New AI Trick Helps Voice Assistants Learn New Accents Without Forgetting Old Ones

Speech recognition AI — the technology behind dictation apps, call-center transcribers, and voice assistants — usually gets trained once on huge amounts of audio, then left alone. But companies often want to teach it something new: a regional accent, a technical vocabulary, a second language. The problem is that when you teach these models something new, they tend to forget what they already knew. Researchers call this "catastrophic forgetting." It's like a chef who learns to make sushi and suddenly forgets how to bake bread.

A team at KU Leuven, led by Steven Vander Eeckt and Hugo Van hamme, has published a fix. Their method, CGaLore, builds on an existing memory-saving technique called GaLore, which shrinks the amount of temporary storage needed while training an AI. Think of it as packing a suitcase with compression bags instead of paying for extra luggage. Their twist: before deciding what to focus on when learning something new, the system checks how sensitive the model's existing knowledge is, and protects the parts it shouldn't disturb. The result — less forgetting, and better results than current best methods, according to their experiments.

The practical payoff is about cost and access. Training or updating a large speech model normally requires expensive, power-hungry hardware that only big companies can afford. Memory-efficient methods like this lower that bar, which means smaller companies, universities, and even clinics or nonprofits could adapt speech AI to their own users — say, a hospital tuning dictation software to local medical terms. Less memory also means lower running costs and less energy use.

The catch: this is a research paper accepted at an academic conference (SLT 2026), not a product you can use today. Results come from lab tests, and real-world audio is messier — background noise, dozens of overlapping accents, poor phone connections. The technique also still requires technical expertise to apply. So expect this thinking to show up in future voice tools rather than a download this week.