JP2003323191A - Access system to internet homepage adaptive to voice - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a client device for recognizing a voice inputted from a client device having a voice input function and realizing homepage reading with a voice and an access system to an Internet homepage adaptive to voice. <P>SOLUTION: This access system to an Internet homepage adaptive to voice has, in Internet homepage reading for accessing a homepage from the client device having a voice input function by voice, a sound analyzing means for receiving a voice input, performing a sound analysis such as voice encoding, extracting a characteristic parameter and transmitting the extracted characteristic parameter to a homepage providing site (1), a voice recognition means in the provider site for the homepage adaptive to voice for receiving the characteristic parameter, using a sound model, etc., to perform voice recognition, converting the recognized voice into a symbol string and transmitting the symbol string to the client device (2), and a receiving means with which the client device receives the homepage and displays the homepage on a screen or the like (3). <P>COPYRIGHT: (C)2004,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention recognizes voice input in a natural language from a client device which can be connected to the Internet and has a voice input function, and realizes homepage browsing by voice. Corresponding Internet home page access system.

[0002]

2. Description of the Related Art The latest computer-related technologies relating to speech processing are roughly classified into speech coding technology, speech synthesis technology, speech recognition technology, and speaker recognition technology. The voice coding technique is a technique of spectrally analyzing voice to remove redundancy of a voice wave and compressing the voice. The voice synthesis technology is a technology for synthesizing human voice and is used for reading out texts. The speech recognition technology is a technology for recognizing speech as a language, but it is also a developing technology and is not as popular as the first two. The speaker recognition technology is a technology for identifying someone from a voice and is applied to security systems and the like, but it is not widely used due to its nature.

A technique widely used in voice processing is spectral analysis. The spectrum analysis is a standard frequency analysis of acoustic analysis, and the power spectrum analysis is particularly widely used. In the power spectrum analysis, first, the input voice signal is sampled by digital processing, and the sampled data is calculated by DFT (Discrete Fourier Transform) or FFT (Fast Fourier Transform) to obtain a frequency component. By analyzing the obtained frequency components and applying phonological processing, data applicable to speech recognition technology and the like can be obtained.

In most speech recognition, after the sampled speech spectrum is analyzed, it is classified into about 100 types of groups by vector quantization so that it can be processed as a label string. After that, a statistical model called Hidden Markov Model (HMM) is used to analyze the data while performing likelihood calculation, and the result is output. The HMM can be used for both the phonological model and the word model, which are the models at the core of the continuous speech recognition technology by unspecified speakers.

The function of the HMM is to observe the input speech pattern to find the best matching symbol or phoneme sequence. In the language model, the appearance probability of a symbol string obtained by statistically analyzing a large amount of text data by HMM or the like is used.

For example, there is a voice recognition system using an acoustic model and a language model. This system has a Japanese dictionary of 60,000 words, and the words contained in it are
It includes acoustic model and language model as statistical data necessary for probability calculation. The voice input from the microphone is converted to digital data, the candidates are narrowed down by comparing it with a dictionary of 60,000 words, and the steps of high-speed large classification, language analysis and detailed voice matching are taken, and finally analyzed with an acoustic model and a language model. The speech recognition is performed by selecting the highest-probability analysis result as the recognition result.

If the above speech recognition is roughly divided into two functions, it consists of acoustic analysis and speech recognition. In acoustic analysis, audio coding, noise processing, correction, etc. are performed. In speech recognition, speech recognition is performed on acoustically analyzed encoded speech data by acoustic processing or linguistic processing, and a word or character string with the highest probability is extracted. An acoustic model and a language model are used for this acoustic processing and language processing. A dictionary is a dictionary in which acoustic patterns, words, character strings, etc. used in each model are registered, and it is possible to improve the recognition rate by expanding the dictionary.

