DNN Speech Recognition Using Heterogeneous Wideband Narrowband Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Automatic speech recognition (ASR) systems face challenges in obtaining suitable training data due to the scarcity of wideband signals, leading to suboptimal approaches like down-sampling or up-sampling, which can introduce errors or fail to utilize the full information in heterogeneous data.
Innovation Solution
A deep neural network (DNN) is trained using heterogeneous data that includes both wideband and narrowband signals, where missing features in narrowband signals are automatically populated using zero-padding or mean-padding, allowing the system to perform recognition tasks regardless of the input signal quality.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If wideband signals are down-sampled to narrowband signals for training, then the ASR system can be trained using available narrowband processing methods, but the additional information in wideband data is lost
Solution Approach 1:
The patent changes the parameter representation by introducing a heterogeneous data framework where wideband and narrowband signals are represented with different feature dimensions. Wideband signals use full feature vectors while narrowband signals use reduced feature vectors, allowing the system to process both types without information loss or complex conversion.
2Loss of information
If narrowband signals are up-sampled to extend bandwidth, then the ASR system can utilize more frequency information, but the procedures become complicated and introduce errors
Solution Approach 1:
The patent segments the training data into distinct heterogeneous categories (wideband and narrowband) with different feature representations. Instead of attempting to uniformly up-sample or convert all signals, the system processes each segment according to its native characteristics, simplifying the overall processing pipeline while preserving information.
3Reliability
If two separate ASR systems are built (one for wideband and one for narrowband), then each system can be optimized for its specific data type, but the overall system complexity increases and narrowband data abundance cannot be fully exploited
Solution Approach 1:
The patent creates a universal ASR system that can handle both wideband and narrowband inputs through a single heterogeneous data framework. The system uses adaptive feature vector processing where the feature dimension is automatically adjusted based on input signal type, eliminating the need for separate systems while maintaining optimization for each data type.
4Reliability
If a DNN is trained only on wideband signals, then the system can achieve high recognition accuracy for wideband input, but it cannot effectively process narrowband signals due to lack of training data
Solution Approach 1:
The patent introduces dynamic adaptability by training the DNN on heterogeneous data with varying feature dimensions. The network learns to dynamically adjust its processing based on the input signal characteristics, enabling it to effectively handle both wideband and narrowband signals rather than being static and specialized for only one type.
Data Source
AI summary
Technologies pertaining to training a deep neural network (DNN) for use in a recognition system are described herein. The DNN is trained using heterogeneous data, the heterogeneous data including narrowband signals and wideband signals. The DNN, subsequent to being trained, receives an input signal that can be either a wideband signal or narrowband signal. The DNN estimates the class posterior probability of the input signal regardless of whether the input signal is the wideband signal or the narrowband signal.