Voice Recognition Dynamic Filter Bank Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing voice recognition systems face inefficiencies due to fixed filtering and normalization parameters that do not adequately account for variations in vocal tract length among speakers, leading to suboptimal speech recognition accuracy and long computation times for large coefficient matrices.
Innovation Solution
Determine the runtime pitch of a speaker's voice signal to dynamically adjust acoustic model parameters, such as minimum and maximum frequencies, for more accurate voice recognition analysis.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fixed filtering parameters are used in voice recognition systems, then device complexity is reduced, but speech recognition accuracy deteriorates due to inability to account for variations in vocal tract length among speakers
Solution Approach 1:
The patent implements dynamic adjustment of filtering parameters (fmin and fmax) based on runtime pitch detection and speaker categorization. Instead of using fixed filter banks, the system adapts the frequency range parameters in real-time according to the detected speaker characteristics, allowing the filtering process to dynamically match the speaker's vocal tract length and improve recognition accuracy.
Solution Approach 2:
The system changes the parameters of the filter bank (specifically fmin and fmax) based on detected speaker characteristics. By categorizing speakers into different groups (e.g., male/female, adult/child) based on runtime pitch, the system selects appropriate parameter sets that optimize the frequency range for each speaker type, thereby improving measurement precision without requiring complex individualized adjustments.
2Measurement precision
If vocal tract normalization with large coefficient matrices is performed, then speech recognition accuracy is improved, but computation time increases significantly
Solution Approach 1:
The patent segments the continuous space of vocal tract lengths into discrete categories (e.g., male adult, female adult, male child, female child). Instead of computing a single large normalization matrix for all possible vocal tract lengths, the system creates separate, smaller normalization matrices for each category. This segmentation reduces the computational burden while maintaining accuracy for each specific group.
Solution Approach 2:
The system performs preliminary categorization of the speaker based on runtime pitch detection before applying vocal tract normalization. By determining the speaker category in advance, the system can select the pre-computed normalization matrix that best matches the speaker's characteristics, avoiding the need to compute a large general-purpose matrix during real-time processing.
3Adaptability or versatility
If fixed acoustic model parameters are used, then device complexity is reduced, but adaptability to different speakers deteriorates
Solution Approach 1:
The system changes acoustic model parameters (fmin and fmax) based on detected speaker characteristics. By categorizing speakers into different groups based on runtime pitch and selecting appropriate parameter sets for each category, the system achieves adaptability to different speakers without implementing complex individualized parameter optimization for each speaker.
Data Source
AI summary
Voice recognition methods and systems are disclosed. A voice signal is obtained for an utterance of a speaker. A runtime pitch is determined from the voice signal for the utterance. The speaker is categorized based on the runtime pitch and one or more acoustic model parameters are adjusted based on a categorization of the speaker. The parameter adjustment may be performed at any instance of time during the recognition. A voice recognition analysis of the utterance is then performed based on the acoustic model.