Two-Stage Noise Suppression for Speech Recognition Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current speech recognition systems in intelligent virtual assistants and smart speakers face challenges with low signal-to-noise ratios and non-stationary noise, leading to errors in speech recognition, especially in noisy environments.
Innovation Solution
A two-stage noise suppression method that processes audio signals using a combination of a stationary noise suppression model and a neural network-based approach, where the first stage reduces stationary noise and the second stage addresses non-stationary noise using voice activity detection signals to optimize the noise suppression model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional speech recognition systems are used in noisy environments, then the system structure remains simple, but the speech recognition accuracy deteriorates due to low signal-to-noise ratio and non-stationary noise
Solution Approach 1:
The noise suppression process is divided into two distinct stages: stationary noise suppression using a noise model, and non-stationary noise suppression using neural networks. This segmentation allows each stage to specialize in handling specific noise types, improving overall speech recognition accuracy while maintaining a structured and manageable system architecture.
Solution Approach 2:
Voice activity detection signals serve as an intermediary mechanism that bridges the two noise suppression stages. These signals guide the neural networks in the second stage to selectively process non-stationary noise components, enabling coordinated operation between stages and improving speech recognition reliability without requiring a completely complex unified system.
2Adaptability or versatility
If a single-stage noise suppression method is used, then the system complexity remains low, but the ability to handle both stationary and non-stationary noise simultaneously deteriorates
Solution Approach 1:
The system segments noise handling into two specialized pathways: one for stationary noise using traditional noise modeling techniques, and another for non-stationary noise using neural networks. This segmentation enables the system to adapt to different noise types effectively while keeping each processing module relatively simple and focused.
Solution Approach 2:
The system dynamically switches between different noise suppression strategies based on the characteristics of the input signal. The voice activity detection mechanism enables the system to adapt its processing approach in real-time, handling stationary noise when appropriate and non-stationary noise when detected, thereby improving versatility without requiring all processing components to operate simultaneously.
3Reliability
If neural networks are used for all noise suppression tasks, then the handling of non-stationary noise improves, but the processing time and computational complexity increase
Solution Approach 1:
The system segments the noise suppression task so that only non-stationary noise components are processed by computationally intensive neural networks, while stationary noise is handled by faster traditional methods. This segmentation significantly reduces the overall processing time while maintaining high accuracy for the more challenging non-stationary noise components.
Solution Approach 2:
Instead of applying neural networks to all noise suppression tasks, the system applies them partially—only to the non-stationary noise components identified by voice activity detection. This partial application of the more complex method achieves the necessary accuracy improvement without incurring the full computational cost of applying neural networks universally.
Data Source
AI summary
A speech enhancement method and a speech enhancement system are provided. The speech enhancement method performs two-stage noise suppression by using digital signal processing and neural network approach. The first-stage noise suppression generates artifact signals by reducing stationary noise in the digital audio signals. The second-stage noise suppression performs voice activity detection and further reduces non-stationary noise in the artifact signals. The result of the voice activity detection is fed back to establish or update a noise model used in the first-stage noise suppression.