BRDNN Acoustic Score Calculation for Incremental Speech Recognition

The patent divides the speech signal into multiple windows and processes each window independently through the BRDNN. By segmenting the speech into manageable frames and processing them in sequence rather than requiring the entire speech signal at once, the system enables incremental decoding while maintaining the accuracy benefits of bidirectional processing within each window context.

The patent applies padding frames to both sides of each speech window before processing. This preliminary action provides the BRDNN with bidirectional context information in advance for each window, allowing the network to make accurate pronunciation probability calculations without waiting for the entire speech signal. The padding frames contain pre-calculated acoustic scores that provide contextual information about surrounding speech segments.

The patent recalculates acoustic scores for overlapping padding frames by combining results from adjacent windows using statistical methods. This approach recovers computational efficiency by reusing previously calculated scores for overlapping regions rather than recalculating them independently, reducing redundant computations while maintaining context accuracy.

The patent uses statistical methods (arithmetic mean or geometric mean) to combine acoustic scores from overlapping regions of adjacent windows. This copying and averaging approach ensures consistency across window boundaries by using the same statistical operation for all overlapping frames, maintaining score reliability while enabling parallel processing of non-overlapping regions.