Method for reranking speech recognition results

Patent metadata

Field	Value
Publication number	US-9390710-B2
Application number	US-201514604991-A
Country	US
Kind code	B2
Filing date	Jan 26, 2015
Priority date	Oct 14, 2014
Publication date	Jul 12, 2016
Grant date	Jul 12, 2016

Abstract

Provided is a speech recognition method using machine learning, including: receiving a speech signal as an input, performing speech recognition to generate speech recognition result information including multiple candidate sentences and ranks of the respective candidate sentences; processing the multiple candidate sentences included in the speech recognition result information according to a machine learning model which is learned in advance and changing the ranks of the multiple candidate sentences to re-rank the multiple candidate sentences; and selecting the highest-rank candidate sentence among the re-ranked multiple candidate sentences as a speech recognition result. Particularly, the machine learning model is generated by: receiving the speech signal and a correct answer sentence as inputs; generating the speech recognition result information and a correct answer set; generating learning data by using the correct answer set; and performing the machine learning of changing the ranks of the candidate sentences.

First claim

What is claimed is:

1. A speech recognition method using machine learning, comprising:
receiving a speech signal as an input, performing speech recognition to generate speech recognition result information including multiple candidate sentences and ranks of the respective candidate sentences;
processing the multiple candidate sentences included in the speech recognition result information according to a machine learning model which is learned in advance and changing the ranks of the multiple candidate sentences to re-rank the multiple candidate sentences; and
selecting a highest-rank candidate sentence among the re-ranked multiple candidate sentences as a speech recognition result,
wherein the machine learning model is generated by:
receiving the speech signal and a correct answer sentence as inputs;
performing the speech recognition on the speech signal to generate the speech recognition result information including the multiple candidate sentences and sentence scores representing the ranks of the respective candidate sentences;
adding the correct answer sentence to the speech recognition result information to generate a correct answer set;
extracting features of the candidate sentences and the correct answer sentence included in the correct answer set to generate learning data; and
performing the machine learning of changing the ranks of the candidate sentences according to differences between the features of the candidate sentences and the features of the correct answer sentence based on the learning data, and
wherein the features include speech recognition ranks, a sentence score of the highest-rank candidate sentence, a morpheme bigram, a POS (part of speech) bigram, the number of domain dictionary unregistered words, morphemes/POSs of domain dictionary unregistered words, the number of general dictionary unregistered words, and morphemes/POSs of general dictionary unregistered words.

2. The speech recognition method according to claim 1 , wherein the machine learning is a Rank SVM.

3. The speech recognition method according to claim 1 ,
wherein the correct answer set includes a portion of the multiple candidate sentences, and
wherein the portion of the multiple candidate sentences are candidate sentences of which sentence scores are equal to or higher than a predetermined sentence score among the candidate sentences included in the speech recognition result information.

4. The speech recognition method according to claim 1 , wherein the speech recognition result information is transmitted from a predetermined external speech recognition service server.

Assignees

• Univ Sogang Res Foundation
• Kangwon Nat University University Industry Cooperation Foundation

Inventors

• Seo Jung Yun
• Kang Sang Woo
• Koo Myoung-Wan
• Kim Hark Soo
• Ahn Hyeok Ju
• Song Yeong Kil
• Choi Maeng Sik

Classifications

• G10L15/08 Primary

  Speech classification or search · CPC title

• G10L15/063 Primary

  Training · CPC title

• G10L2015/0635

  updating or merging of old and new templates; Mean values; Weighting · CPC title

• G10L25/27

  characterised by the analysis technique · CPC title