buzzdetect live!
Try out the app here.
buzzdetect live! (GitHub repo) is a little webapp that lets you run buzzdetect analyses from your browser. Why? Because! Mostly, it’s a fun demo tool that I might expand into an educational tool. Somewhat, I would like to be able to check what buzzdetect thinks of my acoustic environment in realtime.
It doesn’t really make sense as a research tool; the real buzzdetect is way faster and this one can’t work in batches. It doesn’t really make sense as a results exploration tool. SeeNote is way cooler and way more powerful. It doesn’t really make sense as a field tool, because if you’re pointing your phone at a bee you presumably know that the bee is there, and thus don’t need to detect its buzzes (and could just use, you know, a microphone).
And yet, I find it strangely compelling and stupidly entertaining. I’ve already been deep diving into buzzdetect results, but watching the model process in real-time is like having a conversation with it. It’s like I’m learning how my model thinks. And, incidentally, I’ve been finding that my model doesn’t think remotely in the same way as I do. You’ll notice some strange patterns in the neuron activations. For example, all the neurons are strongly anticorrelated, even though we train the model to be simultaneous-event-aware. That’s probably YAMNet’s fault and there’s not much we can do about it if it’s baked into the embeddings. It’s also interesting to try to figure out what the model is looking for by humming like a bee and modulating the duration, sequence, pitch shifting, raspiness… Yes, I’m sitting in my room buzzing at my phone. I’m doing research, mom! It’s data science, sheesh you don’t get it!