JP4468264B2 - Methods and systems for multilingual name speech recognition - Google Patents

Description

  The present invention relates generally to speech recognition processes. The present invention is particularly useful for multi-language name speech recognition using a personal electronic device, but is not necessarily limited thereto.

  Personal electronic devices such as mobile phones, personal digital assistants (PDAs), and small wireless paging devices have become popular throughout the industrialized world. Millions of users currently rely on such devices for quick and easy access and communication of electronic information. The weight and size of these devices generally increases their convenience by making them easy to carry, for example in a pocket or purse. However, the downside of these devices is that the haptic interfaces on devices such as keypads and buttons are often very small and unwieldy.

  Thus, speech recognition is a valuable function for many personal electronic devices. For example, the voice recognition function allows a car driver to issue a simple command to the personal electronic device without taking his eyes off the road. Also, voice recognition can be performed easily and often faster than the time required to enter instructions on a small keypad, so voice recognition can access, for example, address book entries in a PDA. Convenience can be improved.

  Thus, the speech recognition system is a popular means for executing computer programs and accessing databases. However, the miniaturization of personal electronic devices may limit the performance of an integrated speech recognition system. Effective speech recognition often requires relatively large databases and significant processing speed, but the memory capacity and processing power of small electronic devices are generally limited. In order to overcome these limitations, personal electronic device speech recognition systems are usually customized only for limited specific situations. For example, such systems often rely on speakers and interpret speech patterns only for specific speakers, as described in more detail below. Also, such systems are often language dependent and are designed for only a limited vocabulary. These design compromises allow the system to function fairly well with the limited resources of personal electronic devices for specific purposes.

Speech recognition systems generally work by matching input words with an acoustic model stored in a database. The matched acoustic model is then matched against an entry in the dictionary database to complete word and sentence recognition. Acoustic models often consist of Hidden Markov Models (HMM). The HMM is a statistical description including an average vector and a variance vector, and describes speech units such as words and phonemes. The HMM pattern matching is then used to determine whether the acoustic model in the speech recognition database matches the words entered by utterance. HMMs are generally based on a probability function consisting of several complex Gaussian probability distribution functions (PDF) called Gaussian mixture. Speech pattern matching is therefore a process of matching Gaussian mixing with input speech words. Thus, the available sophistication of HMM pattern matching acoustic models is an important variable factor that speech recognition system designers must consider when making the necessary compromise between performance and memory and processing resources. .

Another compromise in speech recognition systems relates to the system's ability to recognize multiple users' speech. Thus, speech recognition systems are further classified as either speaker independent systems or speaker dependent systems. While speaker-independent systems are designed to recognize the speech of any speaker in any language, speaker-dependent systems are educated to recognize the speech of a single speaker. The Speaker-independent systems typically include an acoustic database containing HMMs obtained from multiple training speakers. The HMM obtained from the training speaker's speech is a Gaussian mixture parameter that is intended to be representative of speech patterns found in larger speaker groups. Such systems are generally less accurate than speaker dependent systems. This is because the speech model must be compromised to accommodate different speech attributes, and a speaker-independent system will not fit the specific speech attributes of any particular speaker using that system Because.

  Speaker dependent systems are tuned to recognize specific speech patterns of individual speakers. Typically, during the training routine, the speaker loads scripts containing various speech patterns into the speaker-dependent system. The training speech is then tailored to the script so that the system can be adjusted to the speaker's unique speech attributes so that the system recognizes the speaker's speech more accurately during speech recognition. become. However, speaker dependent systems are often undesirable in situations where many people need to use speech recognition systems. For example, a voice recognition system built into a mobile phone allows a user to speak a command and operate the device by the command being recognized by the phone. However, the primary user of a mobile phone may want many friends, colleagues, or family members to be able to use the phone's voice recognition function. Because such a second user of the phone may need a voice recognition function for only a short period of time, the second user must first let his phone recognize his / her voice before using the voice recognition function. That is inconvenient.

