Intent Determination Using Acoustic and Textual Feature Vectors
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional intelligent personal assistant systems struggle to accurately determine the intent behind spoken user utterances, particularly when utterances have textually identical representations but differ in intent, leading to inappropriate responses.
Innovation Solution
A method and system that analyze both textual and acoustic properties of spoken user utterances using speech-to-text analysis, generating textual and acoustic feature vectors, and employing a neural network to determine the intent, combining these features to provide intent-specific responses.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If conventional IPA systems analyze only textual data from digital audio signals, then the system complexity remains low, but the intent determination accuracy deteriorates when utterances are textually identical but have different intents
Solution Approach 1:
The patent merges textual features and acoustic features into a unified feature set for intent determination. The system combines text-to-speech analysis results with acoustic property analysis results to form enhanced feature vectors that are fed into the neural network, allowing the system to differentiate between utterances that are textually identical but have different acoustic characteristics and intents.
Solution Approach 2:
The patent adds an acoustic dimension to the traditional text-only analysis by extracting and analyzing acoustic properties such as pitch, tone, and speech patterns. This additional dimension allows the system to distinguish between different intents that would be indistinguishable through text analysis alone, thereby improving intent determination accuracy without significantly increasing system complexity.
2Measurement precision
If the system analyzes both textual and acoustic properties using neural networks, then the intent determination accuracy improves, but the processing time and computational resources increase
Solution Approach 1:
The patent performs preliminary analysis by separately processing textual data through text-to-speech analysis and acoustic data through acoustic property extraction before combining them. This preliminary separation allows for optimized processing of each feature type independently, and the results are then integrated into enhanced feature vectors that are efficiently processed by the neural network, reducing overall processing time.
Solution Approach 2:
The patent segments the intent determination process into distinct modules: text-to-speech analysis, acoustic property extraction, feature vector generation, and neural network classification. This segmentation allows each module to be optimized independently and processed in parallel where possible, thereby reducing the total processing time while maintaining high accuracy.
3Ease of operation
If the system provides intent-specific responses based on acoustic analysis, then the user interaction quality improves, but the device complexity increases
Solution Approach 1:
The patent implements a universal intent determination system that handles both text-based and acoustic-based intent recognition through a single neural network architecture. This multi-functional approach allows the system to process various types of utterances (questions, statements, commands) with different intents using the same underlying framework, thereby improving user interaction quality without proportionally increasing device complexity.
Data Source
AI summary
A method and server for determining an intent associated with a spoken user utterance are disclosed. The method includes determining at least one speech unit where each speech unit has textual data representative of a word or pause and has a corresponding segment of the digital audio signal. For each speech unit, the method includes generating a respective textual feature, generating a respective acoustic feature vector and generating a respective enhanced feature vector. The method also includes employing a neural network (NN) configured to determine the intent of the spoken user utterance by inputting into the NN the enhanced feature vectors. The NN has been trained to estimate a probability of the intent being of a given type.