Abstract:
본 발명에 따른 강한 음성 인식 시스템은 전처리(pre-processing) 과정인 MPDR 빔포머를 사용하여 음원을 향상시킨 후, 향상된 음원 신호들과 노이즈 신호들의 합성신호에 대하여 HIVA 학습 알고리즘을 적용하여 음원 신호에 대한 특징 벡터를 추출하는 것을 특징으로 한다. 상기 음성 인식 시스템은 신호 왜곡을 최소화시키고 언믹싱 매트릭스에 대한 컨버전스를 향상시키기 위하여, HIVA 학습 알고리즘을 수행함에 있어서, non-holonomic constraint와 최소 왜곡 원칙(Minimal Distortion Priciple; 이하 'MDP'라 한다)을 적용하는 것을 특징으로 한다. 또한, 상기 음성 인식 시스템은 향상된 음원과 노이즈 음원을 이용하여 학습 과정에서 손실된 특징들(Missing Features)을 파악하고 이를 보상하는 것을 특징으로 한다. 전술한 특징들에 의하여, 본 발명에 따른 강한 음성 인식 시스템은 하모닉 주파수 의존성을 이용한 독립 벡터 분석 알고리즘을 기반으로 하여 노이즈 등에 강한 시스템을 제공하게 된다.
Abstract:
PURPOSE: A method for removing an interested sound source and a method for recognizing a voice are provided to estimate a mixed noise signal effectively, by removing the interested sound source from a mixed signal provided through two microphones installed in an acoustic signal mixed environment. CONSTITUTION: A voice recognition device initializes a vector(12). The device learns the vector to remove an interested sound source from a mixed signal(10). The device is initialized, and generates a mixed noise signal by removing an interested sound source signal from an input mixed signal(14). When the mixed noise signal is generated, the device generates a mask, by comparing the mixed noise signal with the input mixed signal in a time-frequency domain(16). [Reference numerals] (10) Learn a separation vector(w(K)) for separation of a target sound source; (12) Initialize the separation vector(w(K)); (14) Remove a target sound source; (16) Generate a mask using a mixed noise sound source and an input sound source
Abstract:
본 발명에 따르는 관심음원 제거방법은, 두 개의 마이크 각각으로부터의 입력 혼합신호를 제공받아 단구간 푸리에 변환하여 시간-주파수 영역으로 변환하는 단계; 상기 시간-주파수 영역의 입력 혼합신호들로부터 관심음원을 제거하기 위한 관심음원제거 벡터를 설정하는 단계; 상기 관심음원제거 벡터를 이용하여 입력 혼합신호에서 관심음원을 제거하여 혼합된 잡음신호를 생성하는 단계;를 구비함을 특징으로 한다.
Abstract:
A robust speech recognition system according to the present invention improves a sound source by using an MPDR beamformer in a pre-processing process, applies an HIVA learning algorithm to the composed signals of the improved sound source signals and noise signals, and extracts a feature vector of the sound source signals. The speech recognition system applies a non-holonomic constraint and a minimal distortion principle when performing the HIVA learning algorithm to minimize signal distortion and improve convergence of a non-mixing matrix. In addition, the speech recognition system checks for missing features in the learning process by using an improved sound source and a noise sound source and compensates for the same. By the aforementioned features, the robust speech recognition system provides a system resistant to noise on the basis of an independent vector analysis algorithm using harmonic frequency dependency. [Reference numerals] (200) Signal input unit;(210) Signal converting unit;(220) Pre-processing unit;(230) Sound source extracting unit;(246) Mask generating unit;(248) Loss property compensation output unit;(250) DCT converting unit;(260) Voice recognition unit;(AA,BB) Log unit
Abstract:
PURPOSE: A method for separating a blind source according to independent vector analysis by using a feed forward network and a device thereof are provided to resolve a problem according to independency between frequencies without heuristic technology. CONSTITUTION: An ST(Short-Time) Fourier transformer(100) converts mixed signals of a TD(Time-Domain) into mixed signals of an FD(Frequency-Domain). An FF unmixing filter network(104) separates the mixed signals of the FD into source signals. An inverse ST Fourier transformer(105) converts the separated source signals into source signals of the TD. An MPDR beam-former(102) receives the mixed signals of the FD from the ST Fourier transformer. The MPDR beam-former generates predetermined mixed signals of the FD. The MPDR beam-former provides the generated mixed signals of the FD to the FF unmixing filter network. [Reference numerals] (100) ST Fourier transformer; (102) MPDR beamformer; (104) FF unmixing filter network; (105) Reverse ST Fourier transformer