Building an AI-powered Football Commentator
<p>If you’ve ever played a game like FIFA or PES, you’ll know that sometimes the pre-recorded commentary leaves you less immersed than you’d like.</p>
<p>A commentator could be reading the team sheet, and suddenly they interrupt themselves by screaming a player’s name.</p>
<p>Or maybe you’ve just heard the same match summary one too many times over the course of a few seasons.</p>
<p>These tiny annoyances made me wonder if there could be a way to improve that experience.</p>
<p>So I decided to try my hand at building a proof of concept which solves these problems.</p>
<p>The flow I wanted was:</p>
<ol>
<li>Input some largely unstructured football data (e.g. match stats)</li>
<li>Have a commentary script generated using those stats, in the style of an experienced commentator</li>
<li>Then have a realistic-sounding voice read out the generated script</li>
</ol>
<p>I selected a few different technologies for this project:</p>
<ul>
<li><a href="https://nextjs.org/" rel="noopener ugc nofollow" target="_blank">React/Next.js</a> for the frontend and some APIs (I had never built with React before, I am firmly an Angular dev so I wanted to see what all the fuss was about)</li>
<li><a href="https://openai.com/product" rel="noopener ugc nofollow" target="_blank">GPT 3.5 Turbo</a> for the commentary generation, as it is both impressive in its generations and cheap per token generated</li>
<li><a href="https://beta.elevenlabs.io/" rel="noopener ugc nofollow" target="_blank">ElevenLabs</a> text-to-voice service, to generate a realistic-sounding voice rather than a synthetic robotic voice</li>
</ul>
<p><a href="http://If you’ve ever played a game like FIFA or PES, you’ll know that sometimes the pre-recorded commentary leaves you less immersed than you’d like. A commentator could be reading the team sheet, and suddenly they interrupt themselves by screaming a player’s name. Or maybe you’ve just heard the same match summary one too many times over the course of a few seasons. These tiny annoyances made me wonder if there could be a way to improve that experience. So I decided to try my hand at building a proof of concept which solves these problems. The flow I wanted was: Input some largely unstructured football data (e.g. match stats) Have a commentary script generated using those stats, in the style of an experienced commentator Then have a realistic-sounding voice read out the generated script I selected a few different technologies for this project: React/Next.js for the frontend and some APIs (I had never built with React before, I am firmly an Angular dev so I wanted to see what all the fuss was about) GPT 3.5 Turbo for the commentary generation, as it is both impressive in its generations and cheap per token generated ElevenLabs text-to-voice service, to generate a realistic-sounding voice rather than a synthetic robotic voice">Website</a> </p>