ASR Evaluation Dataset Segmentation for Speaker Adaptation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing ASR engines perform poorly in specific environments and for certain groups of speakers due to lack of adaptation to individual characteristics such as age, gender, ethnicity, and recording conditions, as standard datasets fail to account for these variations.

Innovation Solution

A computer-implemented method that accesses ASR test datasets, samples them based on user-defined attributes like speaker characteristics and recording conditions, and generates additional samples if necessary to create a tailored dataset for evaluating and optimizing ASR engine configurations.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If standard datasets are used for ASR evaluation, then the evaluation process is simple and fast, but the evaluation results do not reflect specific speaker characteristics and recording conditions

Engineering Contradiction:
Improveevaluation accuracyVSAvoiddataset configuration complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments the evaluation dataset into multiple subsets based on speaker characteristics (age, gender, ethnicity) and recording conditions. Each subset is evaluated separately to provide specialized performance metrics for different user groups and environments, rather than using a single monolithic dataset.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies local quality by creating tailored evaluation datasets with specific characteristics (e.g., elderly speakers, female speakers, noisy environments) to evaluate ASR performance in localized contexts. This allows the system to optimize for specific local requirements rather than general performance.

Inventive Principle:
Principle #3Local quality

2Adaptability or versatility

If ASR engines are trained on diverse datasets, then the engine performs well across different speaker groups, but the training time and computational resources increase

Engineering Contradiction:
Improvespeaker group adaptabilityVSAvoidtraining time
Core Design Contradiction:
Adaptability or versatilityVSLoss of time

Solution Approach 1:

The patent performs preliminary evaluation using pre-existing diverse datasets to identify performance gaps before final training. This preliminary action allows the system to focus training resources on specific underperforming speaker groups rather than retraining on all data uniformly.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent changes training parameters dynamically based on evaluation results. When specific speaker groups are identified as underperforming, the system adjusts training parameters (learning rate, data weighting, model architecture) to prioritize training on those specific groups, optimizing the balance between diversity and training time.

Inventive Principle:
Principle #35Parameter changes

3Reliability

If the evaluation dataset is filtered by specific attributes, then the evaluation reflects specific user scenarios, but the dataset size decreases

Engineering Contradiction:
Improveevaluation relevanceVSAvoiddataset size
Core Design Contradiction:
ReliabilityVSQuantity of substance

Solution Approach 1:

The patent segments the evaluation process into multiple stages: first evaluating on broadly defined subsets (e.g., by age group or gender), then drilling down into more specific sub-groups. This segmentation allows maintaining sufficient dataset size at each stage while progressively increasing evaluation relevance to specific scenarios.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies partial action by filtering datasets for specific attributes (age, gender, ethnicity, recording conditions) to create targeted evaluation sets. While this reduces the overall dataset size, it increases the relevance and reliability of evaluation results for the specific user scenarios being tested.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP4287178A1Improved performance evaluation of automatic speech recognition engines
Publication Date: 2023.12.06 VIVOKA
  • EP4287178A1 patent drawingFigure 1
  • EP4287178A1 patent drawingFigure 2
  • EP4287178A1 patent drawingFigure 3

AI summary

The invention relates to a computer-implemented method comprising the steps of: accessing one or more Automatic Speech Recognition (ASR) test dataset, each ASR test dataset comprising one or more ASR test sample, each ASR test sample comprising: an audio sample, or its representation; a reference textual transcription of the audio sample; at least one attribute; receiving from a user at least one value of said at least one attribute; sampling said one or more ASR test dataset into a sampled ASR test dataset according to said at least one value of said at least one attribute; calculating a value of a performance metric for at least one ASR engine configuration using said sampled ASR test dataset.