CN100568222C - Divergence elimination language model - Google Patents

Abstract

A language model for a language processing system such as a speech recognition system, structured as a function of associated characters, word phrases, and contextual tokens. A method and apparatus for generating a training corpus for training the language model, and a system or module for using the disclosed language model.

Description

disambiguation language model

Background of the invention

The present invention relates to language modeling. More particularly, the invention relates to creating and using a language model for minimizing ambiguity during character recognition, such as input speech.

Accurate speech recognition requires more than an acoustic model to select the correct word spoken by the user. In other words, if a speech recognizer had to select or determine which word was pronounced, the speech recognizer would obviously not perform satisfactorily if all words had the same pronunciation. A language model provides a method or means of specifying which sequences of words in a vocabulary are possible, or generally, information about the similarity of various sequences of words.

Speech recognition is often viewed as a top-down form of language processing. Two general forms of language processing include "top-down" and "bottom-up". Top-down language processing starts by identifying the largest unit of language, such as a sentence, and processes it by classifying it into smaller units, such as phrases, which in turn are subdivided into smaller units, such as words. In contrast, bottom-up language processing starts with words and builds larger phrases and/or sentences from there. Both forms of language processing can benefit from language models.

A well-known classification technique uses an N-gram language model. Because N-word columns can associate large amounts of data, N-word correlations usually provide shallow structure that suppresses both syntax and semantics. Although N-gram language models can perform well for general dictation, homophones can make large errors. A homonym is an element of a language code, such as a character or syllable, that is, one of two or more elements that sound similar but have different spellings. For example, when a user is spelling characters, the speech recognition module may output wrong characters because some characters sound the same. Likewise, the speech recognition module may output erroneous characters for different characters that sound similar to each other when pronounced (such as "m" and "n").

The problem of ambiguity is especially prevalent in languages such as Japanese or Chinese, which are primarily written in the Kanji writing system. The characters of these languages are many complex pictographs representing sounds and meanings. These characters formed a limited number of syllables, which in turn produced a large number of homonyms, greatly increasing the time required to generate documents through dictation. In particular, incorrect homophone characters must be identified and correct homonym characters inserted in the file.

There is thus a continuing need to develop new methods for minimizing ambiguity in pronouncing homonyms and similar-sounding sounds with different meanings. With the development of technology, speech recognition is provided in more applications, which requires a more accurate language model.

Summary of the invention

Speech recognizers typically use a language model such as an N-gram language model to improve accuracy. A first aspect of the invention involves generating a language model that is particularly useful when a speaker is recognizing a character or characters (eg a syllable) such as when spelling a word. The language model helps disambiguate homophones and different characters that sound similar to each other. The language model is constructed from a training corpus containing a string (which can be a single character) of associated elements, a word phrase with strings (which can be a single word), and a contextual token. Using a wordlist or dictionary, the training corpus can be automatically generated by forming a partial sentence or phrase for each wordphrase containing the wordphrase, context tokens, and a string of wordphrases. In another embodiment, a phrase is generated for each token of the word phrase.

Another aspect of the present invention is a system or module using the above-described language model for recognizing spoken characters. Combined with contextual markers in related word phrases when speaking a string, the speech recognition module determines how the user is spelling or recognizing characters. The speech recognition module will only output the recognized characters, not contextual tokens or associated word phrases. In yet another embodiment, the speech recognition module compares the recognized character to a recognized word phrase to verify that the correct character has been recognized. If the recognized character is not in the recognized word phrase, the output character is a character of the recognized word phrase.

Brief description of the drawings

Accompanying drawing 1 is a block diagram of a language processing system.

Figure 2 is a block diagram of a typical computing environment.

Accompanying drawing 3 is a block diagram of a typical speech recognition system.

Accompanying drawing 4 is the flowchart of a kind of method of the present invention.

Accompanying drawing 5 is a block diagram for realizing the method of accompanying drawing 4.

Figure 6 is a block diagram of a voice recognition module and an optional character verification module.

