Dynamic Speaker Set Reduction for Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The accuracy of speaker recognition decreases with an increasing number of registered speakers, leading to higher possibilities of misidentification, as not all participants may need to be considered as constant target speakers.
Innovation Solution
An information processing method that detects speech segments, extracts feature quantities, compares them with stored values, and removes registered speakers with low similarity thresholds, allowing for dynamic adjustment of registered speakers based on consecutive speech segments and predetermined conditions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If the number of registered speakers is increased to cover all participants, then the completeness of speaker identification is improved, but the accuracy of speaker recognition decreases due to higher possibility of misidentification
Solution Approach 1:
The system dynamically adjusts the set of registered speakers during the recognition process. Instead of using a fixed large set of all participants, the system continuously updates which speakers are currently active based on speech detection in consecutive segments, thereby maintaining high accuracy while preserving completeness of identification.
Solution Approach 2:
The system changes the parameter of the registered speaker set over time. By evaluating speech presence in consecutive segments and applying deletion conditions, the system transforms a static large speaker set into a dynamic reduced set, improving recognition accuracy while maintaining adaptability to actual speaking participants.
2Adaptability or versatility
If all participants are preregistered as target speakers, then no speaker is missed in identification, but computational complexity and processing time increase due to comparing against all registered speakers in every segment
Solution Approach 1:
The system extracts and removes unnecessary speakers from the registered speaker set. By applying deletion conditions to eliminate speakers who are not currently active, the system reduces the number of comparisons needed in each segment while maintaining coverage of all actual speakers, thereby reducing computational complexity.
Solution Approach 2:
The system temporarily discards speakers from the active registered set when they are not detected in consecutive segments, and can recover them when重新 detected. This dynamic discarding and recovering mechanism reduces unnecessary computational comparisons while ensuring all speakers are identified when they speak.
3Difficulty of detecting and measuring
If the threshold for speaker similarity is lowered to include more potential speakers, then the sensitivity of speaker detection is improved, but the accuracy of speaker recognition decreases due to increased misidentification
Solution Approach 1:
The system dynamically adjusts the effective threshold by changing the registered speaker set. Instead of using a low fixed threshold that causes misidentification, the system adaptively reduces the candidate pool to only those speakers currently active in consecutive segments, achieving high sensitivity without sacrificing accuracy.
Data Source
AI summary
The information processing method in the present disclosure is performed as below. At least one speech segment is detected from speech input to a speech input unit. A first feature quantity is extracted from each speech segment detected, the first feature quantity identifying a speaker whose voice is contained in the speech segment. The first feature quantity extracted is compared with each of second feature quantities stored in storage and identifying the respective voices of registered speakers who are target speakers in speaker recognition. The comparison is performed for each of consecutive speech segments, and under a predetermined condition, among the second feature quantities stored in the storage, at least one second feature quantity whose similarity with the first feature quantity is less than or equal to a threshold is deleted, thereby removing the at least one registered speaker identified by the at least one second feature quantity.