Multi-Turn Voice Interaction Context Scoring for Vehicle Systems
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current voice interaction systems in vehicles struggle to accurately process multi-turn voice commands, often leading to over-rejection of normal instructions and poor user experience due to limited semantic rejection capabilities and blunt context inheritance, which results in inappropriate responses or failures to recognize user intentions.
Innovation Solution
A voice interaction method that calculates a context score and Natural Language Understanding (NLU) score for each turn of voice based on previous turns, using a combination of template and model context scores, and NLU processing to determine appropriate reply or rejection results, thereby improving the vehicle's ability to understand user commands and reduce over-rejection.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If the vehicle operating system uses a simple semantic rejection system to process voice commands, then the device complexity is reduced, but the measurement precision of understanding user intentions deteriorates, leading to over-rejection of normal instructions
Solution Approach 1:
The semantic rejection system is divided into multiple independent modules: a template matching module for predefined patterns, a context inheritance module for multi-turn dialogue state tracking, and a neural network module for semantic scoring. Each module handles specific aspects of voice command processing, allowing the system to maintain low complexity while achieving high precision through coordinated module operations.
Solution Approach 2:
The system transitions from single-turn independent processing to multi-turn contextual processing by introducing a context dimension. Context variables are inherited and updated across dialogue turns, enabling the system to distinguish between abnormal instructions requiring rejection and normal instructions requiring execution based on contextual understanding rather than isolated pattern matching.
2Speed
If the system processes each voice command independently without considering context, then the processing speed is improved, but the reliability of instruction execution deteriorates due to inability to distinguish normal from abnormal instructions in multi-turn dialogue
Solution Approach 1:
Context variables are pre-initialized and continuously updated during multi-turn dialogue before the final judgment is made. The system maintains a context state that accumulates information from previous turns, allowing rapid processing of current commands while relying on pre-established contextual understanding to ensure reliable distinction between normal and abnormal instructions.
3Device complexity
If the system uses blunt context inheritance to handle multi-turn dialogue, then the device complexity is reduced, but the measurement precision of context understanding deteriorates, resulting in inappropriate responses
Solution Approach 1:
The context inheritance mechanism is made dynamic by introducing confidence scores that vary with context relevance. Instead of uniformly inheriting all context variables, the system dynamically adjusts the weight and applicability of inherited context based on the current dialogue state and relevance to the current command, improving context understanding precision without requiring overly complex mechanisms.
Data Source
Figure 1~2
Figure 3~4
Figure 5
AI summary
The present disclosure provides a voice interaction method, a server and a voice interaction system. The voice interaction method is applied in the server. The voice interaction method includes: receiving multi-turn voice that is input in order of time and forwarded by a vehicle operating system, the multi-turn voice including a current turn of voice and previous turns of voice; calculating a context score and an NLU score of the current turn of voice in accordance with the multi-turn voice; obtaining a reply result or a rejection result of the current turn of voice in accordance with the current turn of voice and the context score and the NLU score of the current turn of voice; and performing a voice interaction between the vehicle operating system and a user in accordance with the reply result or the rejection result of the current turn of voice. In the voice interaction method, the vehicle operating system can be controlled to process the user's voice in an appropriate way, thereby reducing problems such as over-rejection and improving user experience.