Dual Deep Neural Network Scoring for Automatic Speech Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition systems face challenges in reducing computational complexity and real-time factors while maintaining accuracy, particularly when using deep neural networks as scoring mechanisms.
Innovation Solution
The implementation of a method that utilizes two deep neural networks with varying complexity, where the second network shares trained weights from the first network and includes fewer parameters and hidden layers, allowing for dynamic selection based on phonetic class conditional probabilities, entropy, and phonetic look-ahead information to optimize computation and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a first deep neural network with many parameters is used for scoring, then accuracy is improved, but computational complexity increases
Solution Approach 1:
The system segments the scoring task by using multiple DNNs with different complexities (first DNN with many parameters, second DNN with fewer parameters) instead of a single monolithic network. This allows the system to divide computational work based on specific needs.
Solution Approach 2:
The system dynamically selects between the first and second DNN based on phonetic class conditional probabilities, entropy, and phonetic look-ahead information. This dynamic adaptation allows the system to use computational resources efficiently while maintaining accuracy when needed.
2Measurement precision
If a first deep neural network with many parameters is used for scoring, then accuracy is improved, but real-time factor worsens
Solution Approach 1:
The system dynamically adapts between first and second DNN based on phonetic class conditional probabilities, entropy, and phonetic look-ahead information. This allows the system to use the computationally lighter second DNN when accuracy requirements are met, improving real-time performance.
Solution Approach 2:
The system changes the parameter complexity of the DNN being used by selecting between first DNN (many parameters) and second DNN (fewer parameters) based on acoustic characteristics. This parameter adaptation allows balancing accuracy and speed in real-time processing.
3Device complexity
If a second deep neural network with fewer parameters is used, then computational complexity is reduced, but accuracy deteriorates
Solution Approach 1:
The system dynamically selects between first and second DNN based on phonetic class conditional probabilities, entropy, and phonetic look-ahead information. This allows the system to use the second DNN (fewer parameters) only when it provides sufficient accuracy, switching to the first DNN when higher accuracy is needed.
Solution Approach 2:
The system uses phonetic class conditional probabilities, entropy calculations, and phonetic look-ahead information as intermediary mechanisms to determine when the second DNN is sufficient and when the first DNN is needed, mediating between computational efficiency and accuracy requirements.
4Measurement precision
If multiple deep neural networks are used for scoring, then accuracy is maintained or improved, but device complexity increases
Solution Approach 1:
The system segments the scoring function across multiple DNNs with different complexities, allowing each network to be optimized for specific purposes while collectively maintaining high accuracy.
Solution Approach 2:
Both the first and second DNNs are trained to perform the same phonetic classification task, with the second DNN serving as a lighter alternative. This multi-functionality allows the system to use either network depending on requirements, managing complexity while maintaining accuracy.
Data Source
AI summary
A system and method for speech recognition is provided. Embodiments may include receiving an audio signal at a first deep neural network (“DNN”) associated with a computing device. Embodiments may further include receiving the audio signal at a second deep neural network (“DNN”) associated with a computing device, wherein the second deep neural network includes fewer parameters than the first deep neural network. Embodiments may also include determining whether to select an output from the first deep neural network or the second deep neural network and providing the selected output to a decoder with an overall objective of speeding up ASR.