FIG. 1 shows a basic model of the speech recognition engine 1, and the recognition decoder in the figure corresponds to the above speech recognition processing. Voice a (analog voice) input through a microphone is encoded by acoustic analysis 110, and encoded voice b
(Digital voice) converted to voice recognition decoder 120
Passed to. The speech recognition decoder 120 analyzes the coded speech b using the system acoustic model 121, the system dictionary 122, and the system language model 123, and extracts some candidate words. The word with the highest probability is the speech recognition result c
Is extracted as.

Although some devices have been developed to perform voice operations by word level (for example, words such as "right" and "left"), at the natural language level, a large-scale voice recognition system is required. Not used for simple devices.
The reason is that the amount of encoded voice data is large,
The point is that the system of the voice recognition decoder becomes large. The inability to separate the acoustic analysis processing and the speech recognition decoder into separate processing also narrows the field of application using voice data input.

Internet home page access by voice input is not currently performed in a client device having a small memory and a low CPU speed. For example, in i-mode (registered trademark), operations on the Internet are not performed by voice, but are performed by a manually operated input device. At present, it is not possible to realize voice support while having a voice input function.

As a patent for browsing a homepage by voice recognition by a telephone, there is "homepage access device and homepage access method" (Japanese Patent Laid-Open No. 2001-109687). This is to acquire a home page from an Internet access server based on a home page address voice-input from a facsimile terminal or a general telephone terminal and a keyword of a browse item of the home page. In other words, the base base (access server) for voice recognition
Is provided between the user and the website providing site to enable voice input. By using this method, it is possible to access the home page from the client device.

[0012]

The problem to be solved by the present invention is to realize a high recognition rate for voice input without using a learning function in light of the actual problem, and to provide a light system. It is an object to provide a client device and a method for enabling access to a voice-enabled Internet home page by voice recognition.

[0013]

Means for Solving the Problems In order to solve the above-mentioned problems, the invention described in claim 1 is: (1) Human beings who access a home page by voice from a client device having a voice input function. Acoustic analysis means for receiving voice input, performing voice analysis, noise processing, acoustic analysis such as correction, extracting characteristic parameters from voice data, and transmitting the characteristic parameters to a website providing site, (2) the characteristic parameters In response to this, voice recognition is performed using an acoustic model, a dictionary, and a language model, specific voices are converted into symbol strings, necessary operations are performed on the home page, and the selected information is transmitted to the client device. Speech recognition means at the providing site, (3) The client device receives the home page and displays it on the screen or sounds. Receiving means for performing such output is a voice-enabled Internet homepage access system characterized by having.

According to the second aspect of the invention, the speech recognition includes means for creating a rule grammar for analyzing the characteristic parameters to generate a symbol string for each home page and performing voice recognition for each home page. The access system for a voice-enabled Internet home page according to claim 1, wherein:

According to a third aspect of the present invention, in a client device having a voice input function used for browsing an internet home page for accessing a home page by voice, (1) receiving a human voice input, encoding the voice,
A client device having a voice input function, characterized by comprising acoustic analysis means for performing acoustic analysis such as noise processing and correction, extracting characteristic parameters from voice data, and transmitting the characteristic parameters to a website providing site. .

In the system of the present invention, the acoustic analysis unit is incorporated in the client device. On the other hand, at sites that provide voice-enabled Internet homepages, voice recognition is performed based on the digital signals (encoded voice data) obtained by the acoustic analysis unit, the symbol strings required for homepage operations are extracted, and homepage operations are performed. . In order to speed up the transmission rate of voice data, a digital signal generated by a client device is used by generating a feature parameter by extracting a feature amount representing a feature of voice data that does not reduce the recognition rate. On the other hand, for sites that provide voice-enabled Internet homepages, we will develop a rule grammar that performs voice recognition corresponding to feature parameters for each homepage to speed up and reduce weight.

