Method and device for optimizing speech recognition system

A speech recognition and optimization method technology, applied in speech recognition, speech analysis, instruments, etc., can solve the problems that optimization technology is difficult to analyze comprehensively and accurately locate, and achieve the effect of overcoming artificial subjective experience, facilitating promotion, and reducing optimization costs

CN105679309AActive Publication Date: 2016-06-15科大讯飞(北京)有限公司
7 Cites 4 Cited by

Patent Information

Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Publication Date
2016-06-15

Smart Images

  • Figure 1
    Figure 1
  • Figure 2
    Figure 2
  • Figure 3
    Figure 3
Patent Text Reader

Abstract

The invention relates to a method and a device for optimizing a speech recognition system, and belongs to the technical field of speech recognition. The method comprises the following steps: building a model which comprises a call layer, a navigation layer and an interaction layer, wherein the call layer and the navigation layer are correlated, and the navigation layer and the interaction layer are correlated; getting a fact table used for multi-dimension multi-index analysis through specific model processing operation according to the model; calculating the indexes evaluating the ability of the system through on-line analytical processing OLAP according to the fact table, and locating an abnormal factor pulling down the indexes; and removing the abnormal factor to optimize the system. According to the invention, a model is introduced, a machine is fully utilized to carry out automatic operation, and analysis can be made through an existing mature database analysis method. Thus, manpower is saved greatly, and the defects of artificial subjective experience, low efficiency of test under a complex system, incomplete coverage and ambiguous location are overcome.
Need to check novelty before this filing date? Find Prior Art

Description

technical field

[0001] The present invention relates to the technical field of speech recognition, and more specifically relates to an optimization method and device for a speech recognition system. Background technique

[0002] In speech recognition technology, the core measurement index is the recognition rate, which reflects the core competitiveness of the entire product and system. Usually, the recognition rate of a certain product has been tested before it is launched, so that the recognition rate can reach theoretically usable or ideal results. However, in the actual application process, due to the influence of the actual environment, individual products usually require The entire system is optimized, so that the recognition rate of products and systems can reach an ideal level.

[0003] At present, in the optimization process, the existing recognition rate is firstly counted, then the recognition rate and the information related to the recognition rate are analyzed, ...

Examples

Embodiment 1

[0046] see figure 1 , an embodiment of the present invention provides a method for optimizing a speech recognition system, including:

[0047] Step 110: Construct a model, which includes a call layer, a navigation layer and an interaction layer.

[0048] Wherein, the communication layer and the navigation layer are related to each other, and the navigation layer and the interaction layer are related to each other.

[0049] The model is mainly composed of three layers: call layer, navigation layer and interaction layer. The call number or session number is used to identify and associate layers or layers. Further, the call layer and the navigation layer are related to each other through the call number, and the navigation layer and the interaction layer are related to each other through the session number. The optimization method is a "multi-dimensional" and "multi-indicator" analysis method established on the basis of the model. Each dimension has downward inheritance, and ea...

Embodiment 2

[0059] The embodiment of the present invention provides a specific implementation method of the above step 120, wherein the fact table of the multi-dimensional multi-index analysis includes an interaction layer fact table, a navigation layer fact table, and a call layer fact table. The method may include the steps of:

[0060] Step 121: According to the interaction layer fact table information of the interaction layer, calculate and judge the interaction layer base table information to generate an interaction layer fact table, and the interaction layer base table information includes statistical information, tag information and dimension information .

[0061] For example, the fact table information of the interaction layer has the tag information of "whether to help", "it is true when the interaction recognition result is 'help', otherwise it is false", when the interaction recognition result is "help", the generated interaction The flag information of "whether to help" in t...

Embodiment 3

[0077] see image 3 ,

[0078] The embodiment of the present invention provides a specific implementation method of the above step 130, which may include the following steps:

[0079] Step 131: Analyze the indicators of the call layer to obtain indicators or sets of indicators that are not in the ideal state, and use OLAP, such as using OLAP technical means such as slicing, dicing, drilling, and rotation, to obtain abnormal indicators and abnormal dimensions of the call layer , the abnormal index of the communication layer is an indicator that the communication layer is not in an ideal state, and the abnormal dimension of the communication layer is a dimension that affects that the communication layer is not in an ideal state.

[0080] Step 132: Analyze the relevant indicators and dimensions of the navigation layer according to the abnormal indicators and dimensions of the communication layer, and use OLAP, such as using OLAP slice, slice, drill, rotate and other technical me...