Speaker Verification Accuracy via Transcript-Adapted GMM Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Text-independent speaker verification systems face challenges in accuracy due to the need to model a wide range of phonemes and contexts, leading to incorrect verification results when the feature space of the test utterance is not equally well-modeled by the universal background model and the speaker model.
Innovation Solution
Adapting the data structures of the universal background model (UBM-GMM) and the enrolled speaker Gaussian mixture model (ENR-GMM) responsive to a textual transcript of the spoken words, generating adapted models that improve verification accuracy by comparing enrolled and universal speaker probabilities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If text-independent speaker verification is used to allow free speech without restricting words, then ease of operation is improved, but measurement precision deteriorates due to high probability assignments to feature subspaces not present in test utterances
Solution Approach 1:
The patent segments the GMM by selecting only those Gaussian components that correspond to phonemes actually present in the test utterance (based on ASR transcript). This segmentation reduces the model from considering all possible phonemes to only relevant ones, thereby improving verification accuracy while maintaining text-independent operation.
Solution Approach 2:
The patent performs preliminary action by using automatic speech recognition to identify the phoneme sequence before speaker verification. This preliminary phoneme identification allows the system to pre-select relevant GMM components, avoiding the problem of high probability assignments to irrelevant feature subspaces.
2Adaptability or versatility
If a context-independent model is used to model speakers for a large variety of phonemes and contexts, then adaptability is improved, but measurement precision deteriorates due to unequal modeling of feature space
Solution Approach 1:
The patent introduces dynamics by making the GMM context-dependent rather than static. The model dynamically adapts to the specific phoneme sequence in each test utterance by selecting relevant components, allowing the same base model to accurately handle different phonetic contexts without sacrificing precision.
Solution Approach 2:
The patent changes the effective parameters of the GMM by selecting subsets of Gaussian components based on phoneme recognition results. This parameter selection approach allows the model to adapt its characteristics to match the specific utterance being verified, improving measurement precision while maintaining versatility.
Data Source
AI summary
A sampled speech data sequence contains words spoken by a speaker. A sequence of feature vectors is generated characterizing spectral distribution of sampled speech data. A textual transcript of the words spoken by the speaker is obtained. Data structures of a universal background model of a Gaussian mixture model (UBM-GMM) and of an Enrolled speaker Gaussian mixture model (ENR-GMM) are adapted responsive to the textual transcript, to generate an adapted UBM-GMM and an adapted ENR-GMM, respectively. An enrolled speaker probability is generated based on the sequence of feature vectors and the adapted ENR-GMM, and a universal speaker probability is generated based on the sequence of feature vectors and the adapted UBM-GMM. A speaker verification indication of whether the speaker is an enrolled speaker is generated by comparing the enrolled speaker probability to the universal speaker probability.