LSTM Recurrent Projection Layers for Acoustic Sequence Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Acoustic modeling systems face challenges in accurately generating phoneme representations of acoustic sequences due to high dimensionality of data, which hinders efficient training and processing, and limits the addition of LSTM memory blocks without increasing training and processing time.
Innovation Solution
The implementation of long short-term memory (LSTM) layers with recurrent and non-recurrent projection layers to reduce data dimensionality, allowing for efficient processing and training by projecting LSTM outputs into lower-dimensional spaces, thereby facilitating the inclusion of additional LSTM memory blocks without compromising training and processing efficiency.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If high dimensionality acoustic data is processed using traditional methods, then processing accuracy is maintained, but training time and processing time increase significantly
Solution Approach 1:
The patent extracts and removes redundant dimensions from the high-dimensional acoustic feature representations by projecting them into a lower-dimensional space using projection layers. This extraction of essential features while discarding redundant information reduces the data dimensionality, thereby decreasing training time and processing time while preserving the accuracy needed for phoneme representation.
2Measurement precision
If additional LSTM memory blocks are added to improve phoneme representation accuracy, then processing accuracy improves, but training time and processing time increase
Solution Approach 1:
By extracting and removing redundant features through projection layers, the patent reduces the dimensionality of data flowing through LSTM memory blocks. This allows for the addition of more LSTM blocks to improve accuracy without proportionally increasing processing time, as each block operates on compressed, lower-dimensional representations.
Solution Approach 2:
The patent transforms the data from high-dimensional space to lower-dimensional space through projection layers. This dimensionality change reduces the computational burden on LSTM memory blocks, enabling the system to incorporate additional blocks for improved accuracy while maintaining processing efficiency.
3Productivity
If data dimensionality is reduced using projection layers, then processing efficiency improves, but information loss may occur
Solution Approach 1:
The patent changes the parameters of the projection layers during training to optimize the balance between dimensionality reduction and information preservation. By adjusting projection matrices and regularization parameters, the system learns to retain the most informative features while discarding redundant ones, thereby maintaining processing efficiency without significant information loss.
Solution Approach 2:
The patent employs feedback mechanisms during training where the system monitors the quality of phoneme representations and adjusts the projection layer parameters accordingly. This feedback loop ensures that dimensionality reduction does not lead to excessive information loss, as the projection parameters are continuously optimized to preserve essential acoustic features.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on computer storage media, for generating phoneme representations of acoustic sequences using projection sequences. One of the methods includes receiving an acoustic sequence, the acoustic sequence representing an utterance, and the acoustic sequence comprising a respective acoustic feature representation at each of a plurality of time steps; for each of the plurality of time steps, processing the acoustic feature representation through each of one or more long short-term memory (LSTM) layers; and for each of the plurality of time steps, processing the recurrent projected output generated by the highest LSTM layer for the time step using an output layer to generate a set of scores for the time step.