A feature of the system of the present invention is that the acoustic analysis section and the speech recognition section (speech recognition decoder) are separated, and the former is included in the client device and the latter is included in the site. Further, in the present invention, the voice itself is not transmitted from the voice input client device (for example, i-mode (registered trademark) mobile phone), but the feature parameter is created and the feature parameter is transmitted. The feature parameter is obtained by extracting the feature amount of the voice data to the extent that the voice recognition rate is not lowered and parameterizing the set. Specifically, depending on the characteristic parameters, normal voice data 32KB / sec (= 16KH
It is possible to reduce the communication volume to 3.75% (1.2KB / sec) of z, 16bit). The acoustic analysis function for this purpose is converted into hardware and installed in the client device.

On the other hand, at a site providing a home page, a voice recognition decoder performs a voice recognition process on characteristic parameters. The speech recognition decoder performs speech recognition using a command and control type rule grammar as a normal acoustic model and a language model. The rule grammar is developed for each home page, and an acoustic model, a language model, and a dictionary are created for each home page. For this reason, it is possible to reduce the weight and speed of the site voice recognition system, and also achieve a high recognition rate.

The site executes a normal homepage processing operation based on the result of voice recognition of the feature parameter by the voice recognition decoder, and transmits a screen (information) corresponding to the execution result to the user's client device. . Except for voice input and related processing, normal homepage success is used as it is.

[0020]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described with reference to the drawings. In the following, the client device used by the user will be referred to as an i-mode (registered trademark) mobile phone, but basically any client device having a voice input device and capable of connecting to the Internet is basically used. A common explanation holds.

FIG. 2 is a system block diagram for explaining a method of accessing a voice-enabled Internet home page according to the present invention. Button 3 on i-mode (registered trademark) mobile phone 3
Although the home page can be browsed by directly connecting to the home page providing site 4 by 20 operations, the present invention is based on voice input. First, a website providing site
To connect to 4, use standard 320 button operations. When the connected site 4 has a voice recognition system that complies with the standards of the present invention, this is indicated by the mobile phone 3 from the site 4.
Is displayed on the display 330 or sound is output from the speaker. After that, the information can be acquired by voice. The procedure will be described.

1) When connected to the site 4, the main menu is normally displayed on the display 330 and the menu is selected by voice. In this case, if there is a number,
Although only numbers such as “1” and “2” may be pronounced and input, natural language voice input such as “select number 1” and “I want information for number 2” may be performed. 2) When there is a voice input, the hardware-based acoustic analyzer 310 analyzes the voice and extracts the characteristic parameter p,
Send feature parameter p to site 4. 3) The site 4, which has received the characteristic parameter p, determines whether the HP system 400 is a voice input or a key input, and in the case of voice input, passes the characteristic parameter p to the voice recognition engine 410. 4) The voice recognition engine 410 receiving the characteristic parameter p
Voice recognition is performed using the rule grammar 430 corresponding to the screen 420 currently displayed on the mobile phone 3, and the voice recognition result c
(Symbol string) is generated and passed to the HP system 400. HP
The system 400 selects the screen 420 from the voice recognition result c,
Send screen i (screen 420 of site 4) to mobile phone 3. 5) The transmitted screen i is displayed on the display 330.

The above is a series of voice input procedures when the present invention is used. Next, let's take a closer look at voice recognition. FIG. 3 is a diagram showing a voice recognition process of the present invention.
The illustrated acoustic analyzer 310 is included in the mobile phone and the voice recognition engine 410 is included in the site. Especially in the present invention, the voice recognition decoder 41 plays an important role on the site side.
Is 1. Because the acoustic analysis is already done on the mobile phone.

When the voice a is input by the mobile phone, the acoustic analyzer 310 analyzes the analog voice a and performs encoding, noise processing, correction and the like. After that, only the characteristic portion is extracted from the encoded voice within a range that does not deteriorate the voice recognition rate, and the characteristic parameter p is created. This feature parameter p can be reduced to a maximum of 3.75% compared to normal voice data.

