Speaker Verification ANN Sub-Phonetic Discrimination
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing speaker verification and identification systems require scoring against both speaker-specific and generic acoustic models, increasing processing time and complexity, especially on devices with limited compute resources, and face challenges in setting appropriate decision thresholds due to score variability.
Innovation Solution
A method using an artificial neural network (ANN) that discriminates between sub-phonetic classes uttered by a target speaker and other speakers by modifying an existing ANN to include additional output nodes for each sub-phonetic class, allowing the network to output probabilities for both the target and anti-speaker, thus eliminating the need for a separate generic acoustic model during the testing phase.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If speaker-specific and generic acoustic models are used for scoring, then verification accuracy is improved, but processing time increases
Solution Approach 1:
The patent combines the speaker-specific acoustic model and generic acoustic model into a single neural network model. The neural network is trained to simultaneously learn speaker-specific characteristics and general speech patterns, eliminating the need for separate scoring against multiple models and reducing processing time while maintaining verification accuracy.
Solution Approach 2:
The neural network is designed to perform multiple functions: it acts as both the speaker-specific model and the generic model simultaneously. By training the network with a combination of speaker-specific data and general speech data, it achieves multi-functionality, scoring speech samples against both model types through a single unified structure.
2Reliability
If speaker-specific and generic acoustic models are used for scoring, then verification accuracy is improved, but device complexity increases
Solution Approach 1:
The patent merges two separate acoustic modeling systems into a single neural network architecture. Instead of maintaining and managing separate speaker-specific and generic models, the system uses one unified neural network that incorporates both types of knowledge, thereby reducing system complexity while preserving the accuracy benefits of dual-model approaches.
3Reliability
If decision threshold tuning accounts for score variability from multiple models, then verification reliability is improved, but operational complexity increases
Solution Approach 1:
The patent changes the parameter representation by having the neural network output a single confidence score instead of multiple scores from separate models. This parameter transformation simplifies the decision threshold tuning process, as only one threshold needs to be established rather than managing threshold relationships between multiple model scores, while the network internally handles the variability through its training on diverse data.
Data Source
AI summary
In one embodiment, a computer system stores speech data for a plurality of speakers, where the speech data includes a plurality of feature vectors and, for each feature vector, an associated sub-phonetic class. The computer system then builds, based on the speech data, an artificial neural network (ANN) for modeling speech of a target speaker in the plurality of speakers, where the ANN is configured to discriminate between instances of sub-phonetic classes uttered by the target speaker and instances of sub-phonetic classes uttered by other speakers in the plurality of speakers.