ASR-NLU Integration for Speech Recognition Latency

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing speech-based conversation interaction frameworks, relying on separate Automatic Speech Recognition (ASR) and Natural Language Understanding (NLU) models, face errors that are amplified and incur high computational costs, leading to delayed recognition processes.

Innovation Solution

A method that integrates ASR and NLU models by converting speech into text, obtaining intention and slot information, and using a conversation state machine to generate interaction information, thereby reducing processing time and computational complexity.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If separate ASR and NLU models are used for speech recognition, then the system can perform semantic analysis, but the recognition process is delayed and computational costs increase

Engineering Contradiction:
Improverecognition accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The patent combines ASR and NLU models into an integrated speech recognition system. The NLU model directly processes acoustic features from the ASR model without requiring separate text output, enabling end-to-end speech-to-intention recognition and reducing processing delays while maintaining accuracy.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The system performs preliminary semantic analysis by integrating the NLU model to process ASR outputs in real-time. This preliminary processing of semantic information during the recognition pipeline reduces subsequent computational overhead and accelerates the overall recognition process.

Inventive Principle:
Principle #10Preliminary action

2Measurement precision

If a large model is used for speech recognition, then recognition accuracy is improved, but calculation amount increases and system cost rises

Engineering Contradiction:
Improverecognition accuracyVSAvoidcomputational cost
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

By merging ASR and NLU models into a unified architecture, the system eliminates redundant processing steps. The integrated model shares computational resources and processing pipelines, reducing overall calculation requirements while maintaining the accuracy benefits of large models.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent extracts and processes only the essential semantic information needed for speech recognition through the integrated NLU model. This selective extraction of critical features reduces unnecessary computational overhead while preserving recognition accuracy.

Inventive Principle:
Principle #2Taking out (Extraction)

3Measurement precision

If separate ASR and NLU models are used, then comprehensive semantic analysis is achieved, but error propagation occurs and system complexity increases

Engineering Contradiction:
Improvesemantic analysis accuracyVSAvoidsystem complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The integration of ASR and NLU models into a unified system reduces the number of independent components and interfaces. This consolidation simplifies the system architecture while maintaining comprehensive semantic analysis capabilities, reducing error propagation points between separate models.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentUS11322153B2Conversation interaction method, apparatus and computer readable storage medium
Publication Date: 2022.05.03 BAIDU ONLINE NETWORK TECH (BEIJIBG) CO LTD
  • US11322153B2 patent drawing
  • US11322153B2 patent drawing
  • US11322153B2 patent drawing

AI summary

A conversation interaction method and apparatus, and a computer-readable storage medium are provided. The method includes: converting a speech to be recognized into a first text; inputting the first text into a semantic analysis model, to obtain intention information and slot information of the first text; and inputting the intention information and the slot information of the first text into a conversation state machine, to obtain interaction information corresponding to the first text. By using a semantic analysis model, intention information and slot information of a first text are obtained directly from the first text. The process in the existing technology, where a semantic analysis model needs to be used immediately after a language model, is avoided, thereby shortening processing time and making it possible to respond faster to a user. Further, by using the above scheme, calculation complexity and the cost of a whole system are reduced.