  Finally, speech recognition acoustic models are usually designed specifically for a single language. Therefore, a speech recognition system that can recognize multilingual speech requires a large number of acoustic models, which also increases the memory requirements and sophistication of the system.

  Recently, bilingual speech recognition devices have been developed for personal electronic devices. Thus, for example, a bilingual user of a mobile phone can call a name from an address book stored on the phone using one of two languages, English and Mandarin. Due to the acoustic models and vocabulary databases specific to the different languages used on these devices, users generally have to switch the phone language mode to one specific language before using the speech recognition function. Don't be. However, it is inconvenient that a specific language must be selected in advance, for example, when the address book includes variously mixed names of two languages or other contact information. Also, because a specific language must be pre-selected, the system uses speech recognition to create a two-part name mixed with multiple languages, for example, a name whose first name is English and whose last name is Mandarin Chinese. It cannot be specified.

  Thus, multilingual names can be recognized without the need for manual switching between language modes, and speaker-independent speech recognition that effectively uses the limited resources of personal electronic devices. There is a need for improved methods and systems.

Therefore, according to one aspect of the present invention, there is provided an improved speech recognition method for multilingual names, the step of storing text representing a plurality of names consisting of characters in an electronic device, and each of the names Using a step of identifying at least one language and a plurality of language-specific character / sound converters (hereinafter referred to as “language-specific character / sound converters”) Converting to a phonetic unit; receiving a word spoken by a microphone associated with the electronic device; converting the word to a feature vector; and converting the feature vector into the order of at least one name And a step of collating with a series of pronunciation units.

  The multi-language includes Mandarin Chinese, and the step of identifying at least one language for each of the names is whether the name is composed of Chinese alphabet characters or Roman alphabet characters. Preferably, the method includes a step of determining and a step of determining whether the name of the Roman alphabet is a Chinese sword.

The multi-language is preferably composed of a Western language and Chinese.
The plurality of language-specific character / sound converters are preferably composed of a Chinese character / sound converter and a Western language character / sound converter.

Preferably, the Chinese character / sound converter depends on the context, and the Western language character / sound converter does not depend on the context.
The step of matching the feature vector with a series of phonetic units ordered by at least one name comprises comparing the feature vector with the sequence of phonetic units ordered and a Gaussian mixture parameter in an automatic speech recognition engine. It preferably comprises the step of decoding the feature vector.

The automatic speech recognition engine preferably uses a beam search Viterbi algorithm.
Preferably, the name comprises a contact list component stored in the electronic device.

  According to another aspect, the present invention is a multilingual name speech recognition method for receiving a word spoken by a microphone associated with an electronic device, and converting the word into a feature vector. And comparing the feature vector with a series of pronunciation units that are at least one name and in the order of names stored in the electronic device as a representation of characters. At least one language of the name is identified from the characters, and the name is then converted into the ordered series of pronunciation units using a plurality of language-specific character / sound converters.

  According to yet another aspect, the present invention is a system for multilingual name recognition, comprising a microprocessor, at least one memory operably connected to the microprocessor, and the microprocessor. A microphone operatively connected to the The microprocessor executes a code stored in the memory, receives a spoken word with the microphone, converts the word into a feature vector, and the feature vector has at least one name. It operates to collate with a series of pronunciation units that are in the order of names stored in the memory as character representations. At least one language of the name is identified from the letters, and the name is then used to produce the ordered series of pronunciations using a plurality of language-specific letter / sound converters operably connected to the microprocessor. Has been converted to a unit.

Preferably, the name comprises a contact list component stored in the system.
The system is preferably operatively connected to either a mobile phone or a personal digital assistant.

  In this specification, including the claims, the terms “comprising”, “including”, “consisting of”, or similar terms mean non-exclusive inclusions and therefore consist of many elements. The method or apparatus does not include only those elements, but can easily include other elements not described.

