Unified DNN for Keyword Detection with Shared Hidden Layers
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speech recognition techniques using Hidden Markov Models (HMMs) face inefficiencies in calculating output probabilities for triphones and monophones, leading to increased computational costs due to the need for separate Deep Neural Networks (DNNs) for each, which complicates the learning and execution processes.
Innovation Solution
A keyword detector that uses a single Deep Neural Network (DNN) with shared input and hidden layers for both triphones and monophones, where the output layer includes separate neurons for each, allowing for simultaneous learning and reducing computational requirements by leveraging correlations between triphone and monophone parameters.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If separate DNNs are used for triphones and monophones, then output probability calculation accuracy is improved, but device complexity and computational load increase
Solution Approach 1:
The patent merges the previously separate DNNs for triphones and monophones into a single unified DNN. This unified DNN shares common parameters including the input layer structure, hidden layer architectures, and optimization algorithms, while maintaining separate output layers for triphone and monophone predictions. This merging reduces device complexity and computational load while preserving the accuracy benefits of having specialized networks for each phoneme type.
Solution Approach 2:
The unified DNN is designed to perform multiple functions simultaneously - it can predict both triphone output probabilities and monophone output probabilities within a single network structure. The network accepts the same input features and produces multiple types of output predictions, making it a universal model that replaces multiple specialized models. This multi-functionality is achieved through shared weight matrices and bias terms that are optimized for both prediction tasks.
2Measurement precision
If separate DNNs are learned independently for triphones and monophones, then recognition accuracy is improved, but learning time increases
Solution Approach 1:
The learning processes for triphones and monophones are merged into a single optimization routine. The unified DNN uses shared parameters that are updated simultaneously through backpropagation based on both triphone and monophone prediction errors. This combined learning approach reduces the total learning time compared to independent training of separate networks, while still achieving high recognition accuracy through the specialized output layers.
Solution Approach 2:
The unified DNN enables continuous learning of both triphone and monophone representations in a single iterative process. During each training iteration, the network processes training data and updates its parameters continuously for both phoneme types simultaneously, rather than alternating between separate training phases. This continuous dual-learning approach maintains high recognition accuracy while significantly reducing the total time required to train the system.
3Measurement precision
If separate DNNs are used for triphones and monophones, then output probability precision is improved, but computational load during execution increases
Solution Approach 1:
The inference process merges the previously separate computational paths into a single forward propagation through the unified DNN. The network computes both triphone and monophone output probabilities in one pass, sharing the computational workload across the hidden layers. This eliminates the need to run two separate networks during execution, reducing the computational load and power consumption while maintaining the precision benefits of specialized predictions for each phoneme type.
Data Source
AI summary
A keyword detector includes a processor configured to calculate a feature vector for each frame from a speech signal, input the feature vector for each frame to a DNN to calculate a first output probability for each triphone according to a sequence of phonemes contained in a predetermined keyword and a second output probability for each monophone, for each of at least one state of an HMM, calculate a first likelihood representing the probability that the predetermined keyword is uttered in the speech signal by applying the first output probability to the HMM, calculate a second likelihood for the most probable phoneme string in the speech signal by applying the second output probability to the HMM, and determine whether the keyword is to be detected on the basis of the first likelihood and the second likelihood.