Voice Processing Device Dynamic Feature Update
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional speaking person recognition technologies face challenges in accurately identifying voices over time due to changes in voice quality, such as the Lombard effect, and inefficiently manage registered features, leading to erroneous determinations and increased processing loads.
Innovation Solution
A voice processing device that utilizes a deep neural network to calculate features from input voice signals and compares them with registered features, updating or adding them based on similarity thresholds to maintain accurate identification, while managing feature updates to reduce memory capacity and processing load.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If registered features are continuously updated with new voice data, then voice recognition accuracy is improved, but memory capacity consumption increases
Solution Approach 1:
The system changes the state of registered features from static to dynamic by implementing automatic updates when similarity thresholds are met. When a new voice sample sufficiently resembles an existing registered feature, the system updates that feature's parameters to incorporate the new data, thereby improving recognition accuracy without unbounded memory growth
Solution Approach 2:
The system discards redundant voice data by updating existing registered features rather than continuously adding new ones. When similarity exceeds the threshold, the new voice sample is used to refresh/update the existing feature representation, effectively discarding the need to store separate copies while recovering and maintaining recognition accuracy
2Reliability
If multiple registered features are maintained for each person, then voice recognition reliability is improved, but processing load increases
Solution Approach 1:
The system implements dynamic feature registration where the number and state of registered features per person is not fixed but adapts based on similarity thresholds. This dynamic approach allows the system to maintain multiple features when necessary for reliability while avoiding unnecessary feature proliferation that would increase processing load
Solution Approach 2:
The system uses similarity threshold parameters to control feature registration decisions. By adjusting these threshold parameters, the system can optimize the balance between maintaining sufficient features for reliable recognition and limiting the total number of features to manage processing load efficiently
3Adaptability or versatility
If voice features are updated frequently, then adaptation to voice quality changes is improved, but erroneous determinations increase
Solution Approach 1:
The system applies preliminary anti-action by establishing similarity thresholds that prevent premature or inappropriate feature updates. Before updating a registered feature, the system first checks whether the new voice sample meets the similarity criterion, thereby preventing erroneous determinations that would result from updating with insufficiently similar or incorrect voice samples
Solution Approach 2:
The system uses feedback through similarity threshold comparison to control the feature update process. Each new voice sample is compared against existing registered features, and updates only occur when the similarity feedback indicates sufficient match, thereby adapting to voice quality changes while maintaining reliability by filtering out poor matches
Data Source
AI summary
A voice processing device includes a calculation unit and a determination processing unit. The calculation unit calculates a first feature being a feature of an input voice signal. When a similarity between the first feature and a second feature out of one or more registered features having been registered is equal to or larger than a first threshold, the determination processing unit makes determination that the input voice signal is a voice of a first registered person out of registered persons. The first registered person corresponds to the second feature. When the similarity is equal to or larger than the first threshold and smaller than a second threshold, the determination processing unit adds the first feature to the registered features or updates the registered features with the first feature.