Natural Language Understanding Prediction Model for Vehicle Speech

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Traditional natural language understanding (NLU) technologies are time-consuming and cannot be efficiently parallelized with automatic speech recognition (ASR), and they struggle with accurate prediction for specific users due to their single-model concept.

Innovation Solution

A method that involves receiving a speech request, recognizing it frame by frame, using a prediction model trained on user language habits to predict the text during recognition, and comparing the recognition text with the prediction text to output a natural language understanding result, employing multiple machine learning models including deep learning for improved accuracy.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If traditional NLU approach is used with frame-by-frame ASR recognition and sequential processing, then recognition accuracy can be maintained, but NLU processing time increases and parallelization with ASR becomes difficult

Engineering Contradiction:
Improverecognition accuracyVSAvoidNLU processing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent applies preliminary action by performing NLU prediction on the ASR recognition result before the ASR processing is completely finished. The prediction model generates predicted text and intent during the ASR frame-by-frame recognition process, allowing NLU to be performed in advance rather than waiting for ASR completion. This enables parallelization and reduces total processing time while maintaining accuracy through subsequent verification.

Inventive Principle:
Principle #10Preliminary action

2Device complexity

If traditional single-model NLU approach is used, then system complexity is reduced, but prediction accuracy for specific users deteriorates

Engineering Contradiction:
Improvesystem complexityVSAvoidprediction accuracy
Core Design Contradiction:
Device complexityVSMeasurement precision

Solution Approach 1:

The patent applies dynamics by implementing a flexible model selection mechanism that can dynamically choose between a first prediction model (trained on general data) and a second prediction model (trained on user-specific data). The system determines which model to use based on whether user information is available, allowing the system to adapt its complexity based on the specific task requirements and user context, thereby optimizing both accuracy and complexity management.

Inventive Principle:
Principle #15Dynamics

3Ease of operation

If NLU is performed after ASR recognition completes with timeout mechanism, then processing simplicity is maintained, but response time increases

Engineering Contradiction:
Improveprocessing simplicityVSAvoidresponse time
Core Design Contradiction:
Ease of operationVSSpeed

Solution Approach 1:

The patent applies preliminary action by initiating NLU prediction during the ASR recognition process rather than waiting for completion. The prediction model processes the ASR result in parallel with ongoing ASR frame-by-frame recognition, generating predicted text and intent earlier than traditional sequential approaches. This maintains processing simplicity while significantly reducing response time through parallel execution.

Inventive Principle:
Principle #10Preliminary action

Data Source

PatentEP4086893B1Natural language understanding method and device, vehicle and medium
Publication Date: 2024.08.21 ZHAOQING XIAOPENG NEW ENERGY INVESTMENT CO LTD
  • EP4086893B1 patent drawingFigure 1
  • EP4086893B1 patent drawingFigure 2
  • EP4086893B1 patent drawingFigure 3

AI summary

A natural language understanding method is disclosed, which includes: receiving a speech request and recognizing the speech request frame by frame to obtain a frame text and a recognition text; performing prediction on the frame text using a prediction model to obtain a prediction text during the frame-by-frame recognition; processing the prediction text to obtain a first processing result; and comparing the recognition text with the prediction text after the frame-by-frame recognition is finished, and when the recognition text matches with the prediction text, outputting the first processing result as a natural language understanding result for the speech request. The speech request from a user can be recognized and subjected to natural language understanding, meanwhile, through prediction in advance and NLU, NLU can be completed in advance during timeout waiting. A natural language understanding device, a vehicle and a storage medium are also disclosed.