Day 2 of our Kiswahili ASR challenge stream focused on Conformer + CTC-based models and wow, the progress we made was huge!

✅ What we did:

• Explained CTC models and how they work in the context of speech recognition
• Shared easy-to-understand resources for learning about CTC and Conformer architectures
• Wrote custom inference code to run models fine-tuned on Swahili
• Improved our score from 0.55 to 0.44 on the leaderboard 🔥

💡 But here’s the best part one of the stream viewers, @Joseph_gitau, coded live along with the stream and even beat my score! That moment really highlighted the value of urgency and real-time collaboration.

If you're thinking of getting involved, join the streams live! The learning, engagement, and results hit differently when you're there in real time. That said, the YouTube replays are still there for anyone catching up:

📹 Replay: Zindi ML Live: Kiswahili ASR - From CTC Theory to Real Improvements (Part 2)

📺 Channel: https://www.youtube.com/@koleshjr

📆 Next Streams: https://www.twitch.tv/koleshjr/schedule

Let’s keep pushing forward, see you in the next one!