Detailed Description of Illustrative Embodiments

FIG. 1 shows a language processing system 10 that receives a language input 12 and processes the language input 12 to provide a language output 14 . For example, the language processing system 10 may be embodied as a speech recognition system or module that receives linguistic input 12 in a language spoken or recorded by a user. The language processing system 10 processes spoken language and provides as an output recognized words and/or characters in the form of textual output.

During processing, speech recognition system or module 10 may access a language model 16 to determine which word, in particular, which homonym or other similar-sounding element of the spoken language. The language model 16 encodes a specific language, such as English, Chinese, Japanese, and so on. In the illustrated embodiment, the language model 16 may be a statistical language model, such as an N-gram language model, a context-free grammar, or a mixture of the above, all of which are known in the art. A main aspect of the invention is a method of creating and constructing a language model 16 . Another major aspect is the use of the method in speech recognition.

Before discussing the invention in detail, it is useful to provide an overview of the operating environment. Figure 2 and the associated discussion provide a brief overview of one suitable computing environment 20 in which the present invention may be practiced. The computing environment 20 is only one example of a suitable computing environment and is not intended to limit the scope of use or functionality of the invention. Neither should the computing environment 20 be interpreted as having any dependency or requirement relating to any one or combination of components within the illustrated exemplary operating environment 20 .

The invention operates with numerous other general purpose or special purpose computing system environments or configurations. Examples of known computing systems, environments, and/or configurations suitable for use with the present invention include, but are not limited to, personal computers, server computers, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, set-top boxes , programmable consumer electronic devices, network PCs, minicomputers, mainframes, distributed computing environments including any of the above systems or devices, and similar devices. Furthermore, the present invention can be used in telephone systems.

The invention may be described generally in the context of computer-executable instructions, such as program modules, being executed by a computer. Generally, program modules include routines, programs, objects, components, data structures, etc. that perform particular tasks or implement particular abstract data types. The invention also has application in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network. In a distributed computing environment, program modules may be located in both local and remote computer storage media including memory storage devices. The tasks performed by the programs and modules are explained below with the aid of the figures.

Referring to FIG. 2 , an exemplary system for implementing the present invention includes a general purpose computing device in the form of computer 30 . The computer 30 is comprised of, but is not limited to, a processing unit 40 , a system memory 50 , and a system bus 41 coupling the system components including the system memory to the processing unit 40 . The system bus 41 may be any of several types of bus structures including a memory bus or memory controller, a peripheral bus, and a local bus using any of the bus structures. By way of example, and not limitation, such architectures include Industry Standard Architecture (ISA) bus, Micro Channel Architecture (MCA) bus, Enhanced ISA (EISA) bus, Video Electronics Standards Association (VESA) local bus, and the A Peripheral Component Interconnect (PCI) bus called an add-in board bus.

Computer 30 typically includes a variety of computer readable media. Computer readable media can be any available media that can be accessed by computer 30 and includes both volatile and nonvolatile media, removable and non-removable media. By way of example, and not limitation, computer readable media may comprise computer storage media and communication media. Computer storage media includes volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data medium. Computer storage media including, but not limited to, RAM, ROM, EEPROM, flash memory or other storage technology, CD-ROM, Digital Versatile Disc (DVD) or other optical disk storage, cassette tapes, magnetic tape, magnetic disk storage or other disk storage device, or any other medium that can be used to store the desired information and that can be accessed by computer 20 . Communication media typically embodies computer readable instructions, data structures, program modules or other data in a modulated data signal such as a carrier wave or other transport structure and includes any information delivery media. The term "modulated data signal" means a signal that has one or more bursts or altered in such a manner as to encode information in the signal. By way of example, and not limitation, communication media includes wired media such as a wired network or direct-wired connection, and wireless media such as audio, FR, infrared and other wireless media. Combinations of any of the above should also be included within the scope of computer readable media.

System memory 50 includes computer storage media in the form of volatile and/or nonvolatile memory such as read only memory (ROM) 51 and random access memory (RAM) 52 . A Basic Input/Output System 53 (BIOS) is typically stored in ROM 51 and contains basic routines that help transfer information between components within computer 30, such as during start-up. RAM 52 typically contains data and/or program modules that may be immediately accessed and/or operated on by processing element 40 . By way of example, and not limitation, FIG. 2 shows an operating system 54 , application programs 55 , other program modules 56 and program data 57 .

