my creativity and approached it as a creative experiment. This mindset gave me the freedom to explore more.

Explore, Learn, Enjoy