Abstract:
본 발명은 음성인식 시스템에서의 화자 적응 기술에 관한 것으로서, 본 발명에 따른 진화 학습에 의한 화자 적응 장치는, 음성인식 시스템이 음성인식을 수행하고 있는 인식 모드에서, 미리 결정된 환경 파라미터를 이용하여 인식 대상 음성데이터의 특징벡터에 대해 특징 변환을 수행하는 특징 변환부; 상기 인식 대상 음성데이터를 저장하는 음성 데이터베이스; 및 상기 음성인식 시스템이 음성인식을 수행하지 않고 있는 대기 모드에서, 상기 음성 데이터베이스에 저장된 음성데이터를 이용하여 기존 음향모델에 대한 모델 변환을 수행하는 모델 변환부를 포함하는 것을 특징으로 하여, 사용자 편의성 및 음성인식 시스템의 인식 성능을 개선함은 물론, 화자 적응 및 환경 적응을 동시에 수행하는 이점을 제공한다.
Abstract:
PURPOSE: A wideband active circuit with a feedback structure is provided to remarkably improve the bandwidth by forming a feedback structure which feeds back a signal of an output terminal to an active load. CONSTITUTION: An active load unit(100) provides load varying according to control voltage, and an active circuit unit(200) is connected between the active load unit and the ground. The active circuit unit outputs a signal corresponding to a preset bandwidth among the input signals. A feedback circuit unit(300) is formed between an output terminal of the active circuit unit and the active load unit, and supplies the signal of the output terminal of the active circuit unit to the active load unit. The active load unit is formed between the power voltage terminal and an output terminal in cascode structure.
Abstract:
PURPOSE: A voice recognition system using a speaker adaptive apparatus by properly performing model conversion is provided to apply an unsupervised adaptation type for the adaption of the voice recognition system. CONSTITUTION: A feature converting unit(102) converts a feature for a feature vector of recognition target voice data through a predetermined environmental parameter in a voice recognition mode. A voice database(106) stores recognizing subject audio data. A model converting unit(108) converts the model about the existing voice model through the stored voice data.
Abstract:
PURPOSE: A voice model adaptive method, an apparatus thereof using ML-LST and a voice recognition method using a noise voice model are provided to reduce calculation cost by applying a closed-form algorithm in a process presuming a noise parameter. CONSTITUTION: A linear spectrum feature extractor(420) extracts a linear spectrum data about a feature vector. A cepstrum feature extractor(430) converts linear spectrum data into cepstrum data. An occupation probability calculation unit(440) presumes Gaussian occupation probability by using a clean voice model. A noises parameter estimation unit(460) presumes the noises parameter by using the clean noise model. A noise voice model generator(470) uses the noise parameter.
Abstract:
A method for exploring a ground condition ahead of the tunnel face is provided to obtain the ground condition quickly and easily by using a transducer as an acoustic wave source. A rotation movement apparatus includes a source unit(30), a receiver unit(40), a rotating unit(50), and an external case. The source unit generates an acoustic wave and converts an electrical pulse signal into the acoustic wave. The receiver unit receives the acoustic wave and converts the acoustic wave into the electrical signal. The source unit and the receiver unit are connected to a mounting member through coupling members, respectively. The rotating unit includes a worm, a worm wheel, a support mount, and an implementation member. The worm wheel is rotated by the worm. The support mount supports the worm and the worm wheel. The implementation member is mounted on the worm wheel. The source unit, the receiver unit, and the rotating unit are contained in the external case.