The computer 30 also includes other removable/non-removable, volatile/nonvolatile computer storage media. By way of example only, FIG. 2 shows a hard drive 61 reading from or writing to non-removable, non-volatile magnetic media, and a hard drive 61 reading from a removable, non-volatile magnetic disk 72. A disk drive 71 that reads from or writes to, and an optical disk drive 75 that reads from or writes to a removable, non-volatile optical disk 76 such as a CD ROM or other optical media. Other removable/non-removable, volatile/nonvolatile computer storage media that can be used in typical operating environments include, but are not limited to, cassette tapes, flash memory cards, digital versatile discs, digital video tapes, solid-state RAM , solid state ROM and similar components. Hard disk drive 61 is typically connected to system bus 41 through a non-removable memory interface such as interface 60 , and magnetic disk drive 71 and optical disk drive 75 are typically connected to system bus 41 through a removable memory interface such as interface 70 .

The drives and their associated computer storage media, discussed above and shown in FIG. 2 , provide storage of computer readable instructions, data structures, program modules and other data for the computer 30 . In FIG. 2 , for example, a hard drive 61 is illustrated storing an operating system 64 , application programs 65 , other program modules 66 and program data 67 . Note that these parts can be the same as or different from operating system 54, application programs 55, other program modules 56 and program data 57. Operating system 64, application programs 65, other program modules 66 and program data 67 are given different labels to illustrate, at least, they are different versions.

A user may enter commands and information into the computer 30 through input devices such as a keyboard 82, a microphone 83, and a pointing device 81 such as a mouse, trackball or touch pad. other input devices

(not shown) may include a joystick, game pad, satellite reflector, scanner or similar device. These and other input devices are typically connected to processing unit 40 through a user input interface 80 coupled to the system bus, but may also be connected through other interfaces and bus structures, such as parallel ports, game ports or universal serial bus (USB). A display 84 or other type of display device is also connected to system bus 41 via an interface such as video interface 85 . In addition to the display, the computer can also include other peripheral output devices such as speakers 87 and a printer 86 , and the peripheral output devices are connected through an output peripheral interface 88 .

Computer 30 operates in a network environment using a local connection to one or more remote computers, such as remote computer 94 . Remote computer 94 may be a personal computer, a handheld device, a server, a router, a network PC, a statistical device or other public network node, and typically includes many or all of the components described above in connection with computer 30 . The logical connections depicted in Figure 2 include a local area network (LAN) 91 and a wide area network (WAN) 93, and may include other networks. Such networking environments are commonplace in offices, enterprise-wide computer networks, intranets and the Internet.

When used in a LAN networking environment, computer 30 is connected to LAN 91 through a network interface or adapter 90 . When used in a WAN networking environment, computer 30 typically includes a modem 92 or other means for establishing communications over WAN 93, such as the Internet. Modem 92, which may be internal or external, is connected to system bus 41 through user input interface 80 or other suitable structure. In a network environment, program modules described relative to the computer 30, or portions thereof, can be stored in the remote memory storage device. By way of example, and not limitation, FIG. 2 shows a remote application 95 residing on a remote computer 94 . It will be appreciated that the network connections shown are typical and other means of establishing a communications link between the computers may be used.

A typical implementation of a speech recognition system 100 is shown in FIG. 3 . The speech recognition system 100 includes a microphone 83, an analog-to-digital (A/D) converter 104, a training module 105, a feature extraction module 106, a dictionary storage module 110, an acoustic model 112 along a senone tree, a tree search engine 114, language model 16 and a general language model 111. It should be noted that the entire system 100, or a portion of the speech recognition system 100, can be implemented in the environment shown in FIG. 2 . For example, microphone 83 can be used as an input device to computer 30 through an appropriate interface and through A/D converter 104 . The training module 105 and the feature extraction module 106 can be hardware modules in the computer 30, or software modules stored in any information storage device disclosed in FIG. 2, and accessed by the processing unit 40 or other suitable processors. In addition, the dictionary storage module 110, the acoustic model 112 and the language models 16 and 111 are also preferably stored in any storage device as shown in 2. Furthermore, the tree search engine 114 is implemented in the processing component 40 (which may include one or more processors) or executed by a dedicated speech recognition processor used by the computer 30 .