In order that the present invention may be readily understood and practiced, preferred embodiments will be described with reference to the accompanying drawings. In the accompanying drawings, the same reference numerals denote the same elements.
FIG. 1 is a schematic diagram illustrating functional components of a system 100 for multilingual name speech recognition according to one embodiment of the present invention. The system 100 operates as follows. The character / sound converter 105 converts the text of the name into an ordered series of pronunciation units. This name is usually one of many names stored as a representation of individual characters on personal electronic devices such as mobile phones and personal digital assistants (PDAs). For example, these names may be stored as part of an electronic device address book or contact list. The character / sound converter 105 first identifies at least one language for the name entered into the system 100. This name is then converted into a series of pronunciation units in the order stored in the open vocabulary dictionary 110. The system 100 also includes a mixed language hidden Markov model (HMM) set 115. The HMM set 115 includes Gaussian mixing parameters representing selected speech patterns in at least two languages.

  After a plurality of names and a series of pronunciation units in the order associated with them are input to the open vocabulary dictionary 110, the system 100, when any of those names is uttered to an input unit such as the microphone 120, the names Can recognize the spoken expression. Microphone 120 can be operatively connected to a voice activated device (VAD). The feature extractor 125 then extracts the feature vector of the spoken name according to conventional speech recognition techniques well known in the art. The feature vector is then decoded by an automatic speech recognition (ASR) engine 130 that compares the feature vector and Gaussian mixing parameters. The ASR engine 130 is further supported by a dynamic grammar network 135. This network 135 is built with the open vocabulary dictionary 110 and guides the search for pronunciation models during the speech recognition process. Finally, the matching name from the public vocabulary dictionary is output from the system 100. This matched name can then be used by the electronic device to retrieve a personal telephone number or other contact information from a contact list, for example.

  Therefore, the present invention is useful in applications that require speech recognition of words and names mixed with multiple languages. For example, in China, ASR capable mobile phones are appearing that are speaker independent Chinese (eg, Mandarin or Cantonese) and English. However, these prior art systems can generally only work with a single language model at a time. For example, if a user wants to use the ASR function to retrieve information in the address book using an English name, the user must first set the ASR function to English. Next, when the same user tries to search for information in the address book using the standard Chinese name, the user first sets the ASR function to standard Chinese before the standard Chinese name can be searched. There must be. However, many mobile phone users in China have a bilingual name in the phone address book where the first part of the name is English and the second part of the name is Mandarin Chinese Can be seen. Thus, prior art ASR systems cannot automatically recognize spoken representations of such bilingual bipartite names. On the other hand, the present invention can recognize such bilingual two-part names and does not require the user to manually switch the ASR from one language to the other.

FIG. 2 is a table showing various names of two different languages and pronunciations consisting of a series of pronunciation units in order. For example, the first name, ie

Is composed only of Mandarin Chinese (kanji), followed by its pronunciation composed of a series of pronunciation units in order including individual Chinese phonemes 205. The next name “John Stone” consists only of English, followed by its pronunciation including individual English phonemes 210. The third name, namely

Is the last name in Mandarin Chinese (Kanji),

And the English name "Jacky", so it is a bilingual two-part name. Nevertheless, the method and system of the present invention can also define the pronunciation of that name, including both English phonemes 210 and Chinese phonemes 205. The features of the present invention that enable such pronunciation parsing of bilingual bipartite names without requiring the user to manually switch languages are described below.

  FIG. 3 is a schematic diagram showing the operation and components of the mixed character / sound converter 105 introduced in FIG. As an example, the mixed character / sound converter 105 shown in FIG. 3 operates to convert characters written in either English or Mandarin Chinese. First, the mixed character / sound converter 105 includes an alphabet identifier 305 that identifies the alphabet used to define at least a portion of the written name stored in the device. When the name storage part is composed of Chinese characters 310, the Chinese characters 310 are directly input to the language-limited standard Chinese character / sound converter 315. However, if the storage portion of the name is composed of English characters 320, the name may be written in either Chinese ping or English. Therefore, that part of the name is further classified by the beep identifier 325. The pinyone discriminator 325 uses a 408 syllable pinyone dictionary that basically identifies all Chinese names represented by a pinyone (except the tone). If the English character 320 is a Chinese ping sound, the English character 320 is input to the standard Chinese character / sound converter 315. However, when the English character 320 is an English word, the English character 320 is input to the language-limited English character / sound converter 330. Both the standard Chinese character / sound converter 315 and the English character / sound converter 330 are operable to convert the names into a series of language-limited pronunciation units in a unique order. It will be apparent to those skilled in the art that other character / sound converters 105 that convert characters in various other languages are also possible with the present disclosure. Thus, the character / sound converter 105 of the present invention can parse bilingual bipartite names into a series of pronunciation units in a single order.

