Microphone Array Speech Separation Using Key Phrase Masking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Speech recognition systems face challenges in accurately isolating desired speech signals from background noise and interfering speech in noisy environments with unknown source directions, especially in reverberant settings where sound arrives from multiple directions, making source separation difficult.
Innovation Solution
A system utilizing a microphone array and signal separation filter that employs key phrase recognition to infer acoustic mixing parameters, generating a time-frequency mask to separate desired speech signals from noise and interference, without requiring knowledge of microphone configuration, speaker location, or room acoustics, by processing signals through a method involving spectral analysis, template matching, and dynamic time warping.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If signals from multiple microphones are combined to enhance speech from a given direction, then noise and interference from other directions are reduced, but the system requires knowledge of source direction which is often unknown
Solution Approach 1:
The system performs self-calibration by automatically determining acoustic mixing parameters from the recorded signals themselves, without requiring external configuration information about microphone positions or room acoustics. The calibration process uses the speech signals and noise patterns present in the environment to infer the mixing matrix and its pseudoinverse, enabling the system to adapt to its specific acoustic environment autonomously.
2Reliability
If traditional source separation methods are used in reverberant environments, then separation performance degrades because sound arrives from multiple directions, but the patent achieves effective separation without requiring knowledge of room acoustics or speaker location
Solution Approach 1:
The system transforms the separation problem from the time domain to the frequency domain using short-time Fourier transform, where it can operate on spectral components. By working in the frequency domain and using pseudoinverse operations on the mixing matrix, the system can effectively separate sources even in reverberant conditions where traditional time-domain methods fail.
Solution Approach 2:
The system automatically adapts to different acoustic environments by calibrating itself using the actual signals present in that environment. It determines the acoustic mixing parameters specific to each room and configuration without requiring manual input about speaker locations, microphone positions, or room acoustics, making it universally adaptable to various settings.
3Measurement precision
If the system processes signals through spectral analysis and template matching to separate speech from noise, then speech recognition accuracy improves, but processing time and computational complexity increase
Solution Approach 1:
The system divides the speech signal into short-time spectral components using short-time Fourier transform, processing each frequency bin independently. This segmentation allows parallel processing of multiple frequency components and enables efficient application of the separation filter to each component, reducing overall processing time while maintaining accuracy.
Solution Approach 2:
The system applies separation processing selectively based on detected speech activity. It uses energy-based voice activity detection to identify time frames containing speech, and only applies the computationally intensive separation operations during these periods, rather than continuously processing all input signals, thus reducing unnecessary computational overhead.
Data Source
AI summary
An apparatus for enhancing a speech signal includes a spectral analysis circuit that generates time-frequency spectra of signals from a microphone array, a signal separation filter, and a hardware processor. The hardware processor identifies a key phrase in the signals and aligns a template of the key phrase to time-frequency spectra of a microphone in the array. A mask is generated from the aligned template and applied to the time-frequency spectra. The masked spectra are used to determine acoustic parameters that, in turn, are used to generate coefficients for configuring the source separation filter which is then used to process the spectra to generate an enhanced speech signal. The apparatus maybe used as a front end for a speech recognition engine, for example.