# US 9,437,208 B2 · General sound decomposition models > Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt - Number: US 9,437,208 B2 - Document ID: US9437208B2 - Status: Granted - Kind: B2 - Application number: 13/908,904 - Filed: 2013-06-03 - Published: 2014-12-04 - Granted: 2016-09-06 - Earliest claimed: 2013-06-03 - Assignees: Adobe Systems Incorporated - Applicants: Adobe Systems Incorporated - Inventors: Dennis L. Sun; Gautham J. Mysore - Web page: https://patentgrep.com/patent/US9437208B2 ## CPC - G10L 25/51: for comparison or discrimination - G10L 15/00: Speech recognition (G10L17/00 takes precedence) - G10L 21/0208: Noise filtering - G10L 25/27: characterised by the analysis technique - G06F 18/2133: based on naturality criteria, e.g. with non-negative factorisation or negative correlation - G10L 25/81: for discriminating voice from music - G10L 25/84: for discriminating voice from noise - G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING - G06F: ELECTRIC DIGITAL DATA PROCESSING (computer systems based on specific computational models G06N) ## Abstract Sound decomposition models are described. In one or more implementations, a plurality of individual models is generated for respective ones of a plurality of sound sources. The plurality of models is collected to form a universal audio model that is configured to support sound decomposition of sound data through use of one or more of the models. The plurality of models is not generated using a sound source that originated at least a portion of the sound data. ## Claims ### Claim 1 A method implemented by one or more computing devices, the method comprising: receiving a sound signal originated by one or more sound sources; fitting a universal audio model having a collection of individual models of sound sources to sound data associated with the sound signal, the fitting performed by selecting from the models in the collection based on correspondence of the sound data to respective said models, wherein the selecting is performed based on a penalty that encourages the correspondence and use of fewer of models in instances of strong correspondence between a subset of the models and the sound data as opposed to instances of weak correspondence between the subset of the models and the sound data; and decomposing the sound data using the selected models of the universal audio model. ### Claim 2 (depends on 1) A method as described in claim 1, wherein the universal audio model does not include a model trained on a specific sound source that originated at least a portion of the sound data. ### Claim 3 (depends on 1) A method as described in claim 1, wherein the decomposing is performed to also include use of training data learned on a specific sound source that originated at least a portion of the sound data. ### Claim 4 (depends on 1) A method as described in claim 1, wherein the selecting is performed based on a penalty that encourages block sparsity. ### Claim 5 (depends on 1) A method as described in claim 1, wherein each of the models is learned based on non-negative matrix factorization (NMF) or latent component analysis. ### Claim 6 (depends on 1) A method as described in claim 1, wherein the decomposing is performed based on non-negative matrix factorization (NMF) or latent component analysis. ### Claim 7 (depends on 1) A method as described in claim 1, wherein the decomposing is performed to support audio denoising, music transcription music remixing, or audio-based forensics. ### Claim 8 A method implemented by one or more computing devices, the method comprising: receiving sound signals originated from a plurality of sound sources; generating a plurality of individual models for respective ones of the plurality of sound sources, the generating comprising determining frequency representations from a plurality of examples of each of the plurality of sound sources; and collecting the plurality of models to form a universal audio model that is configured to support sound decomposition of a separate instance of sound data through use of one or more of the models, the plurality of models not being generated using a sound source that originated at least a portion of the separate instance of sound data, the universal audio model further configured to utilize generalized training data stored in the universal audio model in the absence of specific training data stored in the universal audio model for the sound decomposition of the separate instance of sound data. ### Claim 9 (depends on 8) A method as described in claim 8, further comprising decomposing the sound data using selected models of the universal audio model. ### Claim 10 (depends on 9) A method as described in claim 9, wherein the decomposing is performed to also include use of training data learned on a specific sound source that originated at least a portion of the sound data. ### Claim 11 (depends on 9) A method as described in claim 9, wherein the decomposing is performed based on non-negative matrix factorization (NMF) or latent component analysis. ### Claim 12 (depends on 8) A method as described in claim 8, wherein the universal audio model is configured to support selection of particular ones of the plurality of models based on a penalty that encourages the correspondence and use of fewer of models in instances of strong correspondence between a subset of the models and the sound data as opposed to instances of weak correspondence between the subset of the models and the sound data. ### Claim 13 (depends on 12) A method as described in claim 12, wherein the selecting is performed based on a penalty that encourages block sparsity. ### Claim 14 (depends on 8) A method as described in claim 8, wherein each of the models is learned based on non-negative matrix factorization (NMF) or latent component analysis. ### Claim 15 A system comprising: at least one module implemented at least partially in hardware and configured to collect a plurality of models to form a universal audio model that is configured to support sound decomposition of sound data through use of one or more of the models; and one or more modules implemented at least partially in hardware and configured to fit the universal audio model to the sound data to decompose the sound data by selecting from the models in the collection using a penalty term that encourages block sparsity. ### Claim 16 (depends on 15) A system as described in claim 15, wherein the universal audio model does not include a model trained on a specific sound source that originated at least a portion of the sound data. ### Claim 17 (depends on 15) A system as described in claim 15, wherein one or more modules are configured to perform the decomposing to also include use of training data learned on a specific sound source that originated at least a portion of the sound data. ### Claim 18 (depends on 15) A system as described in claim 15, wherein the penalty encourages use of fewer of models in instances of strong correspondence between a subset of the models and the sound data as opposed to instances of weak correspondence between the subset of the models and the sound data. ### Claim 19 (depends on 15) A system as described in claim 15, wherein each of the models is learned based on non-negative matrix factorization (NMF) or latent component analysis. ### Claim 20 (depends on 15) A system as described in claim 15, wherein one or more modules are configured to estimate a likely contribution of each source to portions of the sound data based at least in part on the selected models.