To allow the present invention to function without the user having to manually switch the language model of the system 100, the mixed language HMM set 115 includes at least two acoustic model sets, one for each of the two languages. Contains. For example, according to the above embodiment of the present invention that recognizes both English and Mandarin Chinese, the HMM set 115 includes two monolingual acoustic model sets: a Mandarin Chinese model that depends on context, Combined with a relationship-independent English model. Here, the context refers to a sounding unit that is adjacent to the right and / or left of any sounding unit. In Chinese, these units are called “initial” and “final” as described in more detail below. The three-tone model is a pronunciation model that considers both the left adjacent pronunciation unit and the right adjacent pronunciation unit. If two pronunciation units have the same identity but different left or right contexts, they are considered different triphones.

  One feature that distinguishes Chinese from Western languages such as English is that all Chinese characters are single syllables with consonant / vowel (C / V) structure plus tone. Therefore, syllable recognition is the basis of the configuration of most Chinese speech recognition systems. There are a total of 1254 syllables (408 atonal syllables) in Chinese, which are 22 “vowels” (ie consonants before vowels in syllables) and 38 “vowels” (ie vowels in syllables). Obtained from various combinations of later consonants). There are 21 true vocals and one so-called “zero”. According to a preferred embodiment of the present invention, the zero initial is treated as a true initial. Considering the situation where only limited training data is available, it is generally seen that for Chinese speech, the simultaneous articulation effect within a syllable is significantly greater than the simultaneous articulation effect between syllables. This is due to the Chinese single syllable structure. In the syllable, the acoustic characteristics of the initial are highly dependent on the final, but the final characteristics are almost independent of the final. For example, the initial “t” in the syllable “ta” is pronounced very differently from the same initial in another syllable “tu”, but the final “a” in the syllable “ta” is “cha”. It is pronounced almost the same as "a". Therefore, a reasonable approach in Chinese speech recognition assumes that both the simultaneous articulation effect between syllables and the dependence of the final on the preceding phoneme in the syllable can be ignored, and the initial It is assumed that there is a right-to-left context dependency, and that the final is not dependent on the context. Thus, the preferred embodiment of the present invention uses 155 subsyllables including 117 vocals and 38 finals. Each syllable is then broken down into a pair of subsyllables. An example of such syllable decomposition used in the Chinese acoustic model of the preferred embodiment of the present invention is shown in Table 1.

  In order to reduce the size of the English acoustic model in the HMM set 115 and thus reduce the overall complexity and computational requirements of the system 100, the preferred Chinese / English embodiment of the present invention is a context-independent English acoustic model. Is used. In addition, 40 single notes are used as a basic English modeling unit. One such single note material is the Carnegie Mellon University (CMU) pronunciation dictionary. The CMU pronunciation dictionary contains about 127,000 English words with corresponding pronunciations. The CMU pronunciation dictionary also defines 39 individual phonemes in English. Other dictionaries may be used instead of the above dictionary.

