Neural Network Speech Recognition Model Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional automatic speech recognition (ASR) systems fail to effectively model dependencies between frames in audio data sequences, relying on generalized approximations that do not utilize the specific context of the current sequence, leading to suboptimal recognition performance.
Innovation Solution
The use of artificial neural networks (ANNS) to dynamically generate customized speech recognition models during the recognition process, taking into account the specific context of individual frames by processing a window of previous frames and state information, thereby generating probability density functions (PDFs) that model observation probabilities and state transition probabilities without relying on discrete state clustering.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional ASR systems use generalized approximations without modeling dependencies between frames, then device complexity is reduced, but recognition performance deteriorates
Solution Approach 1:
The system dynamically generates customized speech recognition models during the recognition process by using neural networks to generate probability density functions that adapt to the specific context of individual frames. This dynamic model generation allows the system to capture dependencies between frames while maintaining computational efficiency, resolving the contradiction between improved recognition performance and reduced model complexity.
Solution Approach 2:
The patent applies local quality by generating customized models for each frame based on its specific context rather than using a single generalized model for all frames. The neural network generates probability density functions that are tailored to the local characteristics of each frame, allowing the system to improve recognition performance where it matters most without uniformly increasing complexity across the entire system.
2Reliability
If offline state clustering is used to simplify the system, then device complexity is reduced, but recognition performance deteriorates due to loss of specific sequence context
Solution Approach 1:
The system eliminates the need for offline state clustering by using neural networks to automatically generate probability density functions during the recognition process itself. The system serves its own modeling needs dynamically, adapting to each sequence without requiring pre-computed clustered states, thus improving recognition performance while simplifying the system setup by removing the offline clustering step.
3Reliability
If non-linear feature transformations are applied to improve recognition, then recognition performance is improved, but system latency increases
Solution Approach 1:
The patent replaces traditional non-linear feature transformations with a neural network-based probability density function generation approach. This substitution maintains the ability to capture complex patterns and improve recognition performance while reducing computational overhead and system latency, as the neural network can be efficiently trained and deployed for real-time processing.
Data Source
AI summary
Features are disclosed for using an artificial neural network to generate customized speech recognition models during the speech recognition process. By dynamically generating the speech recognition models during the speech recognition process, the models can be customized based on the specific context of individual frames within the audio data currently being processed. In this way, dependencies between frames in the current sequence can form the basis of the models used to score individual frames of the current sequence. Thus, each frame of the current sequence (or some subset thereof) may be scored using one or more models customized for the particular frame in context.