Title:
SPEAKING PERSON SEPARATION METHOD AND APPARATUS BASED ON RECURRENT NEURAL NETWORK AND ACOUSTIC FEATURES
Document Type and Number:
WIPO Patent Application WO/2020/258661
Kind Code:
A1
Abstract:
A speaking person separation method based on recurrent neural network and acoustic features, comprising acquiring, by means of speech recognition, a word vector set of speech data to be recognized, recognizing and acquiring an MFCC feature vector set of the speech data to be recognized, and performing full connection on the sets, in order to obtain a combined feature vector (S120); encoding the combined feature vector to obtain an encoded result (S130); decoding the encoded result to obtain a split result corresponding to the combined feature vector (S140); performing prediction of speaking person changeover on the split result to obtain a speaking person recognition results corresponding to speaking person changeover symbols (S150); subjecting the speaking person recognition results to clustering to obtain speaking person classification results (S160); and sending the speaking person classification results to an upload terminal corresponding to the speech data to be recognized (S170).
Inventors:
WANG JIANZONG (CN)
JIA XUELI (CN)
JIA XUELI (CN)
Application Number:
PCT/CN2019/117805
Publication Date:
December 30, 2020
Filing Date:
November 13, 2019
Export Citation:
Assignee:
PING AN TECH SHENZHEN CO LTD (CN)
International Classes:
G10L21/0272; G10L17/06; G10L25/24; G10L25/30
Foreign References:
| CN110444223A | 2019-11-12 | |||
| CN106683661A | 2017-05-17 | |||
| CN108766440A | 2018-11-06 | |||
| CN109036454A | 2018-12-18 | |||
| CN105427858A | 2016-03-23 | |||
| CN108320732A | 2018-07-24 | |||
| CN109584903A | 2019-04-05 | |||
| US6895376B2 | 2005-05-17 | |||
| US20190156837A1 | 2019-05-23 |
Attorney, Agent or Firm:
SHENZHEN TALENT PATENT SERVICE (CN)
Download PDF: