Speech Recognition Using Block-Parallel Bidirectional LSTM-RNN
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Bidirectional Long Short-term Memory Recurrent Neural Networks (BLSTM-RNN) face a significant time delay in processing speech recognition, making real-time recognition impossible due to the need to wait for the entire sentence to be read before performing bidirectional processing.
Innovation Solution
The method involves dividing the frame-level speech feature sequence into blocks with a master and slave block structure, allowing parallel input into a preset bidirectional LSTM-RNN model for both forward and backward recognition, enabling immediate processing and reducing time delay by using 'past' and 'future' speech information simultaneously.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If bidirectional LSTM-RNN is used for speech recognition, then recognition accuracy is improved, but time delay increases making real-time recognition impossible
Solution Approach 1:
The patent divides the speech feature sequence into multiple blocks, where each block contains a master block and a slave block. The master block is processed immediately while the slave block is processed in parallel, eliminating the need to wait for the entire sequence. This segmentation allows bidirectional processing to occur in real-time by breaking down the global dependency into local parallel operations.
Solution Approach 2:
The patent performs preliminary processing by dividing the speech sequence into blocks and pre-identifying master and slave blocks before actual recognition. This allows the system to prepare processing pipelines in advance, so when speech data arrives, the bidirectional LSTM-RNN can immediately process master blocks while slave blocks are being prepared, reducing overall latency.
2Measurement precision
If the entire sentence is read before bidirectional processing, then accurate recognition is achieved, but processing speed decreases
Solution Approach 1:
The patent segments the speech sequence into multiple independent blocks that can be processed in parallel. Each block contains a master block processed immediately and a slave block processed in parallel, eliminating the sequential bottleneck. This allows the system to maintain high processing speed while achieving bidirectional processing accuracy through localized forward-backward computations within each block.
Solution Approach 2:
The patent introduces a parallel processing dimension by creating master-slave block pairs that operate simultaneously. Instead of processing the entire sequence in a single temporal dimension, the system adds a parallel execution dimension where multiple blocks are processed concurrently, significantly improving processing speed while maintaining recognition accuracy through the bidirectional LSTM-RNN architecture.
Data Source
AI summary
Disclosed are a speech recognition method, apparatus, computer device and storage medium. The method includes: performing a framing and an acoustic feature extraction of a speech-information-to-be-tested according to a default rule to obtain a frame-level speech feature sequence; dividing the frame-level speech feature sequence into n blocks sequentially; inputting all blocks into a preset bidirectional LSTM-RNN model parallelly to obtain an output result of the corresponding neuron in an output layer of the preset bidirectional LSTM-RNN model corresponding to the forward recognition result and backward recognition result of each block to obtain a speech recognition result of the speech-information-to-be-tested. The present application can improve the speech recognition effect significantly and reduce the time delay of the speech decoding effectively.