Speech noise reduction method and device based on artificial intelligence and computer device

Patent metadata

Field	Value
Publication number	US-10867618-B2
Application number	US-201715856180-A
Country	US
Kind code	B2
Filing date	Dec 28, 2017
Priority date	Apr 14, 2017
Publication date	Dec 15, 2020
Grant date	Dec 15, 2020

Abstract

Embodiments of the present disclosure provide a speech noise reduction method and a speech noise reduction device based on artificial intelligence and a computer device. The method includes the followings. A first noisy speech to be processed is received. The first noisy speech to be processed is pre-processed, to obtain the first noisy speech in a preset format. The first noisy speech in the preset format is sampled according to a sampling rate indicated by the preset format, to obtain first sampling point information of the first noisy speech. A noise reduction is performed on the first sampling point information through a deep-learning noise reduction model, to generate noise-reduced first sampling point information. A first clean speech is generated according to the noise-reduced first sampling point information.

First claim

What is claimed is:

1. A speech noise reduction method based on artificial intelligence, performed by one or more computing devices and comprising:
receiving a first noisy speech to be processed;
pre-processing the first noisy speech to be processed, to obtain the first noisy speech in a preset format;
sampling the first noisy speech in the preset format according to a sampling rate indicated by the preset format, to obtain first sampling point information of the first noisy speech;
obtaining the deep-learning noise reduction model after training;
using the first sampling point information of the first noisy speech as an input of the deep-learning noise reduction model and performing a noise reduction on the first sampling point information in a time domain through the deep-learning noise reduction model, to generate noise-reduced first sampling point information having a corresponding time point as an output of the deep-learning noise reduction model; and
generating a first clean speech according to the noise-reduced first sampling point information,
wherein the sampling rate indicates the number of the sampling points included in a speech with a period of 1 second,
wherein obtaining the deep-learning noise reduction model after the train comprises:
collecting a second clean speech and a second noisy speech corresponding to the second clean speech;
pre-processing the second clean speech and the second noisy speech corresponding to the second clean speech, to obtain the second clean speech in the preset format and the second noisy speech in the preset format;
dividing the second clean speech in the preset format and the second noisy speech in the preset format according to a preset length to obtain divided second clean speeches and divided second noisy speeches;
sampling the divided second clean speeches and the divided second noisy speeches with the sampling rate indicated by the preset format, to obtain second sampling point information of the second clean speech and third sampling point information of the second noisy speech; and
using the third sampling point information of the second noisy speech as an input of a deep-learning model, and using the second sampling point information of the second clean speech having a corresponding time point as an output of the deep-learning model, so as to train the deep-learning model, to obtain the deep-learning noise reduction model.

2. The method according to claim 1 , wherein collecting the second noisy speech corresponding to the second clean speech comprises:
adding noise to the second clean speech to collect the second noisy speech corresponding to the second clean speech.

3. The method according to claim 1 , wherein the preset format is a pulse code modulation format with a sampling rate of 16000 Hz, 16-bit quantization, and a single channel.

4. The method according to claim 1 , wherein the preset format is a pulse code modulation format with a sampling rate of 16000 Hz, 16-bit quantization, and a single channel.

5. The method according to claim 1 , wherein the preset format is a pulse code modulation format with a sampling rate of 16000 Hz, 16-bit quantization, and a single channel.

6. The method according to claim 2 , wherein the preset format is a pulse code modulation format with a sampling rate of 16000 Hz, 16-bit quantization, and a single channel.

7. A computer device, comprising:
one or more processors;
a storage device, configured to store one or more programs;
a receiver, configured to receive a first noisy speech to be processed;
wherein the one or more processors are configured to read the one or more programs stored in the storage device to perform acts of:
pre-processing the first noisy speech to be processed, to obtain the first noisy speech in a preset format;
sampling the first noisy speech in the preset format according to a sampling rate indicated by the preset format, to obtain first sampling point information of the first noisy speech;
obtaining the deep-learning noise reduction model after training;
using the first sampling point information of the first noisy speech as an input of the deep-learning noise reduction model and performing a noise reduction on the first sampling point information in a time domain through the deep-learning noise reduction model, to generate noise-reduced first sampling point information having a corresponding time point as an output of the deep-learning noise reduction model; and
generating a first clean speech according to the noise-reduced first sampling point information,
wherein the sampling rate indicates the number of the sampling points included in a speech with a period of 1 second,
wherein obtaining the deep-learning noise reduction model after the train comprises:
collecting a second clean speech and a second noisy speech corresponding to the second clean speech;
pre-processing the second clean speech and the second noisy speech corresponding to the second clean speech, to obtain the second clean speech in the preset format and the second noisy speech in the preset format;
dividing the second clean speech in the preset format and the second noisy speech in the preset format according to a preset length to obtain divided second clean speeches and divided second noisy speeches;
sampling the divided second clean speeches and the divided second noisy speeches with the sampling rate indicated by the preset format, to obtain second sampling point information of the second clean speech and third sampling point information of the second noisy speech; and
using the third sampling point information of the second noisy speech as an input of a deep-learning model, and using the second sampling point information of the second clean speech having a corresponding time point as an output of the deep-learning model, so as to train the deep-learning model, to obtain the deep-learning noise reduction model.

8. The computer device according to claim 7 , wherein collecting the second noisy speech corresponding to the second clean speech comprises:
adding noise to the second clean speech to collect the second noisy speech corresponding to the second clean speech.

9. The computer device according to claim 7 , wherein the preset format is a pulse code modulation format with a sampling rate of 16000 Hz, 16-bit quantization, and a single channel.

10. A non-transitory storage medium comprising computer executable instructions, wherein when the computer executable instructions are executed by a computer processor, a speech noise reduction method based on artificial intelligence is executed, and the method comprises:
receiving a first noisy speech to be processed;
pre-processing the first noisy speech to be processed, to obtain the first noisy speech in a preset format;
sampling the first noisy speech in the preset format according to a sampling rate indicated by the preset format, to obtain first sampling point information of the first noisy speech;
obtaining a deep-learning noise reduction model after training;
using the first sampling point information of the first noisy speech as an input of the deep-learning noise reduction model and performing a noise reduction on the first sampling point information in a time domain through the deep-learning noise reduction model, to generate noise-reduced first sampling point information having a corresponding time point as an output of the deep-learning noise reduction model; and
generating a first clean speech according to the noise-reduced first sampling point information,
wherein the sampling rate indicates the number of the sampling points included in a speech with a period of 1 second,
wherein obtaining the deep-learning noise reduction model aft

Assignees

• Baidu online network technology beijing co ltd

Inventors

• Zou Wei
• Li Xiangang
• Cui Weiwei
• Hu Jingyuan

Classifications

• G06N3/09

  Supervised learning · CPC title

• G10L21/0208 Primary

  Noise filtering · CPC title

• G06N20/00

  Machine learning · CPC title

• G10L21/0264

  characterised by the type of parameter measurement, e.g. correlation techniques, zero crossing techniques or predictive techniques · CPC title

• G10L21/02 Primary

  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) · CPC title