Machine-Learned Voiceprint Hashing for Negative-List Speaker Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing fraud detection systems in telecom and call centers suffer from high latency and computational intensity, particularly in large negative lists, making real-time fraudster identification challenging, and existing normalization techniques are not optimized for multi-target speaker detection.
Innovation Solution
A system utilizing Locality Sensitive Hashing (LSH) and a novel Negative List Normalization (NL-NORM) methodology to efficiently reduce similarity score computations and calibrate match scores, combined with machine learning models like GMM, DNN, and PLDA, to enhance fraud detection accuracy and reduce detection time.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional similarity score computation methods are used for negative list detection, then detection accuracy can be maintained, but detection time and computational intensity become excessively high
Solution Approach 1:
The patent segments the negative list into multiple cohorts or subsets, allowing the system to perform similarity computations against smaller groups rather than the entire list. This segmentation enables parallel processing and significantly reduces the computational burden and detection time while maintaining detection accuracy through comprehensive coverage of all cohorts.
Solution Approach 2:
The system performs preliminary actions by pre-computing and storing similarity scores or embeddings for the negative list cohorts before actual detection occurs. This preprocessing creates ready-to-use representations that can be quickly compared against probe utterances, eliminating the need for computationally intensive real-time computations during detection.
2Measurement precision
If comprehensive normalization techniques are applied to all speakers in the negative list, then score calibration accuracy improves, but computational complexity increases significantly
Solution Approach 1:
The normalization process is segmented by applying it separately to each cohort or subset of the negative list rather than to all speakers simultaneously. This cohort-based normalization reduces the computational complexity of each normalization operation while collectively providing comprehensive score calibration across the entire negative list through the aggregation of results from multiple normalized cohorts.
3Measurement precision
If the complete voiceprint set is used for matching against all negative list speakers, then recognition accuracy is maintained, but processing latency increases
Solution Approach 1:
The voiceprint set is processed in conjunction with segmented cohorts, where the system efficiently matches voiceprints against divided subsets of the negative list. This segmentation enables the processing to be distributed and parallelized, significantly improving processing speed while maintaining recognition accuracy through the comprehensive coverage of all segments.
Solution Approach 2:
The system extracts and utilizes only the essential voiceprint features necessary for accurate matching, removing redundant information from the complete voiceprint set. This extraction of critical features reduces the dimensionality and processing requirements while preserving the core information needed for accurate speaker recognition.
Data Source
AI summary
A system and method for identifying a speaker based on dual calculation of voiceprint match scores, using machine learned voiceprint sets and machine learned hashing models to reduce a voiceprint set, normalizing the joint scores, and performing user identification. The system and method include, among other things, a novel normalization technique designed for multitarget detection that considers scores between all enrolled negative list (NL) utterances and the normalization cohort as a single distribution. The system and method include machine learned hashing models to efficiently find a small subset of utterances from enrolled NL utterances and the normalization cohort that are most similar to the test utterance, so that the number of similarity score computations can be significantly reduced in identifying the speaker.