In the illustrated embodiment, speech is provided by the user as an input to the system 100 to the microphone 83 in the form of an audio sound signal during speech recognition. The microphone 83 converts the audio voice signal into an analog electronic signal, and supplies it to the A/D converter 104 . The A/D converter 104 converts the analog voice signal into a series of digital signals and provides them to the feature extraction module 106 . In one embodiment, the feature extraction module 106 is a conventional array processor that performs spectral analysis on the digital signal and calculates a magnitude for each frequency bin of the spectrum. In one illustrated embodiment, the signal is provided to feature extraction module 106 via A/D converter 104 at a sampling frequency of approximately 16 kHz.

Feature extraction module 106 divides the digital signal received from A/D converter 104 into frames containing a plurality of digital samples. The duration of each frame is about 10 milliseconds. These frames are then encoded by the feature extraction module 106 into a feature vector reflecting spectral features of multiple frequency bands. In the case of discrete and semi-continuous hidden Markov models, the feature extraction module 106 also encodes the feature vector into one or more codewords using vector quantization techniques and a codebook obtained from the training data. Accordingly, the feature extraction module 106 provides its output as a feature vector (or codeword) for each utterance. The feature extraction module 106 provides feature vectors (or codewords) at a frequency of approximately one feature vector or (codeword) every 10 milliseconds.

The distribution of output probabilities is then calculated according to a Hidden Markov Model using the feature vectors (or codewords) of the particular frame being analyzed. These probability distributions are then used to perform a Viterbi decoding process or similar type of processing technique.

After receiving the codeword from the feature extraction module 106 , the tree search engine 114 accesses the information stored in the acoustic model 112 . The model 112 stores an acoustic model, such as a Hidden Markov Model, which represents the speech components detected by the speech recognition system 100 . In one embodiment, the acoustic model 112 includes a senone tree associated with each Markov state in the hidden Markov model. Hidden Markov models represent phonemes in one illustrated embodiment. Based on the senones in the acoustic model 112, the tree search engine 114 determines the most similar phonemes represented by feature vectors (or code words) received from the feature extraction module 106, and representations of speaking patterns received from system users.

Tree search engine 114 also accesses a dictionary stored in module 110 . The information received by the tree search engine 114 based on its access to the acoustic model 112 is used to search the dictionary storage module 110 to determine a word that most likely represents the codeword or feature vector received from the feature extraction module 106 . At the same time, the search engine 114 accesses the language models 16 and 111 . In one embodiment, the language model 16 is a sequence of N characters for identifying the most similar character or characters represented by input speech to a word, comprising a character(s), contextual markers, and a word phrase to recognize characters. For example, the input speech could be "N as inNancy", where "N" (which can also be lowercase) is the desired character, "as in" is the context marker, and "Nancy" is a word phrase related to the character "N" to clarify or identify the desired character. As for the phrase "N as in Nancy", the output of the speech recognition system 100 may only be the character "N". In other words, the speech recognition system 100 determines that the user has selected spelling characters after analyzing the input speech data about the phrase "N as in Nancy". Therefore, context tokens and related word phrases are ignored from the output text. The search engine 114 can remove contextual notations and related word phrases as necessary.

It should be noted that in this embodiment, the language model 111 is a word-N-word sequence for identifying the most similar word represented by a generally dictated input speech. For example, when speech recognition system 100 is embodied as a dictation system, language model 111 provides an indication of the most similar words for general dictation; The language model 111 for the same phrase has a higher value. A higher value for the language model 16 is used as an indication in the system 100 that the user is identifying with contextual tags and word phrases. Thus, for an input phrase with contextual markers, the search engine 114 or other processing components of the speech recognition system 100 ignore the contextual markers and word phrases and output only the required characters. The discussion of the use of language models 16 continues below.

