# US 11,443,757 B2 · Artificial sound source separation method and device of thereof

> Agents: this data is available as JSON from the patentgrep API and MCP server — see https://patentgrep.com/llms.txt

- Number: US 11,443,757 B2
- Document ID: US11443757B2
- Status: Granted
- Kind: B2
- Application number: 16/593,155
- Filed: 2019-10-04
- Published: 2020-01-30
- Granted: 2022-09-13
- Earliest claimed: 2019-08-30
- Assignees: LG Electronics Inc.
- Applicants: LG Electronics Inc.
- Inventors: Hyeonsik Choi
- Web page: https://patentgrep.com/patent/US11443757B2

## CPC

- G10L 21/0272: Voice signal separating
- G06N 3/08: Learning methods
- G06N 20/00: Machine learning
- G06N 20/10: using kernel methods, e.g. support vector machines [SVM]
- H04L 65/80: Responding to QoS
- G06N 3/084: Backpropagation, e.g. using gradient descent
- G10L 25/30: using neural networks
- G10L: SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G06N: COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- H04L: TRANSMISSION OF DIGITAL INFORMATION, e.g. TELEGRAPHIC COMMUNICATION (arrangements common to telegraphic and telephonic communication H04M)

## Abstract

An artificial sound source separation method and device are disclosed. The sound source separation method by the artificial sound source separation device based on dictionary learning generates a dictionary matrix by performing dictionary learning, receives an overlapping sound source in which at least two sound sources are mixed and separates a target sound source from the overlapping sound source based on the dictionary matrix; and detecting the target sound source. The dictionary learning may be performed using a K-SVD algorithm. The intelligent computing device configuring a sound source processing device of the present disclosure may be associated with an artificial intelligence module, drone (unmanned aerial vehicle, UAV), robot, augmented reality (AR) devices, virtual reality (VR) devices, devices related to 5G services, and the like.

## Claims

### Claim 1

An artificial sound source separation method, the method comprising:
generating a dictionary matrix by performing dictionary learning;
receiving an overlapping sound source in which at least two sound sources are mixed;
separating a target sound source from the overlapping sound source based on the dictionary matrix; and
detecting the target sound source,
wherein the dictionary learning is performed using a K-singular value decomposition (K-SVD) algorithm,
wherein the target sound source is separated from the overlapping sound source through non-negative matrix factorization,
wherein the separated target sound source is represented by the dictionary matrix and a gain matrix,
wherein, based on a summed value of gains in the gain matrix being less than a specific threshold, the detection of the target sound source is determined to fail, and
wherein the failed target sound source is used for the dictionary learning.

### Claim 2 (depends on 1)

The method of claim 1, further comprising:
receiving a single sound source data to perform the dictionary learning,
wherein the single sound source data is transformed into a mel-scale.

### Claim 3 (depends on 1)

The method of claim 1, wherein the overlapping sound is transformed into a mel-scale.

### Claim 4 (depends on 1)

The method of claim 1, wherein the dictionary matrix includes a frequency component and the gain matrix includes a time component.

### Claim 5 (depends on 1)

The method of claim 1, wherein an update of the gain matrix is performed to minimize the difference between a vector of the overlapping sound source and the dictionary matrix and the gain matrix.

### Claim 6 (depends on 1)

The method of claim 1, further comprising:
updating the generated dictionary matrix.

### Claim 7

An artificial sound source separation device, comprising:
an input/output unit configured to transmit and receive data;
a memory configured to store the data;
a dictionary learning unit configured to perform dictionary learning; and
a processor configured to:
control the dictionary learning unit to generate a dictionary matrix by performing the dictionary learning using a K-singular value decomposition K-SVD algorithm,
control the input/output unit to receive an overlapping sound source in which at least two sound sources are mixed,
separate a target sound source from the overlapping sound source based on the dictionary matrix, and
detect the target sound source,
wherein the target sound source is separated from the overlapping sound source through non-negative matrix factorization,
wherein the separated target sound source is represented by the dictionary matrix and a gain matrix,
wherein, based on a summed value of gains in the gain matrix being less than a specific threshold, the detection of the target sound source is determined to fail, and
wherein the failed target sound source is used for the dictionary learning.

### Claim 8 (depends on 7)

The artificial sound source separation device of claim 7, wherein the dictionary matrix is stored on the memory.

### Claim 9 (depends on 7)

The artificial sound source separation device of claim 7, wherein the input/output unit transmits and receives the data through a wireless communication network.

### Claim 10 (depends on 9)

The artificial sound source separation device of claim 9, wherein the data is transmitted through a PUSCH and the data is transmitted through a PDSCH.

### Claim 11 (depends on 7)

The artificial sound source separation device of claim 7, wherein the dictionary learning unit performs the dictionary learning using deep neural networks (DNN).

### Claim 12 (depends on 7)

The artificial sound source separation device of claim 7, wherein the processor further configured to:
control the input/output unit to perform a random-access procedure to a base station, and
control the input/output unit to receive an UL grant from the base station.

### Claim 13 (depends on 12)

The artificial sound source separation device of claim 12, wherein the processor controls the input/output unit further to perform an uplink beam management (BM) procedure.