In this stream, we kicked off a brand new challenge: building a Kiswahili speech-to-text (ASR) system for real-world, low-resource environments.

🧠 Here's what we did:

• Broke down what ASR is for beginners
• Explored the Zindi competition and the real-world impact of offline voice tech
• Researched past ASR competitions and stalked a winning solution (spoiler: it was mine 😅)
• Created a training-free baseline by hunting for pre-trained Swahili models on HuggingFace
• Let AI write the code for us to simulate a beginner workflow
• Faced a hilarious number of bugs (oops) but got it working in the end
• Landed a Top 5 leaderboard position with zero training 🎉

FYI: The 3rd Placed sub is not shown on the stream but it is a hunted model as well(So keep hunting haha)

Tomorrow we explore: conformer-ctc-asr baselines

📺 Watch the full replay here: https://youtu.be/rZuEj6JBZOY?si=QDg2S_5qzqD714cX

📌 Subscribe to catch future ML streams: https://www.youtube.com/@koleshjr

📆 Join live sessions (Tues/Wed/Thurs): https://www.twitch.tv/koleshjr/schedule