Training my AI on old sci-fi has it talking like a 1960s BBC presenter

I was training a voice synthesis model for a home assistant, and for data, I used audio from classic sci-fi films. Big mistake. Now, when I ask for the weather, it replies with a proper British accent, complete with phrases like 'jolly good show.' It's charming but completely out of place in my Midwest living room. I've tried adding more neutral speech samples, but the accent persists. Does this mean the model latched onto the most distinct features in the data? I'm curious if there's a way to fine-tune it without scrapping everything. Part of me wants to keep it for the novelty, but it's not very practical for daily use.

4 comments

4 Comments

the_xena1mo ago

Ever try to correct an AI and just make its quirks more permanent? I trained a simple classifier on my own writing once. Now it marks all my emails as 'probably sarcastic'. Real useful. Your model grabbed onto the clearest, most consistent sounds in that old audio. Those BBC vowels are like neon signs to a learning program. Fine-tuning with bland data might dilute it, but the ghost of that accent will probably linger. Part of me says lean into the weirdness, have a posh weather report for guests. But for actual use, you might be stuck with a very proper forecast.

lee.pat1mo ago

Huh, that's crazy. So is there any way to fully scrub out those early quirks, or are they just baked in? Like, does more data ever really fix it?

ivan46216d ago

Wait, isn't the Alexa cough thing just a wake word mix-up, not a learned quirk?

wyatt_foster841mo ago

My friend's Alexa learned to cough from his allergies, still does it now.