Voice Recognition Device Noise Likelihood Recalculation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Voice recognition systems face performance degradation when feature values used in creating noise patterns differ significantly from those encountered during voice recognition, leading to erroneous noise recognition as voice.
Innovation Solution
A voice recognition device that calculates first and second feature vectors from input voice data, uses acoustic and discriminant models to assess phoneme likelihoods and noise degrees, and recalculates noise likelihoods to ensure accurate pattern matching with vocabulary words, thereby preventing misrecognition.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If the likelihood of the garbage model is corrected based on feature values of input data, then erroneous recognition of noise as voice is reduced, but when feature values are greatly different between noise data used in garbage model creation and noise data at recognition time, the likelihood of the standard pattern of voice becomes higher than that of the standard pattern of noise
Solution Approach 1:
The invention changes the parameter used for noise likelihood calculation from raw acoustic likelihood to a corrected value that incorporates noise degree information. Specifically, it calculates a corrected noise likelihood by adding the maximum phoneme likelihood to the noise degree (which is the difference between noise likelihood and voice likelihood from the discriminant model). This parameter transformation allows the system to adapt to different noise conditions by dynamically adjusting the noise likelihood based on the actual noise characteristics detected during recognition, rather than relying solely on pre-created garbage models trained on specific noise data.
2Reliability
If standard patterns for all noises are created in advance, then recognition accuracy under various noise conditions improves, but it becomes difficult to create standard patterns for all types of noise
Solution Approach 1:
The invention enables the voice recognition device to automatically adjust its noise likelihood calculations without requiring external preparation of comprehensive noise standard patterns. The discriminant model (trained once on voice and noise data) serves the system continuously by providing noise degree measurements during recognition. The system self-adjusts by calculating corrected noise likelihoods using the formula: corrected noise likelihood = maximum phoneme likelihood + noise degree. This eliminates the need for manual creation and maintenance of multiple noise standard patterns while maintaining high recognition accuracy across diverse noise conditions.
3Measurement precision
If margin sections are added before and after utterance for voice section cutting, then accurate voice section extraction is achieved, but the margin sections may include only ambient environmental noise
Solution Approach 1:
The invention introduces the discriminant model as an intermediary component that evaluates whether margin sections contain voice or noise. The discriminant model (separately trained on voice and noise characteristics) provides an objective assessment of the margin sections. Based on this intermediary evaluation, the system can adjust the noise likelihood for margin sections using the corrected calculation method, thereby handling noise-contaminated margins more effectively without compromising voice section extraction accuracy.
Data Source
AI summary
A voice recognition device includes: a first feature vector calculating unit (2) for calculating a first feature vector from voice data input; an acoustic likelihood calculating unit (4) for calculating an acoustic likelihood of the first feature vector by using an acoustic model used for calculating an acoustic likelihood of a feature vector; a second feature vector calculating unit (3) for calculating a second feature vector from the voice data; a noise degree calculating unit (6) for calculating a noise degree of the second feature vector by using a discriminant model used for calculating a noise degree indicating whether a feature vector is noise or voice; a noise likelihood recalculating unit (8) for recalculating an acoustic likelihood of noise on the basis of the acoustic likelihood of the first feature vector and the noise degree of the second feature vector; and a collation unit (9) for performing collation with a pattern of a vocabulary word to be recognized, by using the acoustic likelihood calculated and the acoustic likelihood of noise recalculated, and outputting a recognition result of the voice data.