Pruned Acoustic Model for Low Resource Key Phrase Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current key phrase detection systems in low resource environments face challenges with robustness due to high complexity, leading to issues with noise and false accepts, making them unsuitable for widespread implementation in wake on voice systems.
Innovation Solution
A method using a pruned deep neural network-based acoustic model, a start state based rejection model, and a key phrase model, which generates a time series of scores for sub-phonetic units to determine the likelihood of a key phrase, reducing resource usage while maintaining detection accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex key phrase detection models (GMMs with context-dependent phones) are used, then detection accuracy is improved, but device complexity and resource consumption increase
Solution Approach 1:
The patent extracts and removes unnecessary components from the original complex GMM model, keeping only the essential elements needed for key phrase detection. This involves pruning the model to eliminate redundant context-dependent phone models while retaining the core detection functionality, thereby reducing complexity while maintaining accuracy.
Solution Approach 2:
The patent applies different levels of modeling complexity to different parts of the key phrase detection system. Instead of uniformly complex modeling across all phones, it uses simplified models for non-critical segments and more detailed models only where necessary for accurate detection, optimizing the balance between accuracy and resource usage.
2Use of energy by moving object
If simpler key phrase detection techniques are used, then resource consumption is reduced, but robustness deteriorates
Solution Approach 1:
The patent performs preliminary processing and filtering of audio input before applying the simplified detection model. By pre-processing the signal to enhance relevant features and suppress noise early in the pipeline, the simpler model can achieve robust detection performance without requiring complex post-processing or high computational resources.
Solution Approach 2:
The patent optimizes detection parameters such as threshold values, time windows, and feature extraction parameters to maximize the performance of simplified models. By carefully tuning these parameters, the system achieves robust detection with lower computational complexity and reduced power consumption.
3Use of energy by stationary object
If current low resource techniques are used, then power consumption is reduced, but false accepts increase
Solution Approach 1:
The patent introduces intermediary verification steps and additional features that act as mediators between the simplified detection model and the final decision. These intermediaries provide extra validation without requiring full complex model execution, thereby reducing false accepts while maintaining low power consumption.
Data Source
AI summary
Techniques related to key phrase detection for applications such as wake on voice are discussed. Such techniques may include updating a start state based rejection model and a key phrase model based on scores of sub-phonetic units from an acoustic model to generate a rejection likelihood score and a key phrase likelihood score and determining whether received audio input is associated with a predetermined key phrase based on the rejection likelihood score and the key phrase likelihood score.