Researchers built an AI model that learns new accents without forgetting old ones
The new AccentCL framework lets AI identify English accents incrementally as new data arrives.
Most AI systems struggle when you try to teach them new categories after their initial training. They often suffer from 'catastrophic forgetting,' where new information overwrites what they already know. A new research paper describes a system called AccentCL that fixes this issue specifically for recognizing regional English accents.
AccentCL uses the well-known Whisper-Large-v3 model as its foundation. Instead of retraining the whole model from scratch, the researchers designed a way to expand its knowledge incrementally. This is useful because speech data often comes from different sources and recordings, which can confuse standard models.
AccentCL enables robust regional accent classification while allowing new accent categories to be added without full retraining.
The framework addresses two major problems in voice recognition: class imbalance and domain shift. Imbalance happens when a model hears more of one accent than others and starts favoring the majority. Domain shift occurs when recording conditions change, such as background noise or microphone quality variations.
To keep the model sharp, the researchers implemented an 'imbalance-aware' loss function. This forces the system to pay equal attention to all accents regardless of how much training data exists for each. They also used a 'domain mean alignment' technique to keep the model consistent across different types of audio files.
In testing, the team successfully taught the model to identify five common English accents. They then added two new categories—Spanish-accented and Chinese-accented English—without harming the model's accuracy on the original group. The model retained high performance even after these additions.
The system is not perfect. It still faces challenges in balancing accuracy when adding many new accents over a long period. Furthermore, the tests were limited to these specific linguistic categories, and the researchers have not yet demonstrated how it handles other global dialects or extremely noisy environments. Have you ever wondered how an AI determines where a speaker is from just by the sound of their voice?
Source: arXiv cs.CL