The characteristic parameters p (indicated by ○ 1, ○ 2, ○ 3, etc. in the figure represent characteristic amounts, that is, individual characteristic vectors) created by the acoustic analysis device 310 of the mobile phone are transmitted to the site for voice recognition. Is done. The voice recognition decoder 411 that has received the characteristic parameter p performs voice recognition by acoustic probability calculation and linguistic probability calculation while referring to the dictionary 432. An acoustic model for command and control (acoustic model for rule grammar) 431 is used for acoustic probability calculation, and a language model 433 for rule grammar is used for linguistic probability calculation. Particularly, in order to correspond to the characteristic parameter p, the acoustic model 431 for rule grammar is indispensable in the present invention.

The speech recognition result c by the speech recognition decoder 411 is a symbol string (W1, W2, W3 ... In the figure indicate individual words).

In the A / D conversion, the instantaneous value of the waveform sampled by the A / D converter is digitized. This digitization is A /
This is called D conversion. A / D conversion is also called quantization. A
The output of the / D conversion is a pulse code (PC: digital code whose amplitude is quantized by the PCM method). In the calculation of the characteristic parameter, only the characteristic sound is calculated from the PC and corrected to create the characteristic parameter. Therefore,
At this stage, all noises are removed, and only the acoustic digital data necessary for the homepage browsing operation is transmitted to the voice recognition decoder as a characteristic parameter.

As the characteristic parameter, a set of characteristic quantities that well represent the characteristic is used from the voice data. Power spectrum and Δ (delta) usually obtained by spectrum analysis,
ΔΔ (delta delta) is used and is a vector-quantized multidimensional vector quantity. Normally MF every 10ms
CC (12), △ MFCC (8), △△ MFCC
(4), power spectrum, △ power spectrum, △ △
A total of about 27th order feature vectors of the power spectrum are calculated, but in the present invention, finally, about 100 types of groups (labels) are set, and 32 KB / sec of audio data is usually 3.75% data amount (1.2 KB / sec). ) Reduce and send. Specifically, the feature parameter is a feature vector extracted from 16000 (16 KHz) sampling data per second at a window interval of 20 ms every 10 ms.

On the other hand, the speech recognition decoder performs acoustic analysis and language analysis. The models used at this time are the acoustic model for command and control (acoustic model for rule grammar) and the language model for rule grammar, which are registered in the auxiliary storage together with the dictionary, and loaded and executed on the personal computer as necessary. . Since the acoustic model for rule grammar is made simpler than the ordinary acoustic analysis model, the recognition accuracy is slightly lowered, but the memory consumption is small. Furthermore, because it is an acoustic model that matches the screen,
It is possible to prevent deterioration of recognition accuracy.

The language model for the rule grammar is also simpler than the normal language model. Moreover, by using a language model for the rule grammar that corresponds to the screen, we have achieved further weight reduction and recognition accuracy maintenance. The dictionary is divided into a pronunciation dictionary and a language dictionary. In the present invention, since the speech recognition analysis is performed by the rule grammar based on the characteristic parameter, the dictionary is also used as the pronunciation dictionary for the rule grammar (used in the acoustic model for the rule grammar) and the rule grammar. A language dictionary (used in the language model for rule grammar) has been created.
The pronunciation dictionary for rule grammar registered in the present invention is created based on adult men and women excluding children and old people who have many variations.

[0031]

The feature of the present invention is that the acoustic analysis section and the speech recognition decoder section are separated, the acoustic analysis section is incorporated in the client device with a speech input function, and the acoustic analysis result is
The voice data is transmitted to the voice recognition unit as a feature parameter expressed only by a feature vector, and the voice recognition decoder performs voice recognition using a rule grammar (acoustic model, dictionary, language model). The effects of the present invention obtained thereby are as follows.

