# US 11,908,487 B2 · Signal processing apparatus and non-transitory computer readable medium > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 11,908,487 B2 - Document ID: US11908487B2 - Status: Granted - Kind: B2 - Application number: 17/187,559 - Filed: 2021-02-26 - Published: 2022-03-17 - Granted: 2024-02-20 - Earliest claimed: 2020-09-16 - Assignees: Kabushiki Kaisha Toshiba - Applicants: Kabushiki Kaisha Toshiba - Inventors: Takehiko Kagoshima; Daichi Hayakawa - Web page: https://patentgrep.com/patent/US11908487B2 ## CPC - G10L 25/30: using neural networks - H04R 3/005: {for combining the signals of two or more microphones (specially adapted for hearing aids H04R25/407)} - G10L 21/0272: Voice signal separating - G10L 25/03: characterised by the type of extracted parameters - H04R 1/326: {for microphones (H04R1/34 and H04R1/40 take precedence)} - G10L 15/00: Speech recognition (G10L17/00 takes precedence) - H04R 1/406: {microphones} - G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING - H04R: LOUDSPEAKERS, MICROPHONES, GRAMOPHONE PICK-UPS OR LIKE ACOUSTIC ELECTROMECHANICAL TRANSDUCERS; ELECTRIC HEARING AIDS; PUBLIC ADDRESS SYSTEMS (producing sounds with frequency not determined by supply frequency G10K) ## Abstract A signal processing apparatus according an embodiment includes an acquisition unit and an application unit. The acquisition unit acquires M detection signals output from M detector devices having N-fold symmetry (M is an integer equal to or greater than 2, and N is an integer equal to or greater than 2). Each of the M detector devices detects original signals generated from K signal sources (K is an integer equal to or greater than 2) having the N-fold symmetry. The application unit applies a trained neural network to M input vectors corresponding to the M detection signals and outputs K output vectors. The same parameter is set to, of multiple weights included in a weight matrix of the trained neural network, weights that are commutative based on the N-fold symmetry. ## Claims ### Claim 1 A signal processing apparatus comprising: an acquisition unit configured to acquire M detection signals output from M detector devices (M is an integer equal to or greater than 2) having N-fold symmetry integer equal to or greater than 2), each of the M detector devices detecting an original signal generated from K signal sources (K is an integer equal to or greater than 2) having the N-fold symmetry; and an application unit configured to apply a trained neural network to M input vectors corresponding to the M detection signals and output K output vectors, wherein a same parameter is set to, of a plurality of weights included in a weight matrix of the trained neural network, weights that are commutative based on the N-fold symmetry. ### Claim 2 (depends on 1) The apparatus according to claim 1, wherein the detector devices are microphone devices, the detection signals are speech signals, and the output vectors are one of a speech spectrum, a time-frequency mask, an acoustic score, or a linguistic score. ### Claim 3 (depends on 2) The apparatus according to claim 2, wherein the input vectors are complex spectra, and the plurality of weights includes complex numbers. ### Claim 4 (depends on 1) The apparatus according to claim 1, wherein the detector devices are microphone devices, the detection signals are speech signals, and the application unit applies the trained neural network to the M input vectors and outputs, as the K output vectors, K speech spectra respectively corresponding to the K signal sources. ### Claim 5 (depends on 4) The apparatus according to claim 4, further comprising: a speech recognition unit configured to perform speech recognition on the K speech spectra and output K sets of text data relating to speech content. ### Claim 6 (depends on 1) The apparatus according to claim 1, wherein the detector devices are microphone devices, the detection signals are speech signals, and the application unit applies the trained neural network to the M input vectors and outputs, as the K output vectors, K acoustic scores respectively corresponding to the K signal sources. ### Claim 7 (depends on 6) The apparatus according to claim 6, further comprising: a speech recognition unit configured to perform speech recognition on the K acoustic scores and output K sets of text data relating to speech content. ### Claim 8 (depends on 4) The apparatus according to claim 4, further comprising: an A/D conversion unit configured to perform A/D conversion on the M detection signals to convert them into M digital signals; and a Fourier transform unit configured to perform short-term Fourier transform on the M digital signals to generate M frequency spectra as the M input vectors. ### Claim 9 (depends on 6) The apparatus according to claim 6, further comprising: an A/D conversion unit configured to perform A/D conversion on the M detection signals to convert them into M digital signals; and a Fourier transform unit configured to perform short-term Fourier transform on the M digital signals to generate M frequency spectra as the M input vectors. ### Claim 10 (depends on 1) The apparatus according to claim 1, wherein a first direction range in which, of K direction ranges respectively corresponding to the K signal sources when viewed from the M detector devices, two or more signal sources are present, original signals are produced with time intervals therebetween from the two or more signal sources included in the first direction range. ### Claim 11 (depends on 1) The apparatus according to claim 1, wherein the trained neural network includes: M input channels into which the M input vectors are input respectively, and K output channels which output the K output vectors respectively, and if the correspondence between the M input vectors and the M input channels is changed by rotating the M detector devices in such a manner that the M detector devices overlap according to the N-fold symmetry, the correspondence between the K output vectors and the K output channels is changed before and after the rotation, whereas the K output vectors remain unchanged before and after the rotation. ### Claim 12 A signal processing apparatus comprising: a signal processing unit configured to train a neural network into which M input vectors corresponding to M detector devices having N-fold symmetry are input (M is an integer equal to or greater than 2, and N is an integer equal to or greater than 2) and which outputs K output vectors corresponding to K signal sources having N-fold symmetry (K is an integer equal to or greater than 2); a determination unit configured to determine a combination of commutative weights, among a plurality of weights included in a weight matrix of the neural network, in terms of a symmetrical operation based on an order of rotational symmetry N based on a number of microphone devices M, a number of speakers K, and the order of rotational symmetry N; and a training unit configured to assign a same parameter to multiple weights belonging to a same weight combination and train the neural network, and generate a trained neural network. ### Claim 13 (depends on 12) The apparatus according to claim 12, wherein the training unit trains the neural network based on training data, the training data has training data of a plurality of speech patterns corresponding to combinations of the K signal sources, and of the plurality of speech patterns, speech patterns having the N-fold symmetry share a same speech pattern. ### Claim 14 A non-transitory computer readable medium storing instructions that cause a computer to: output K output vectors corresponding to K signal sources having N-fold symmetry based on an input of M input vectors corresponding to M detector devices having N-fold symmetry (each of M, N, and K is an integer equal to or greater than 2) applied to a neural network, wherein a same parameter is set to, of multiple weights included in a weight matrix of the neural network, weights that are commutative based on the N-fold symmetry.