Speech Preprocessing Model Training for Noise-Robust Recognition

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing speech recognition models suffer from reduced recognition rates due to noise, even when clean voices are processed, as the preprocessing of speech signals is independent and not optimized for the recognition models.

Innovation Solution

A preprocessing model building system and method that includes a training data preparation device, recognition rate prediction model, and speech preprocessing model, which use AI to learn and update speech features, generating clean speech data and optimizing the preprocessing model to enhance recognition rates.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If a well-known speech recognition model is used with independent preprocessing, then the system structure remains simple and easy to implement, but the speech recognition rate decreases because the preprocessing is not optimized for the recognition model

Engineering Contradiction:

Improvespeech recognition rateVSAvoidpreprocessing model structure

Core Design Contradiction:

ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges the preprocessing model and speech recognition model into a unified end-to-end system. The preprocessing operations (noise suppression, echo cancellation, etc.) are directly integrated with the recognition model, allowing the preprocessing to be optimized specifically for the recognition task rather than being a separate generic stage.

Inventive Principle:

Principle #5Merging (Combining)

Solution Approach 2:

The patent changes the parameters of the preprocessing operations by learning optimal preprocessing parameters through training data. Instead of using fixed preprocessing parameters, the system adapts preprocessing parameters based on the specific characteristics of the speech data and recognition model, thereby improving recognition rate.

Inventive Principle:

Principle #35Parameter changes

2Reliability

If preprocessing is performed independently without optimization for the recognition model, then the implementation is simpler, but the recognition rate suffers because clean voice does not guarantee high recognition rates

Engineering Contradiction:

Improverecognition rate for clean voiceVSAvoidpreprocessing implementation complexity

Core Design Contradiction:

ReliabilityVSEase of manufacture

Solution Approach 1:

The patent performs preliminary optimization of preprocessing operations during the training phase. By training the preprocessing model in advance using labeled speech data and recognition model feedback, the system prepares optimized preprocessing parameters before actual speech recognition, ensuring high recognition rates for clean voice without complex runtime adjustments.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The patent introduces a feedback mechanism where the speech recognition model's output is used to evaluate and adjust the preprocessing operations. The recognition rate serves as feedback to optimize preprocessing parameters, creating a closed-loop system that continuously improves recognition performance based on actual recognition outcomes.

Inventive Principle:

Principle #23Feedback

Data Source

PatentUS12525221B2Preprocessing model building system for speech recognition function and preprocessing model building method therefor

Publication Date: 2026.01.13 HYUNDAI MOTOR CO LTD

AI summary

In an embodiment a system includes a training data preparation device configured to obtain a speech recognition rate of speech data for training using a target speech recognition model, a recognition rate prediction model configured to estimate an expected recognition rate of the target speech recognition model for clean speech data in which noise is removed from the speech data for training and a speech preprocessing model configured to preprocess the speech data for training to obtain the clean speech data and to update the speech preprocessing model based on a recognition rate loss corresponding to a difference between the expected recognition rate and a maximum recognition rate.