AWS for M&E Blog
Revolutionizing fan engagement: Bundesliga generative AI-powered live commentary
Many football (or soccer, in the United States) fans follow teams and leagues from around the world, regardless of their native country or continent. What truly matters is their love for the game and their devotion to specific teams or players abroad. However, fans may encounter difficulty accessing live game updates for their favorite foreign team, in a language they understand and in a persona that resonates with them. This can be due to the lack of local broadcasts airing the game or the absence of sports outlets covering live updates in their desired language. This represents a missed opportunity for leagues and teams to engage a significant portion of their international fan base and keep them connected with the excitement of a live football match, even when they’re situated on the other side of the globe.
This blog post describes a solution to address the challenge of making match updates accessible to everyone everywhere, in real time, and in a localized and personalized language. The Amazon Web Services (AWS) team has worked closely with Sportec Solutions AG, a joint venture between Deltatre and Deutsche Fußball Liga, Germany’s premier football league, to develop this solution.
The AI live commentary (or ticker) solution generates real-time, automated commentaries about match events as they happen. These commentaries are generated in different languages and writing styles simultaneously. Writing styles include “Sports Journalist”, “Casual”, or “Bro (Gen Z)”, among other possibilities.
The following video clip is an example of a goal scored by Demirović of FC Augsburg against SV Darmstadt 98. Notice the ticker entries generated a few seconds after the goal: Two entries in English, each with a different writing style, and one in German, all simultaneously generated by this solution.
The next section provides more detail about how this works under the hood.
How it works
The AI Live Commentary/Ticker solution leverages only the match event data as input to produce multi-language/styles entries.
During every Bundesliga match, approximately 1,600 event data points are generated, representing various actions that occur on the pitch, such as shots, corner kicks, passes, and more. Using this event data, the AI commentary solution leverages generative AI on AWS to simultaneously generate commentaries in multiple languages and writing styles. This process occurs in real-time, as the game unfolds, providing instantaneous and dynamic commentary tailored to the fan reading updates.
Using Amazon Bedrock within a serverless architecture
To develop this solution, we took advantage of the power of serverless on AWS, which offers several benefits. Crucially, it provides seamless scaling capabilities to automatically handle multiple concurrent Bundesliga matches, each packed with fast-paced events. In contrast, outside of match days, the serverless approach allowed us to scale down to zero infrastructure to manage, resulting in a highly cost-effective solution. By embracing this approach, we streamlined infrastructure operations, ensured seamless scalability, and optimized resource utilization, ultimately delivering a robust and efficient AI commentary system.
The following diagram illustrates a simplified, high-level architecture of the AI Commentary solution we’ve developed. Each number represents a distinct step in the process and are explained in subsequent detail.
Simplified design architecture of the solution.
- The live event data captured on the pitch is delivered to our backend solution via the Bundesliga backbone’s data delivery solution, the Datahub, which is not covered in detail here. For more information about Datahub, see this blog post from DFL.
- The data is received and processed by a container running on Amazon ECS Fargate, and important information about the event is extracted. The following example shows the event data for a shot at goal, with some selected attributes describing it. The IDs have been changed for the purpose of this example:
- The processed event data is passed as input to an AWS Lambda function, which then prepares the instructions to the large language model (LLM) and makes an API call to Amazon Bedrock. These instructions constitute the model prompt. A prompt is a specific set of inputs, that guide LLMs on Amazon Bedrock to generate an appropriate response. For more information, see our documentation on prompt engineering guidelines.
In the case of the AI live commentary/ticker solution, the prompt instructs the model to generate a captivating football ticker describing what’s happening on the pitch, based on the input event received. The prompt also specifies in which language to generate the commentary, as well as the style of writing. - Amazon Bedrock receives the API call, which specifies which model to use, the model configuration (such as LLM Temperature), and the user prompt. It then generates a match commentary accordingly.
- The response provided by Amazon Bedrock is processed by the AWS Lambda function, which then publishes it to AWS AppSync. We use AWS AppSync as the GraphQL API that connects the backend of the solution with the downstream consumers of the football ticker in a secure and serverless way. One of the consumers is a frontend web application, which we used for demonstration purposes. The frontend is subscribed to any AWS AppSync data changes, which allows it to instantly display new data entries when published to AppSync.
In addition, all commentaries pushed to AWS AppSync are automatically persisted in an Amazon DynamoDB table, to ensure data scalability and durability. AppSync uses GraphQL resolvers to manage the data flow and integration with Amazon DynamoDB, which enables real-time updates and ensures that all data changes are immediately reflected in the DynamoDB table.
English and German real time description of the match events.
The end-to-end workflow takes on average 7 to 12 seconds to finish. That is the latency from the moment the action actually happened on the pitch to when a ticker entry is visible in the UI. This latency is within the broadcasting delay, and can efficiently sync with video distribution on live streaming platforms as needed.
Next steps for AI Commentary
AWS and Sportec Solutions are actively enhancing the AI Commentary solution. Ongoing developments include integrating additional data sources, weaving compelling narratives into live tickers, and introducing new features such as pre-match previews, post-match reports, comprehensive summaries, and more.
To see the AI Commentary solution in action, watch the following video from the Sports Innovation 2024 conference.
Conclusion
This blog post described how we built a generative AI-based, real-time ticker solution using football event data, and how this makes it possible to reach a global audience in a matter of seconds, in a fully automated way.
At AWS, we’re just getting started with generative AI for sports. The AI Live Commentary/Ticker solution developed in collaboration with Sportec Solutions provides text generation. However, content creation extends beyond text.
In the realm of audio generation, generative AI will create expressive, lifelike voices to deliver play-by-play updates, capturing emotion in speech. While this won’t replace on-air talent, it will allow football reach to scale and make it relevant and compelling to a wide range of audiences in ways not previously possible.
We’re also working on visual asset creation, where generative AI can produce dynamic and engaging visuals. This blog post about super slow-motion video creation provides a detailed example.
Stay tuned for updates as we continue to push the boundaries of innovation in sports. In the meantime, check out our work with the Bundesliga.