Although the speech recognition system 100 described herein uses HMM models and senone trees, it should be understood that this is only an illustrative embodiment. Those of ordinary skill in the art will appreciate that the speech recognition system 100 can take many forms, all that is required is to use the features of the language model 16 and provide as output the text spoken by the user.

It is well known that a statistical N-gram language model produces a probability estimate for a word, given the sequence of words up to that word. (i.e. assume the word's history H). The N character string language model only considers the first (n-1) words in the history H that have an impact on the probability of the next word. For example, a double-array (or 2-array) language model considers the previous word as having an influence on the next word. Therefore, in an N character array language model, the probability of a word appearing is expressed as follows:

P(w/H)＝P(w/w1，w2，…w(n-1)) (1)

where w is an important word;

w1 is the word at position n-1 in front of word w;

w2 is the word at position n-2 preceding word w; and

w(n-1) is the first word preceding word w in the sequence.

At the same time, the probability of a sequence of words is determined based on the increasing probability of each word given its history. Therefore, the probability of a word sequence (w1...wm) is expressed as follows:

$\mathrm{<span\; class="notranslate">\; P</span>}\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; <figure-callout\; id="1"\; label="w"\; filenames="C021065300016H1.png,C021065300019H1.png"\; state="\{\{state\}\}">w</figure-callout></span>}\mathrm{<span\; class="notranslate">\; 1</span>}\mathrm{<span\; class="notranslate">\; \&Lambda;wm</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\underset{\mathrm{<span\; class="notranslate">\; i</span>}<span\; class="notranslate">\; =</span>\mathrm{<span\; class="notranslate">\; 1</span>}}{\overset{\mathrm{<span\; class="notranslate">\; m</span>}}{\mathrm{<span\; class="notranslate">\; \&Pi;</span>}}}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; P</span>}<span\; class="notranslate">\; (</span>{{\mathrm{<span\; class="notranslate">\; w</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}}^{\mathrm{<span\; class="notranslate">\; 20</span>}}<span\; class="notranslate">\; /</span>{\mathrm{<span\; class="notranslate">\; h</span>}}_{\mathrm{<span\; class="notranslate">\; i</span>}}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; )</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; -</span><span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; 2</span>}<span\; class="notranslate">\; )</span>$

N-gram models are obtained by applying an N-gram algorithm to a corpus of textual training data (a collection of phrases, sentences, sentence segments, paragraphs, etc.). An N-word algorithm can use, for example, well-known statistical techniques such as the Katz technique, or the binomial back-end distribution compensation technique. In using these techniques, the algorithm estimates the probability that a word w(n) will follow a sequence of words W1, W2, . . . , W(n-1). Together, these probability values form an N character sequence language model. Certain aspects of the invention described below can be used to construct a standard statistical N-character sequence model.

The first main aspect of the invention is shown in Figure 4 as a method 140 of creating a language model for use in a language processing system to indicate characters. Referring also to FIG. 5 , a system or apparatus 142 comprising means with instructions for implementing method 140 . In general, method 140 includes, for each word phrase of a word phrase list, associating, at step 144, the word phrase's character string and the word phrase with a contextual tag identifying the character string. It should be noted that the character string may consist of a single character. Likewise, a word phrase can contain a single word. For example, for a character string equal to one character and a word phrase equal to one word, step 144 associates one character of the word with the context token in word table 141 for each word. A contextual marker is usually a word or word phrase in a particular language used by the speaker to identify a linguistic element within the word phrase. Examples of contextual markers in English include "as in", "for example", "as found in", "like", "such as", etc. Similar words or word phrases can also be found in other languages, such as の in Japanese and の in Chinese. In one embodiment, step 144 includes constructing a corpus of word phrases 143 . Each word phrase includes a character string, word phrase and context token. Typically, when a single character is associated with a word, the first character is used, although another character of the word may also be used. Examples of such word phrases include "N as in Nancy", "P as in Paul", and "Z as in zebra".