The operation method of the ASR engine 130 that collates the sequence of sounding units in order with the feature vector will be described in more detail. Engine 130 analyzes a series of feature vectors of spoken words received by system 100 using a Viterbi beam search algorithm. The purpose of the engine 130 is to find a series of phonetic units, guided by the grammar network 135 , in which the corresponding Gaussian parameters (Gaussian mixture) of the state sequence are in order that best matches the spoken input speech. Viterbi search is a time-synchronized search algorithm that completely processes time t before proceeding to time t + 1. For time t, each state is updated with the best score from all states at time t-1 (rather than using the sum of all input paths). At the end of the search, the most likely state sequence can be recovered by following these backtracking pointers. Thanks to effective efficiency techniques, it is not necessary to search the entire search space or the entire grid. Instead, only the most probable search state space needs to be searched. A comprehensive HMM set is then created for the system 100. This set is associated with an acoustic model of the final element of the dynamic grammar that is generated online after each time the public vocabulary dictionary is updated. Further details regarding the above algorithm can be found in "Statistical Methods for Speech Recognition" by Jelinek, Frederick (MTT Press 1999 ISBN 0-262-10066-5).

For further explanation of the present invention, FIG. 4 illustrates an exemplary method 400 for converting stored text to a pronunciation unit, according to one embodiment of the present invention including a Mandarin / English public vocabulary dictionary 110. It is the summarized generalized flowchart. In step 405 , the method 400 first stores text representing a plurality of names of characters in an electronic device. In step 410, it is determined whether each name is composed of Chinese alphabet characters or Roman alphabet characters. If the characters constituting the name are kanji, it is specified in step 415 that the language of the name is standard Chinese. However, if the character is a Roman alphabet, the language of the name is not yet determined because the character may be a Chinese pingpong. Thus, in step 420, using a 408 syllable pinyone dictionary that basically identifies all Chinese names represented by pinyongs (excluding tone), it is determined whether the character is a Chinese pinyin. judge. If it is determined that the character is a pop sound, the method 400 proceeds again to step 415 to identify that the language of the name is Mandarin Chinese. Otherwise, at step 425, the language of the name is identified as English.

  If the language is determined to be Mandarin Chinese at step 415, the method 400 continues at Step 430 with the standard Chinese character / sound converter 315 using a sequence of phonetic units whose names are ordered. Convert to However, if it is determined in step 425 that the language is English, the method 400 then uses the English character / sound converter 330 to convert the name to an ordered series of pronunciation units in step 435. To do. Next, the sequence of pronunciation units in order is stored in the open vocabulary dictionary 110.

  FIG. 5 is a generalized flowchart illustrating a method 500 for matching spoken words with names stored in the open vocabulary dictionary 110 according to a preferred embodiment of the present invention. The method 500 first receives the spoken word at the microphone 120 of the electronic device at step 505. The apparatus includes a system 100 for multilingual name speech recognition. In step 510, the word is converted to a feature vector. Then, in step 515, according to the above method, the feature vector of the word is checked against a series of pronunciation units that are in the order of at least one name stored in the open vocabulary dictionary 110.

FIG. 6 is a schematic diagram illustrating an example of a personal electronic device that can execute the speech recognition system 100 of the present invention. This example includes a wireless communication device in the form of a wireless telephone 600 that includes a system 100 for multilingual name speech recognition according to one embodiment of the present invention. The phone 600 includes a radio frequency communication unit 602 connected to communicate with the processor 603. The radiotelephone 600 also includes a keypad 606 and a display screen 605 that are connected to communicate with the processor 603. As will be apparent to those skilled in the art, keypad 606 can be optional since screen 605 can be a touch screen.

