Sound Recognition Using Hidden Vector State Model
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current sound recognition methods in intelligent sound boxes, such as those using hidden Markov models combined with Gaussian mixture models, suffer from low recognition accuracy due to difficulties in distinguishing between preset users and unauthorized individuals, leading to potential misuse of credit card transactions.
Innovation Solution
The implementation of a sound recognition method utilizing a hidden vector state model based on discriminative training with maximum mutual information, which extracts mel-frequency cepstral coefficient sound characteristics to differentiate between registered and unauthorized users, enhancing the accuracy of sound recognition and preventing unauthorized access.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If GMM-HMM technology with maximum likelihood sound recognition method is used, then the sound recognition system can process continuous sound and find most probable words, but the recognition probability of unauthorized users may be larger than preset users under certain factors, causing lower recognition accuracy
Solution Approach 1:
The patent changes the fundamental parameter of the recognition method from maximum likelihood estimation to minimum distance classification. Instead of finding the most probable word sequence based on probability models, the system calculates the actual distance between the input sound feature vector and pre-stored standard sound feature vectors of authorized users, enabling more accurate identification by directly measuring feature similarity rather than relying on probabilistic assumptions
Solution Approach 2:
The patent creates standard sound feature vectors as copies of authorized users' voice characteristics during a registration phase. These stored feature vectors serve as reference templates that are later compared against incoming sound inputs, allowing the system to recognize users by matching their voice patterns against these pre-captured copies rather than attempting probabilistic recognition in real-time
Data Source
AI summary
A sound recognition device includes a collecting module, an extracting module connected with the collecting module, a training module, a storage module, a decoding module and a processor module. The collecting module is for receiving sound information of a specific target and a target to be measured. The extracting module is used for extracting sound characteristics of the specific target and the target to be measured. The training module is connected with the extracting module. Regard the sound characteristics of the sound information of the specific target and the target to be measured as input data of a hidden vector state (HVS) model. The storage module is connected with the training module. The decoding module is used for proceeding a language decoding procedure on sound information of the target to be measured. The processor module is connected with the training module, the storage module and the decoding module.