In another embodiment, another character of a word is associated with the word and a contextual marker, whereas in some languages, such as Chinese, where many words consist of only one, two or three characters, this would have the effect of making each of the words It is helpful for characters to be associated with words in contextual markup. As noted above, a simple way to associate desired characters with corresponding words and context tokens is to form identical word phrases. Thus, given a wordlist 141, a corpus of word phrases 143 for training a language model can easily generate all the required contextual tokens.

Based on the corpus 143 , the language model 16 is constructed using a conventional construction model 146 , such as an N-gram construction model, implementing well-known techniques for constructing a language model 16 . Block 148 represents constructing the language model 16 in method 140, wherein the language model 16 includes, but is not limited to, an N-gram language model, a context-free grammar, or a mixture of the above.

The generated phrase can be assigned an appropriate value, which will generate an appropriate probability value according to the formation of the language model. In the example above, "N as in Nancy" is more likely to be said than the phrase "N as in notch". Thus, yet another feature of the invention includes correcting the probability score for each associated string and word phrase in the language model. The probability score may be tuned for language model 16 generation. In another embodiment, the probability score may be corrected by including a sufficient number of identical word phrases in the corpus 143 to produce an appropriate probability value in the language model for the associated character and word phrase. The probability value can also be a function of the likelihood that the word phrase is used. Typically, there are word phrases that are used more frequently than other words to identify one or more characters. Such word phrases may be specified or otherwise provided with a higher probability value in the language model.

FIG. 6 shows a speech recognition module 180 and language model 16 . The speech recognition module 180 may be of the type described above; however, it should be understood that the speech recognition module 180 is not limited to this embodiment, but may take many forms. As noted above, speech recognition module 180 receives data representing input speech and accesses language model 16 to determine whether the input speech includes phrases with contextual markers. Upon detection of a word phrase with a contextual marker, the speech recognition module 180 provides as output only the single or multiple characters associated with the contextual marker and word phrase, rather than the contextual marker or word phrase. In other words, although the speech recognition module detects the complete phrase "N as in Nancy", the speech recognition module provides only "N" as output. This output is especially useful in dictation systems, where the speaker individually selects the desired single or multiple characters to indicate.

At this point, it should be noted that the language model 16 described above is substantially constructed of associated character strings, word phrases, and contextual tokens, thus allowing the language model 16 to be responsive to input speech having this form. In the embodiment of FIG. 3 , the generic language model 111 may be used as a specific form of input speech without character strings, word phrases, and contextual tokens. However, it should be understood that in both embodiments language models 16 and 111 may be combined if desired.

Upon receipt of input speech and access to language model 16, speech recognition module 180 determines a recognized character string and a recognized word phrase for the input speech. In many cases, the recognized string will be correct due to the use of the language model 16 . However, in yet another embodiment, a character verification module 182 may be included to correct at least some of the errors made by the speech recognition module 180 . The character verification module 182 accesses the recognized character string and the recognized word phrase determined by the speech recognition module 180 and compares the recognized character string and the recognized word phrase, and in particular, verifies that the recognized character string is present in the recognized word phrase. If the recognized character string is not in the recognized word phrase, it is clear that an error has occurred, although the error may be caused by the speaker dictating the wrong phrase such as "M as in Nancy", or it may be misunderstood by the speech recognition module 180 A recognized string or a recognized word phrase. In one embodiment, the character verification module 182 may assume that the error is more likely in the recognized character string and, therefore, replace the recognized character string with characters present in the recognized word phrase. Replacing the recognized character string with characters of the recognized word phrase may be based on a comparison of the sound similarity between the recognized character string and the characters in the recognized word phrase. Thus, the character verification module 182 has access to stored data pertaining to the sounds of individual characters when spoken. Using the characters present in the recognized word phrase, the character verification module 182 compares the sound data of each character stored in the recognized word phrase with the recognized character string. Provides the closest character as output. To those of ordinary skill in the art, character verification module 182 may be included in speech recognition module 180; however, for purposes of explanation, character verification module 182 is illustrated separately.

