Multi-Channel Voice Activity Detection With Compressed Signal Normalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing noise cancellation methods, such as Spectral Subtraction and Voice Activity Detection, fail to effectively remove noise in noisy environments, leading to high error rates in speech recognition systems due to non-linear distortion and sensitivity drift between acoustic channels.
Innovation Solution
A dual-stage noise reduction architecture combining multi-channel and single-channel noise cancellation, with adaptive finite impulse response filters and linear filtering to maintain signal fidelity and balance channel sensitivities, using multiple microphones to separate desired and undesired audio signals.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Object-affected harmful factors
If Spectral Subtraction is used to reduce noise, then noise reduction is achieved, but non-linear distortion is introduced resulting in high error rates in speech recognition
Solution Approach 1:
The patent replaces non-linear spectral subtraction methods with linear filtering approaches. The linear filter processes the acoustic signal without introducing non-linear distortion, thereby maintaining speech recognition accuracy while still achieving noise reduction through the linear transformation of the signal spectrum.
Solution Approach 2:
The patent modifies the filtering parameters dynamically based on the noise environment. By adjusting the linear filter coefficients according to the estimated noise characteristics, the system achieves effective noise reduction while preserving the integrity of the speech signal for accurate recognition.
2Measurement precision
If Voice Activity Detection is used, then desired voice detection is improved, but sensitivity drift between acoustic channels occurs leading to ineffective noise removal
Solution Approach 1:
The patent implements a feedback mechanism that continuously monitors the output of multiple acoustic channels and adjusts the linear filter coefficients accordingly. This feedback loop compensates for sensitivity drift between channels, ensuring that the noise reduction remains effective across varying operating conditions while maintaining accurate voice detection.
Solution Approach 2:
The system dynamically adapts the linear filter parameters based on real-time channel sensitivity measurements. By making the filter characteristics variable rather than fixed, the system compensates for drift in channel sensitivities and maintains optimal performance across different acoustic environments.
3Object-affected harmful factors
If multiple microphones are used to separate desired and undesired audio, then noise reduction capability is improved, but device complexity increases
Solution Approach 1:
The patent divides the noise reduction task into separate processing stages for each acoustic channel. Each channel is processed independently through its own linear filter, and the results are combined. This segmentation approach manages the complexity of multi-channel processing by breaking it down into manageable, independent sub-tasks that can be executed in parallel.
Data Source
AI summary
Systems and methods are described to create a desired voice activity detection signal. A main acoustic signal and a plurality of reference acoustic signals are compressed. The compressed main acoustic signal is normalized by the plurality of compressed reference acoustic signals to create a plurality of normalized compressed main acoustic signals. The plurality of normalized compressed main acoustic signals is processed with a plurality of single channel normalized voice threshold comparators to form a plurality of normalized desired voice activity detection signals. One of the plurality of normalized desired voice activity detection signals is selected from the plurality of normalized desired voice activity detection signals to output as the desired voice activity detection signal.