6 results about "Microphone array" patented technology

The invention discloses a microphone array-oriented channel attention weighted speech enhancement method. The method comprises the following steps: respectively acquiring a path of noisy speech time domain signal from each microphone in a microphone array; constructing a multi-channel speech enhancement model; taking a short-time Fourier transform spectrum of each path of noisy speech of the microphone array as an input feature of a multi-channel speech enhancement model, and pre-training the multi-channel speech enhancement model; and inputting multiple paths of noisy speech acquired by the microphone array into the multi-channel speech enhancement model, and realizing microphone array speech enhancement by using the multi-channel speech enhancement model. According to the method, a UNETnetwork is adopted to process the voice frequency spectrum features, the parameter quantity is small, the calculation complexity is low, attention mechanism weighting operation is adopted for the channel dimension on the basis, attention weighting is carried out on each layer of encoder and decoder in the channel dimension, the features are weighted in deep structure simulation beam forming, and the performance of speech enhancement is improved.

The invention discloses a GIS fault diagnosis method and system based on voiceprint imaging. The diagnosis system comprises an acquisition module, an imaging positioning module, a fault diagnosis module and a PC module. The diagnosis method comprises the following steps: positioning an abnormal sound signal through a sound imaging method; performing eMD adaptive decomposition on the acquired GIS sound signals; and calculating the one-dimensional spectral entropy of each IMF component as a feature element to be substituted into the SVM. Multiple fault types of the GIS can be diagnosed and discriminated at one time, and the invention is simple and rapid; a sound signal radiated by a fault source can be prevented from being covered by other interference sources; the acoustic feature acquisition effect of the microphone array in the expected direction can be improved to a certain extent, and subsequent fault diagnosis is prevented from being affected.

The embodiments of the invention relate to the technical field of data processing and particularly disclose a method, device and system for obtaining an airspace filter parameter matrix. The method comprises the steps that a first pointing matrix under first sub-band frequency is determined according to a pre-obtained first sub-band frequency matrix under the first sub-band frequency, the pre-obtained number of microphones, the pre-obtained position where each microphone is located, and the pre-obtained angle corresponding to the position where each microphone is located; numerical values of elements in a constraint matrix are determined according to the angle corresponding to the position where each microphone is located, wherein the number of the elements in the constraint matrix is thesame as that of the microphones; and the airspace filter parameter matrix under the first sub-band frequency is determined according to the constraint matrix and the first pointing matrix. In this way, it can be further ensured that a microphone array receives sound in the preset direction while small-interval arrangement of the microphone array is met, and voice enhancement is realized.

The invention provides an audition and vision collaborative humanoid robot head orientation method. The method is characterized by mainly comprising the following steps of firstly, acquiring a voice signal through a microphone array and recording time; recognizing by using a human voice recognition method, and counting the number of microphones for collecting human voice; setting a candidate target orientation according to time when the number of the voice microphones is greater than 1; then rotating a camera and acquiring a video; obtaining human face image frames from the video, obtaining a human face recognition result through recognition by means of a human face recognition method, and counting the number of human faces; when the number of the human faces is greater than 1, calculating human face areas and sorting the human face areas; setting candidate target human faces according to the ranking; then, achieving lip movement recognition through a lip movement detection algorithm; setting a target orientation according to the area when the number of the human faces with lip movement is greater than 1; and finally rotating the head of the humanoid robot to realize orientation. According to the method, the humanoid robot can distinguish the human voice from other voices and can accurately interact with the target human voice.