Although the present invention has been described with reference to preferred embodiments, various changes in form and description can be made by those skilled in the art without departing from the spirit and scope of the invention.

Claims (24)

1. system and discern the input voice with the method for pointing character of using a computer by language model, this method comprises:

Create language model, comprising:

Structure training corpus comprises:

For comprising, character and this word phrase of this word phrase is associated to generate the contextual tagging phrase of this training corpus automatically with the contextual tagging of this character of expression identification based on each the word phrase in the word phrase dictionary of Chinese character; And

Use this training corpus to create described language model;

Character when identification is spoken comprises:

Reception has the input voice of character string, wherein character string comprises the contextual tagging phrase that comprises character, comprises contextual tagging phrase based on Chinese character, has word phrase and contextual tagging based on Chinese character, and wherein said contextual tagging has been indicated the ambiguity of delete character;

Under the situation that does not have prompting, detect the contextual tagging phrase of importing in the voice that receives;

The instruction that execution conducts interviews to described language model, wherein said language model comprise N character row language model, and this model has the probabilistic information of the contextual tagging phrase that is generated; And

Export this character with text, and do not have this word phrase and contextual tagging of the contextual tagging phrase that detected.

2. the process of claim 1 wherein that this language model comprises a context-free grammar.

3. the process of claim 1 wherein that association comprises that first character with each word phrase is associated with this word phrase.

4. the process of claim 1 wherein that association comprises another character with at least a portion word phrase, rather than first character, be associated with corresponding word phrase.

5. the process of claim 1 wherein that each character that association comprises at least a portion word phrase is associated with corresponding word phrase.

6. the process of claim 1 wherein that association comprises is associated each character of each word phrase with corresponding word phrase.

7. the method for claim 1 also is included as each relevant character and word phrase correction probability score in the language model.

8. the method for claim 10, wherein contextual tagging comprises " " in the Japanese.

9. the process of claim 1 wherein contextual tagging comprise in the Chinese " ".

10. the process of claim 1 wherein that each word phrase all is a word that comprises at least one character.

11. the process of claim 1 wherein that output character comprises character string output as the probability function that is stored in the language model.

12. the method for claim 11, wherein output character comprises the function of character output as the N character row probability of receive input voice.

13. the method for claim 11, wherein output character comprises the unique function of character output as receive input voice.

14. the method for claim 11, when the character of identification was not present in the word phrase of identification, the character that is output was a character of the word phrase of identification.

15. a computer system that is used to discern the input voice comprises:

The language model of indication contextual tagging phrase and described contextual tagging phrase probabilistic information, wherein said contextual tagging phrase are basically by the relevant character string based on Chinese character, word phrase and the contextual tagging with character string; And

Be used to receive the identification module that the data of voice are imported in indication, described identification module detects the existence of the contextual tagging phrase in the input voice that receives under the situation that does not have the pointing character string as the prompting of text; Visit described language model; And according to the probabilistic information in the language model, output based on the character string of Chinese character as text that at least some detected by the said contextual tagging phrase of user.

16. computer system as claimed in claim 15, it is characterized in that, described identification module is handled the contextual tagging phrase that is detected in the mode that is different from other input voice, and described processing is by only exporting the character string in the detected contextual tagging phrase.

17. computer system as claimed in claim 15 is characterized in that, described language model comprises statistical language model.

18. computer system as claimed in claim 15 is characterized in that, described language model comprises N character row language model.

19. computer system as claimed in claim 15 is characterized in that, described language model comprises the context-free language model.

20. computer system as claimed in claim 15 is characterized in that, described language model output string is as the function that institute's identification string and institute's identified word phrase are compared.

21. computer system as claimed in claim 20 is characterized in that the character string of being discerned does not exist in the word phrase of being discerned, the described character string that is output is the character string of the word phrase discerned.

22. computer system as claimed in claim 15 is characterized in that, each described word phrase is a word.

23. computer system as claimed in claim 21 is characterized in that, each described character string is single character.

24. computer system as claimed in claim 15 is characterized in that, each described character string is single character.

Images