Up to now, acoustic analysis and voice recognition have an inextricably close relationship with each other, and moreover, encoded voice data (digital voice data) obtained as a result of the acoustic analysis.
Due to its large capacity, separation was difficult and acoustic analysis could not be performed in the client device. On the other hand, according to the present invention, since the voice recognition rate is not deteriorated and the characteristic parameter created by extracting the characteristic voice data necessary for recognition of the voice data is used, the client device has the acoustic analysis function. By performing acoustic analysis on the client device and transmitting the characteristic parameter to the site, the voice data (characteristic parameter) can be transmitted in a short time. Moreover, since the audio data to be transmitted is digital, it is possible to prevent the deterioration of the data during transmission.

By providing a rule grammar for each screen (information) corresponding to the characteristic parameter, it becomes possible to perform voice recognition while maintaining a high recognition rate even with a characteristic parameter having a small amount of data. In addition, by creating a rule grammar suitable for the application, not only the memory consumption can be reduced, but also the speech recognition processing time can be shortened and a high speech recognition rate is realized. ing. In particular, since the present invention maintains a high voice recognition rate without using the learning function, it does not require a troublesome initial work for the user and can deal with an unspecified number of sites.

[Brief description of drawings]

FIG. 1 is a system configuration diagram of a speech recognition engine system in a conventional example.

FIG. 2 is a system configuration diagram for explaining a procedure for obtaining a voice-enabled Internet home page from a client device according to the present invention.

FIG. 3 is a system configuration diagram for explaining a voice recognition procedure in the present invention.

[Explanation of symbols]

1 voice recognition engine 110 acoustic analysis 120 voice recognition decoder 121 system acoustic model 122 system dictionary 123 system language model 3 i-mode (registered trademark) mobile phone (Internet compatible client device) 310 acoustic analyzer (hardware acoustic analysis Device) 320 button or data input operation key 330 Display 4 Homepage providing site (site) 400 HP system 410 Speech recognition engine 411 Speech recognition decoder 420 Screen (information, contents) 430 Rule grammar 431 Acoustic model for command & control (for rule grammar) Acoustic model) 432 Dictionary 433 Language model for rule grammar a Voice (analog voice) b Coded voice (digital voice) p Feature parameters (feature set) ○ 1, ○ 2, ○ 3 ... Etc. characteristic quantity c speech recognition result W1, W2, W3 ...... like recognition result the obtained individual words i screen (information)

─────────────────────────────────────────────────── ─── Continued Front Page (51) Int.Cl. ⁷ Identification Code FI Theme Coat (Reference) G10L 15/28 G10L 9/00 N 19/00 9/18 F 21/02 3/00 521W 537J

Claims (3)

[Claims] 1. When browsing an Internet home page for accessing a home page by voice from a client device having a voice input function, (1) receiving a human voice input,
Perform audio analysis such as voice coding, noise processing, and correction,
Acoustic analysis means for extracting a characteristic parameter from voice data and transmitting the characteristic parameter to a website providing site, (2) receiving the characteristic parameter, performing voice recognition using an acoustic model, a dictionary, and a language model to obtain a specific voice. Is converted to a symbol string, perform the necessary operations on the homepage,
It is characterized by having a voice recognizing means in a voice-compatible homepage providing site for transmitting the selected information to the client device, and (3) a receiving means for receiving the homepage by the client device and performing screen display or voice output. Access system for voice-enabled Internet websites. 2. The speech recognition includes a unit for creating a rule grammar for analyzing a characteristic parameter to generate a symbol string for each home page and performing voice recognition for each home page. Access system for voice-enabled Internet homepages described in. 3. A client device having a voice input function used for browsing an internet homepage for accessing a homepage by voice, (1) receiving human voice input, and performing acoustic analysis such as voice encoding, noise processing, correction, etc. A client device having a voice input function, comprising: an acoustic analysis means for performing a characteristic parameter extraction from voice data and transmitting the characteristic parameter to a website providing site.