TWI752682B - Method for updating speech recognition system through air - Google Patents

Abstract

The present invention provides a method for updating speech recognition system through air. Client ASR servers connect with a central ASR cloud server through Internet. New edition of ASR system is stored in the central ASR cloud server for being selected and downloaded by the client ASR servers for using.

Description

Method for updating speech recognition system in cloud

The present invention relates to a method for updating a speech recognition system, in particular to a method for updating the speech recognition system via the cloud.

When a general cloud automatic speech recognition system (ASR, Automatic Speech Recognition) needs to be updated, professionals must bring a USB flash drive into a computer room that controls the cloud automatic speech recognition system to update, which is quite labor-intensive and time-consuming.

Since the cloud automatic speech recognition system is in the cloud, it is more convenient to update the automatic speech recognition system from the cloud. This technology is directly designed by the developer of the cloud automatic speech recognition system and used by customers. The developer puts the new version of the automatic speech recognition system on the cloud, and the customer's cloud automatic speech recognition system selects the new version of the automatic speech through the network. Identify the system for use.

The purpose of the present invention is to provide a method for updating the speech recognition system via the cloud, so that the client ASR server and the central ASR cloud server are connected via the network, and a new version of the automatic speech recognition system can be selected. The content of the method of the present invention is described below.

The client ASR server serves as a system for providing automatic speech recognition in the cloud, and a central ASR cloud server is set up to be connected to the client ASR server via the network.

The new version of the automatic speech recognition system is placed on the central ASR cloud server, and the customer ASR server selects the new version of the automatic speech recognition system for use through the network.

The steps of the automatic speech recognition system in the new version to analyze speech are in the order of audio preprocessing, extraction of speech feature parameters, acoustic model and language model, where the acoustic model and the language model are the main body of the cloud update.

1: Client ASR server

2: Client ASR server

3: Client ASR server

4: Central ASR cloud server

21: Audio preprocessing

22: Extract speech feature parameters

23: Acoustic Model

24: Language Models

31: Speech recognition executive program

32: Decide which version to use based on the profile description

41: Steps

42: Steps

43: Steps

44: Steps

45: Steps

46: Steps

47: Steps

48: Steps

49: Steps

50: Steps

A: version

B: version

C:version

FIG. 1 is an explanatory diagram of the basic structure of the present invention.

FIG. 2 is a schematic diagram of the steps of analyzing speech by the automatic speech recognition system of the present invention.

FIG. 3 is a flow chart of the version selection of the cloud automatic speech recognition system of the present invention.

FIG. 4 is a flow chart of the updated version of the automatic speech recognition system of the present invention via cloud communication.

Figure 1 illustrates the basic architecture of the present invention. The client ASR server 1, the client ASR server 2, and the client ASR server 3 are all systems that provide automatic speech recognition in the cloud, and are connected to the central ASR cloud server 4 of the present invention through the network. The central ASR cloud server 4 of the present invention is directly designed by the developer of the cloud automatic speech recognition system and used by the client ASR server 1, ASR server 2, and ASR server 3. The developer puts the new version of the automatic speech recognition system in the central The ASR cloud server 4 selects a new version of the automatic speech recognition system for use by its customer's cloud automatic speech recognition system via the Internet.

FIG. 2 illustrates the steps of the automatic speech recognition system for analyzing speech, the sequence is audio preprocessing 21 , extraction of speech feature parameters 22 , acoustic model 23 and language model 24 . Among them, the acoustic model 23 and the language model 24 are the main body of the cloud update, and developers focus on this to make the cloud update simple Light and fast.

Please refer to Figure 3 to illustrate the process of how to select the version of the system that provides cloud automatic speech recognition, such as customer ASR server 1, customer ASR server 2, customer ASR server 3, , , etc. The speech recognition system first performs a "speech recognition execution program" 31 , and then decides which version to use 32 according to its profile description. If its profile describes version A, it leads to the acoustic model and language model of version A. If it is describing version B, it leads to the acoustic model and language model of version B. If the cloud version needs to be updated in the future, leave a location for version C.

FIG. 4 illustrates the cloud communication update process between the client ASR server 1, the client ASR server 2, the client ASR server 3, , , etc. and the central ASR cloud server 4. Generally, the client ASR server will actively inquire about the new version on the central ASR cloud server 4 at, for example, two in the morning (step 41 ), and the central ASR cloud server 4 will reply to the new version (step 42 ). The client ASR server compares the version in its configuration file (step 43), and if it is the same as the new version, the cloud update will not be performed. If it is different from the new version, the client's ASR server will request the central ASR cloud server 4 to download the new version (step 44).

The central ASR cloud server 4 calculates the MD5 (Message Digest Algorithm) of the new version of the acoustic and language model that has been packaged into a ZIP (compressed file) (step 45), then downloads it to the client ASR server, and informs its MD5 value (step 46).

The client ASR server performs MD5 operation on the downloaded ZIP (step 47 ), and compares the MD5 value in the response message (step 48 ). Step 48 is to verify whether the downloaded ZIP archive is complete, and the same MD5 value indicates the completeness of the ZIP archive.

Finally, the client ASR server performs ZIP decompression (step 49), and points the description of the "configuration file" to the new version (step 50), and finally restarts the entire system, that is, the cloud update is completed.

The spirit and scope of the present invention are determined by the following patent application scope, and are not limited to the above-mentioned embodiments.

1: Client ASR server

2: Client ASR server

3: Client ASR server

4: Central ASR cloud server

Claims (2)

A method for updating a speech recognition system in the cloud, comprising the following steps: (1) setting at least one client ASR server as a system for providing automatic speech recognition in the cloud, and setting up a central ASR cloud server and the at least one client ASR server via a network (2) A new version of the automatic speech recognition system is placed on the central ASR cloud server, and the at least one client ASR server chooses to download the new version of the automatic speech recognition system via the network for use; (3) the At least one client ASR server actively inquires about the new version on the central ASR cloud server: (4) the central ASR cloud server replies to the new version; (5) the at least one client ASR server compares the version in one of its configuration files , if it is the same as the new version, the cloud update will not be performed; (6) if it is different from the new version, the at least one client ASR server requests the central ASR cloud server to download the new version to complete the update; (7) the central ASR cloud server The terminal packs the new version into a ZIP (compressed file), calculates an MD5 (Message Digest Algorithm) value, and then downloads it to the at least one client ASR server, and informs the MD5 value; (8) the at least one client The ASR server performs an MD5 operation on the downloaded ZIP, and compares and responds to the MD5 value, and the same value as the MD5 indicates the integrity of the ZIP file; (9) The at least one client ASR server performs a ZIP decompression, And point the description of a profile to the new version, and then restart the entire system to complete the cloud update. For the method for updating a speech recognition system in the cloud according to item 1 of the scope of the patent application, in the new version, the The steps of the automatic speech recognition system to analyze the speech are in the sequence of an audio preprocessing, an extraction of speech feature parameters, an acoustic model and a language model, wherein the acoustic model and the language model are the main bodies of the cloud update.

Images