# US 11,308,970 B2 · Voice correction apparatus and voice correction method

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

- Number: US 11,308,970 B2
- Document ID: US11308970B2
- Status: Granted
- Kind: B2
- Application number: 16/674,803
- Filed: 2019-11-05
- Published: 2020-06-18
- Granted: 2022-04-19
- Earliest claimed: 2018-12-14
- Assignees: Fujitsu Limited
- Applicants: Fujitsu Limited
- Inventors: Naoshi Matsuo
- Web page: https://patentgrep.com/patent/US11308970B2

## CPC

- G10L 21/02: Speech enhancement, e.g. noise reduction or echo cancellation (reducing echo effects in line transmission systems H04B3/20; echo suppression in hands-free telephones H04M9/08)
- G10L 21/0208: Noise filtering
- G06N 3/0499: Feedforward networks
- G06N 3/09: Supervised learning
- G06N 20/00: Machine learning
- G10L 21/003: Changing voice quality, e.g. pitch or formants
- G10L 21/0232: Processing in the frequency domain
- G10L 25/30: using neural networks
- G06N 3/048: Activation functions
- G06N 3/08: Learning methods
- G10L 2021/02165: {Two microphones, one receiving mainly the noise signal and the other one mainly the speech signal}
- G10L 2021/02166: {Microphone arrays; Beamforming}
- 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

## Abstract

A voice correction method implemented by a computer, the method includes: obtaining first voice information which is voice information recorded when noise is generated and on which noise suppression processing is performed and second voice information indicating voice information recorded in an environment in which no noise is generated, and generating emphasized information by emphasizing a component of a band corresponding to a band having a low signal noise ratio (SNR) of the first voice information, among bands of the second voice information; performing machine learning on a model based on the first voice information and the emphasized information; and generating corrected voice information by correcting third voice information on which noise suppression processing is performed, based on the machine-learned model.

## Claims

### Claim 1

A voice correction apparatus comprising:
a memory; and
a processor coupled to the memory, the processor being configured to:
execute an emphasis processing configured to
obtain first voice information which is voice information recorded when noise is generated and on which noise suppression processing is performed to yield a noise suppressed first voice information and second voice information indicating voice information recorded in an environment in which no noise is generated, and
generate emphasized information by emphasizing a component of a band, corresponding to a band having a low signal noise ratio (SNR) of the noise suppressed first voice information, among bands of the second voice information;
execute a learning processing configured to perform machine learning on a model based on the noise suppressed first voice information and the emphasized information; and
execute a correction processing configured to generate corrected voice information by correcting third voice information on which noise suppression processing is performed, based on the model machine-learned by the learning processing.

### Claim 2 (depends on 1)

The voice correction apparatus according to claim 1,
wherein the emphasis processing is configured to specify a band with a component having a value less than a threshold value as a band to be emphasized among bands of the second voice information and multiples a component of a band corresponding to the band to be emphasized among the bands of the second voice information by a gain exceeding 1 to generate the emphasized information.

### Claim 3 (depends on 2)

The voice correction apparatus according to claim 2,
wherein the processor is further configured to:
execute an inverse emphasis processing configured to divide a component of a band, corresponding to the band to be emphasized, among bands of the corrected voice information corrected by the correction processing by the gain; and
execute a voice recognition processing configured to perform voice recognition based on information obtained by dividing the component of the band corresponding to the band to be emphasized among the bands of the corrected voice information by the gain.

### Claim 4 (depends on 1)

The voice correction apparatus according to claim 1,
wherein the emphasis processing is configured to subtract the noise suppressed first voice information from the second voice information to generate the emphasized information.

### Claim 5 (depends on 4)

The voice correction apparatus according to claim 4,
wherein the processor is further configured to:
execute an inverse emphasis processing configured to add the third voice information to the corrected voice information corrected by the correction processing; and
execute a voice recognition processing configured to perform voice recognition based on information obtained by adding the third voice information to the corrected voice information.

### Claim 6

A voice correction method implemented by a computer, the method comprising:
obtaining first voice information which is voice information recorded when noise is generated and on which noise suppression processing is performed to yield a noise suppressed first voice information and second voice information indicating voice information recorded in an environment in which no noise is generated;
generating emphasized information by emphasizing a component of a band, corresponding to a band having a low signal noise ratio (SNR) of the noise suppressed first voice information, among bands of the second voice information;
performing machine learning on a model based on the noise suppressed first voice information and the emphasized information; and
generating corrected voice information by correcting third voice information on which noise suppression processing is performed, based on the machine-learned model.