Voice recognition result error correction method

A technology of speech recognition and error correction method, which is applied in speech recognition, speech analysis, natural language data processing, etc., and can solve problems that affect the promotion of speech recognition technology, lack of recognition error correction means, and high error rate of speech recognition.

Inactive Publication Date: 2018-02-23

上海百芝龙网络科技有限公司

5 Cites 28 Cited by

AI-Extracted Technical Summary

Problems solved by technology

Especially in everyday spoken language scenarios, the error rate of speech recognition is high

In the existing technology, all focus on improving the accur...

Abstract

The invention discloses a voice recognition result error correction method which includes the steps: preprocessing voice recognition results; finding out error-prone words and sentences in the voice recognition results or semantically parsing important words and characters to be corrected for a text; performing phonetic notation on the words and characters to be corrected in a complete spelling mode and a phonetic initial spelling mode to obtain unpitched pinyin corresponding to the recognition results of voice to be corrected; determining an optimal candidate text and a sub-optimal candidatetext by the aid of an editing distance algorithm according to the complete spelling mode; determining an optimal candidate text and a sub-optimal candidate text by the aid of the editing distance algorithm according to phonetic initials; combining all the optimal candidate texts and all the sub-optimal candidate texts, and only reserving one repeated candidate item; replacing a text to be corrected by a prospective candidate text, calculating the probability of replaced statements by an n-grama language model and selecting the statement with the highest probability as a final recognition result of the voice to be corrected.

Application Domain

Natural language data processingSpeech recognition +1

Technology Topic

Correction methodLanguage model +5

Image

Examples

Experimental program(1)

Example Embodiment

[0017] See figure 1 , The method of this embodiment includes:

[0018] S11: Perform text operations such as word segmentation, part-of-speech tagging, stop words removal and grammatical analysis on the speech recognition results

[0019] S12: According to existing or future technologies, find out the words and characters to be corrected that are prone to errors or important for text semantic analysis. Pay special attention to the verb-object structured phrases, verbs, nouns and words that do not appear in the dictionary database in the speech recognition results.

[0020] S13: Perform phonetic notation on the words and characters to be corrected, and obtain the corresponding pinyin of the voice recognition result to be corrected, and the corresponding pinyin means no tone.

[0021] This situation is divided into multiple situations, which are elaborated as follows:

[0022] Homophones, take full spelling:

[0023] For example, the result of the speech recognition to be corrected is "look at three tones and three is", and the corresponding pinyin after word segmentation is: kan sansheng san shi

[0024] The pronunciation is not standard, take the first letter of each word:

[0025] For example, the speech recognition result to be corrected is "Looking at the mountains, mountains and mountains is", the corresponding pinyin after the word segmentation is: kan shanshan shan shi, for which only the first letter of each word can be taken k s s s s

[0026] S14: First, determine the optimal candidate text and the sub-optimal candidate text according to the Pinyin Quanpin, using an edit distance confirmation algorithm;

[0027] S15: Secondly, according to the first letter of the pinyin, the edit distance algorithm is used again to determine the optimal candidate text and the suboptimal candidate text.

[0028] S16: Combine all the best candidate texts and sub-optimal candidate texts, and keep only one duplicate candidate, and all are collectively called quasi-candidate texts.

[0029] S17: Replace the quasi-candidate texts with the text to be corrected, use the n-grama language model to calculate the probabilities of each sentence after replacement, and select the highest probability as the final speech recognition result to be corrected

[0030] It is worth noting that although the foregoing content has described the spirit and principle of the present invention with reference to several specific embodiments, it should be understood that the present invention is not limited to the disclosed specific embodiments, and the division of various aspects does not imply these. The features in the aspect cannot be combined. This division is only for the convenience of presentation. The present invention is intended to cover various modifications and equivalent arrangements included within the spirit and scope of the appended claims.

PUM

Description & Claims & Application Information

We can also present the details of the Description, Claims and Application information to help users get a comprehensive understanding of the technical details of the patent, such as background art, summary of invention, brief description of drawings, description of embodiments, and other original content. On the other hand, users can also determine the specific scope of protection of the technology through the list of claims; as well as understand the changes in the life cycle of the technology with the presentation of the patent timeline. Login to view more.