AI-Podcast Recommendations

Amazon Music

Design for AI to personalize podcast discovery. Podcast Picks won Best Application of AI for Amazon Music's 2024 Design Challenge.

DURATION

4 Weeks

TEAM

Meng Shi
Sara Ma

My CONTRIBUTIONS

User Research
UX Design
Usability Testing
Animation

AWARDS

Design Challenge Finalists
Best Application of AI
PROBLEM

Assessing Podcasts Take Too Much Effort

My user interviews revealed listeners spend a lot of effort to assess new podcasts but are unsuccessful in finding a new show to follow. The podcast discovery experience suffers from two key issues that make it hard to sift through new podcasts.

ISSUE #1

Assessing Tone is Hard

Recommendation algorithms don't match tone preferences well. Listeners were disappointed when recommendations were too casual or too formal and are frustrated that they only realized after they started listening to the podcast.

In the same genre, the content varies a lot... I don't want to hear two people shoot the shit. LEON, 38
LANDSCAPE ANALYSIS

Only Snippets and Transcripts

Tone is conveyed on streaming platforms through listening to a snippet or reading a transcript, both of which are time-consuming.

Amazon Music, Spotify, and Apple Podcast's discovery screens showing the ability to listen to snippets of audio and read descriptions
SOLUTION

Express Tone through Keywords and Visuals

the design includes a tone visualization animation in the background, tone summarized as keywords and a sound bite

1. Visually Set the Vibe

Background visualizations abstractly represents the podcast tone.

2. Highlight Tone with Key Words

Straightforward tone indicators at the bottom of the text highlight the podcast style.

3. Keep the Sound Bite

The sound bite remains one of the best ways a listener can sample the podcast.

ISSUE #2

Key Info is Missing

Tone, content, and reviews are 3 key factors listeners use to assess new podcasts that needs to be improved. I learned users also want to know why the podcast is being recommended during card sorting to priortize factors.

Tone Summaries Reduce User Burden

Understanding tone is difficult with only using snippets and transcripts.

AI Personalizes Podcast Summaries

Creator written episode summaries lack quality control, impacting length and relevance to new listeners.

Summarize Reviews from Outside Sources with AI

Only Apple has reviews on their podcast platform despite users wanting to read them.

Show Why a Pocast is Recommended

Listeners want to understand why specific podcasts are being recommended to them, which isn't being provided now.

SOLUTION

AI Extracts Pertinent Information

The current flow requires users to click into podcast shows to access show information. The new design brings relevant and personalized AI-generated podcast information out from the podcast show page into the explore page.

the current podcast page only shows images and titles whereas the new podcast discovery will include key information a user normally would need to go to the podcast show page to discover
DESIGN

A Card Carousel Accommodates Relevant Information

A carousel allows users to access more information quickly if they're interested and skip past podcasts if they're not. I used paper prototype testing to arrive at the final card layout.

four cards that can be swipped through in a carousel to show more podcast info
TESTING

Ramp Up the Personalization

Users wanted clear indicators AI is personalizing content to their preferences.

Previous titles for cards were generic
New content for cards is conversational and personalized
FINAL PROTOTYPE

Presenting Podcast Picks

The final UI and clickable prototype is made by my partner, Meng.

qr-code to access the clickable prototype created in Figma. Also, accessible through url: https://tinyurl.com/Amazonpodcastpicks.

Scan or click for the prototype!

FUTURE CONSIDERATIONS

Creators Deserve a Say in Podcast Picks

In the challenge, our scope was limited to the listener's experience. Given more time, Meng and I would have explored how creators would respond to AI representing them. We would explore creator tools to review and refine AI-generated descriptions.

REFLECTIONS

Lessons in Moving Foward

With only 4 weeks to complete the design challenge from prompt to presentation, we outlined each phase of the project. Inevitably, our design process wasn't linear but we were always able to move forward with each conversation.

Getting Unstuck

Our mentor helped us reframe when we got stuck by reminding us of the bigger picture. I learned to refer back to a Northstar to guide our decisions when stuck.

Prepardness Pays

Preparing objectives, questions, and a weekly progress summary for meetings allowed us to make the most of our mentor's time.