AI Soccer Highlights

Inspiration

We were inspired by the idea of using AI to sift through full-match transcripts and highlight the most thrilling sections. We wanted to make it easier for fans, analysts, and media creators to quickly access the best parts of a match without missing any crucial details.

What it does

This program automatically analyzes an entire match's commentary and generates an "excitement heatmap," pinpointing the most important or thrilling moments. Using those timestamps, it can then produce concise video highlights or short clips of the action, letting users jump straight to the best parts of the game.

How we built it

We took a transcript (in JSON format) with detailed timestamps and commentary. The transcript was split into manageable sections, each covering a small window of match time. GPT-4 was used to read each text snippet and assign an "excitement score," helping us rank how highlight-worthy the segment is. We plotted these scores on a timeline to visualize excitement across the full match duration. Based on the highest scores, it identifies key windows that can be turned into short, engaging highlight reels.

Challenges we ran into

• Transcript inaccuracy: Real-world transcripts often contains errors or missing words.
• Token limits: Contraints on maximum number of tokens.
• Time alignment: Matching text segments with exact video timestamps, especially when the speech-to-text data only provided end times.

Accomplishments that we're proud of

• Turning raw transcript data into an excitement heatmap and highlight reel with minimal manual intervention.
• Successfully leveraging GPT-4 to rate excitement without requiring a custom-trained model from scratch.

What we learned

• Prompt Engineering: Small changes to how we prompt GPT-4 can drastically improve the relevance of the excitement scores and event detection.
• Handling large transcripts efficiently by using strategies to avoid token and performance bottlenecks.

What's next for AI Soccer Highlights

• Video integration: Directly clipping the original match video at the high-scoring timestamps.
• Multi-Lingual Support: Expanding pipeline to handle transcripts in multiple languages.

Built With

• google-cloud
• gpt-4
• python