DNN Acoustic Model Training via Frame Segmentation

The patent segments the acoustic context into central frames (relevant) and side frames (less relevant). The DNN input layer is structured to receive different frame types, and the training process learns to weigh them differently, effectively segmenting the influence of different acoustic contexts on the final recognition result.

The patent applies local quality by treating central frames and side frames differently in the training process. Central frames receive full attention during both pre-training and fine-tuning, while side frames have their connections regularized during fine-tuning. This creates different quality levels of feature utilization across different parts of the input data.

The patent applies preliminary action by first performing pre-training on the entire DNN using all input frames (central and side) without regularization. This establishes a baseline model that captures general acoustic patterns. Only after this preliminary training is complete does the fine-tuning phase apply regularization selectively to side frame connections, preventing over-fitting at the appropriate stage.

The patent applies partial action by selectively regularizing only the connections from side frames during fine-tuning, while leaving the connections from central frames unregularized. This partial application of regularization targets specifically the problematic source of over-fitting (side frames) without unnecessarily constraining the beneficial central frame connections.