Speech Feature Extraction Using Linear Domain Delta Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition technologies face challenges in accurately recognizing speech in noisy and reverberant environments due to the limitations of existing feature extraction methods, particularly the use of delta features in the logarithmic domain, which are not robust enough to handle wide dynamic ranges and transfer characteristics.
Innovation Solution
The proposed solution involves extracting delta and delta-delta features in the linear domain and normalizing them using the average spectrum of overall speech, which addresses the issues of wide dynamic ranges and transfer characteristics, improving robustness against noise and reverberation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If delta features in the logarithmic domain are used for speech recognition, then the recognition accuracy in clean environments is maintained, but the robustness against noise and reverberation deteriorates
Solution Approach 1:
The patent changes the domain parameter from logarithmic to linear domain for delta feature extraction. Specifically, it computes delta features by taking the difference of linear spectra rather than logarithmic spectra, and applies normalization by dividing by the average linear spectrum. This parameter change enables the features to better represent speech characteristics in noisy and reverberant environments while maintaining computational feasibility.
2Reliability
If features robust to noise and reverberation are extracted, then the speech recognition accuracy in challenging environments improves, but the complexity of the feature extraction process increases
Solution Approach 1:
The patent performs preliminary computation of the average spectrum over the entire speech signal before extracting delta features. This average spectrum is then used as a normalization factor for all subsequent delta feature calculations. By preparing this normalization reference in advance, the method avoids complex adaptive processing during feature extraction, thereby reducing overall system complexity while maintaining robustness.
3Reliability
If the average spectrum normalization is applied to delta features in the linear domain, then the robustness against noise and reverberation improves, but the computational overhead increases
Solution Approach 1:
The patent changes the normalization approach by operating in the linear domain rather than the logarithmic domain. Specifically, it computes the average linear spectrum and uses it to normalize the delta features directly in the linear domain. This parameter change simplifies the mathematical operations required compared to traditional logarithmic domain normalization, thereby reducing computational energy consumption while achieving superior robustness.
Data Source
AI summary
A speech feature extraction apparatus, speech feature extraction method, and speech feature extraction program. A speech feature extraction apparatus includes: first difference calculation module to: (i) receive, as an input, a spectrum of a speech signal segmented into frames for each frequency bin; and (ii) calculate a delta spectrum for each of the frame, where the delta spectrum is a difference of the spectrum within continuous frames for the frequency bin; and first normalization module to normalize the delta spectrum of the frame for the frequency bin by dividing the delta spectrum by a function of an average spectrum; where the average spectrum is an average of spectra through all frames that are overall speech for the frequency bin; and where an output of the first normalization module is defined as a first delta feature.