Native Language Identification Using Separate Neural Network Models

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Current methods for native language identification from non-native language speech are limited in accurately distinguishing between different native languages, particularly in second-language acquisition and forensic applications, due to variations in pronunciation, grammar, and vocabulary usage.

Innovation Solution

A system utilizing a plurality of artificial neural networks, including time delay deep neural networks, to extract sufficient statistics and total variability matrices from native and non-native language corpora, generating i-vectors and a multilayer perceptron model for identifying a person's native language based on their non-native language utterance.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Device complexity

If a single universal background model is used for native and non-native language speakers, then the system complexity is reduced, but the measurement precision of native language identification deteriorates

Engineering Contradiction:

Improvesystem complexityVSAvoidnative language identification accuracy

Core Design Contradiction:

Device complexityVSMeasurement precision

Solution Approach 1:

The patent divides the single UBM into two separate UBMs: one trained on native English speaker data and another on non-native English speaker data. This segmentation allows each model to specialize in capturing the specific acoustic characteristics of its training population, thereby improving NLI accuracy without requiring a completely separate system for each speaker type.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by training different UBMs with different characteristics tailored to specific speaker groups. The native speaker UBM captures native pronunciation patterns while the non-native speaker UBM captures accent and pronunciation variations, allowing the system to adapt its reference model to the specific local characteristics of the speaker being analyzed.

Inventive Principle:

Principle #3Local quality

2Measurement precision

If separate UBMs are trained for native and non-native language speakers, then the measurement precision of native language identification is improved, but the device complexity increases

Engineering Contradiction:

Improvenative language identification accuracyVSAvoidsystem complexity

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the single UBM into two specialized models, each trained on distinct corpora (native and non-native speakers). This segmentation enables precise capture of language-specific patterns while maintaining a unified system architecture that processes both speaker types through the same i-vector extraction and classification pipeline.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The patent achieves universality by creating a multi-functional system where two UBMs serve different purposes within the same overall framework. Both models feed into the same i-vector extraction process and classification system, allowing the architecture to handle both native and non-native speakers uniformly while maintaining specialized accuracy for each group.

Inventive Principle:

Principle #6Universality (Multi-functionality)

3Manufacturing precision

If sufficient statistics are extracted using multiple artificial neural networks, then the manufacturing precision of i-vector extraction is improved, but the loss of time increases

Engineering Contradiction:

Improvei-vector extraction precisionVSAvoidprocessing time

Core Design Contradiction:

Manufacturing precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by pre-training both the native and non-native UBMs before actual NLI tasks. This preliminary training allows the models to capture and store the statistical patterns of their respective speaker groups in advance, so that during actual processing, the models can quickly extract i-vectors without requiring real-time computation of complex statistical relationships.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent replaces manual or computationally intensive statistical analysis with automated neural network-based extraction. The UBMs automatically learn and encode the statistical patterns of native and non-native speech, substituting complex manual feature engineering and statistical computation with trained neural networks that can rapidly extract sufficient statistics in a more efficient manner.

Inventive Principle:

Principle #28Mechanics substitution (Replace mechanical system)

Data Source

PatentUS10783873B1Native language identification with time delay deep neural networks trained separately on native and non-native english corpora

Publication Date: 2020.09.22 EDUCATIONAL TESTING SERVICE

AI summary

Systems and methods for identifying a person's native language, are presented. A native language identification system, comprising a plurality of artificial neural networks, such as time delay deep neural networks, is provided. Respective artificial neural networks of the plurality of artificial neural networks are trained as universal background models, using separate native language and non-native language corpora. The artificial neural networks may be used to perform voice activity detection and to extract sufficient statistics from the respective language corpora. The artificial neural networks may use the sufficient statistics to estimate respective T-matrices, which may in turn be used to extract respective i-vectors. The artificial neural networks may use i-vectors to generate a multilayer perceptron model, which may be used to identify a person's native language, based on an utterance by the person in his or her non-native language.