Title:
CONVOLUTIONAL RECURRENT NEURAL NETWORK-BASED SINGLE-CHANNEL REAL-TIME NOISE REDUCTION METHOD
Document Type and Number:
WIPO Patent Application WO/2020/042707
Kind Code:
A1
Abstract:
A convolutional recurrent neural network-based single-channel real-time noise reduction method. Said method comprises: extracting acoustic features from a received single-channel sound signal (S110); performing an iterative operation on the acoustic features in a pre-trained convolutional recurrent neural network model, so as to calculate a ratio mask of the acoustic features (S120); masking the acoustic features by means of the ratio mask (S130); synthesizing the masked acoustic features and phases of the single-channel sound signal, so as to obtain a voice signal (S140). The method is able to decrease the number of neural network parameters, reduce the data storage amount and the demand for system data bandwidth, achieve good noise reduction performance, and improve the real-time performance of single-channel voice noise reduction.
Inventors:
TAN KE (CN)
YAN YONGJIE (CN)
YAN YONGJIE (CN)
Application Number:
PCT/CN2019/090530
Publication Date:
March 05, 2020
Filing Date:
June 10, 2019
Export Citation:
Assignee:
ELEVOC TECH CO LTD (CN)
International Classes:
G10L21/0208
Domestic Patent References:
| WO2013149123A1 | 2013-10-03 |
Foreign References:
| CN109841226A | 2019-06-04 | |||
| CN107452389A | 2017-12-08 | |||
| US20170061978A1 | 2017-03-02 | |||
| CN106847302A | 2017-06-13 |
Other References:
D. S. WANG: "A Deep Convolutional Encoder-Decoder Model for Robust Speech Dereverberation", INTERNATIONAL CONFERENCE ON DIGITAL SIGNAL PROCESSING (DSP, 25 August 2017 (2017-08-25), XP033246182, ISSN: 2165-3577
Attorney, Agent or Firm:
SHENZHEN KUAIMA PATENT & TRADEMARK OFFICE (CN)
Download PDF: