Primary competition visual

Your Voice, Your Device, Your Language Challenge

Helping Africa
1 000 CHF
Challenge completed ~1 month ago
Automatic Speech Recognition
Natural Language Processing
278 joined
73 active
Starti
Jul 22, 25
Closei
Sep 22, 25
Reveali
Sep 22, 25
User avatar
Koleshjr
Multimedia university of kenya
Zindi ML Live – Kiswahili ASR (Part 5): Finetuned Models… But Did They Work?
Platform · 7 Aug 2025, 19:06 · 0

Today’s stream was a reality check. We focused on evaluating our finetuned Kiswahili ASR models to see if they actually improved performance — and unfortunately, they didn’t.

But that’s exactly what makes competitive ML exciting and tough. It’s an iterative grind, and progress doesn’t always come in a straight line.

💡 Key takeaways:

  • Finetuning doesn’t guarantee better results
  • Sometimes your base model choice limits your performance
  • Hyperparameter tuning and architectural changes are next

We discussed these options live and recommended trying new backbones and training setups for better results.

Not every stream ends with a win, but every stream is progress. Embrace the iteration loop it’s the only way to improve

📺 Replay: https://youtu.be/aJeEKZMkwNs?si=uQyQ5SU7ws-huVcc

🎮 Next stream: https://www.twitch.tv/koleshjr/schedule

📢 Subscribe: https://www.youtube.com/@koleshjr

Discussion 0 answers