  The processor 603 includes an encoder / decoder 611 and associated data storage code read only memory (ROM) 612 for encoding and decoding voice or other signals that can be transmitted or received by the radiotelephone 600. The processor 603 also includes a microprocessor 613 connected to the encoder / decoder 611 by a common data address bus 617, a character read only memory (ROM) 614, a random access memory (RAM) 604, a programmable static memory 616, A SIM interface 618 is included. The programmable static memory 616 and the SIM operatively connected to the SIM interface 618 (often referred to as a SIM card), respectively, are selected input text messages, numbers in the phone number and name fields, respectively. And a telephone number database TND (or address / phone book) consisting of name fields for identifiers associated with one of the two. For example, one entry in the telephone number database TND can be 91999111111 (entered in the number field) and its associated identifier “Steven C! At work” in the name field. The SIM card and static memory 616 can also store a password to allow access to the password protection function of the wireless telephone 600. All of the components of the present invention, such as the character / sound converter 105, the open vocabulary dictionary 110, the mixed language HMM set 115, the feature extractor 125, the ASR engine 130, the dynamic grammar network 135, etc. are all code read only memory (ROM). ) 612, character read only memory (ROM) 614, random access memory (RAM) 604, static memory 616, and SIM card may be partially or fully stored. The microprocessor 613 has ports for connection to an alarm 615 that typically includes a keypad 606, a screen 605, an alarm speaker, a vibrator motor, and associated drivers. The microprocessor 613 has a port for connection to the microphone 120 and the communication speaker 640. The character read only memory 614 stores a code for decoding or encoding a text message received by the communication unit 602. In this embodiment, the character read only memory 614 also stores an operating code (OC) for the microprocessor 613 and code for performing functions related to the radiotelephone 600.

  The radio frequency communication unit 602 is a combination transceiver having a common antenna 607. The communication unit 602 has a transceiver 608 connected to an antenna 607 via a radio frequency amplifier 609. The transceiver 608 is also connected to a combined modulator / demodulator 610 that connects the communication unit 602 to the processor 603.

  An example performance of one embodiment of the present invention for English and Mandarin Chinese is shown below. The test database includes words with similar pronunciations such as “cancel” and “castle”, and is composed of feature vectors of spoken words composed of 50 vocabularies. The database contains 9494 Mandarin Chinese words from approximately 200 speakers and 6827 English words from 25 speakers. These words were recorded in six different mobile environments, such as offices, cars, shopping malls, streets, etc., trying to establish a real world environment. The test results are summarized in Table 2. The monolingual result shows the recognition accuracy using the monolingual speech recognition dedicated system. The mixed language result includes the recognition accuracy using the mixed language speech recognition system 100 of the present invention.

  Accordingly, the present invention is an improved speech recognition system 100 that can recognize multilingual spoken names without requiring the user to manually switch the language mode of the system 100. Thus, it is useful, for example, in a multilingual environment where a user can have an electronic address book that includes multilingual names. Since the user does not need to switch language modes, the system 100 can recognize even a composite name composed of a first name in the first language and a second name in the second language. Also, the memory and processing requirements of the system 100 can be saved by using a combined acoustic model that includes components that depend on context and components that do not depend on context. Thus, the system 100 can be operated on a personal electronic device such as a mobile phone or PDA having limited memory and processing resources.

  The above detailed description provides only preferred exemplary embodiments and is not intended to limit the scope, applicability, or configuration of the invention. Rather, this detailed description of the preferred embodiments provides those skilled in the art with an enabling description for implementing preferred exemplary embodiments of the present invention. Obviously, various modifications may be made in the function and arrangement of elements and steps without departing from the spirit and scope of the invention as set forth in the claims.

1 is a schematic diagram illustrating functional components of a system for speech recognition of names in multiple languages, according to one embodiment of the present invention. FIG. 4 is a table showing various names of two different languages and pronunciations consisting of a series of pronunciation units in order with respect to them according to an embodiment of the present invention. The schematic which shows the operation | movement and component of a character / sound converter by the embodiment of this invention. 4 is a generalized flow chart summarizing a method for converting stored text into a pronunciation unit according to an embodiment of the invention including a Mandarin / English public vocabulary dictionary. 4 is a generalized flowchart illustrating a method for matching spoken words against names stored in a public vocabulary dictionary according to an embodiment of the present invention. 1 is a schematic diagram illustrating a personal electronic device in the form of a wireless telephone capable of executing a speech recognition system, according to an embodiment of the present invention.

Claims (7)

A speech recognition method for recognizing names in Chinese and English using a speech recognition system (100), the speech recognition method comprising:
A pronunciation unit conversion step of receiving character text representing a plurality of names that are not speech and converting the character text into a pronunciation unit;
A feature vector conversion step of receiving a spoken language that is a spoken word and converting the spoken language into a feature vector;
A collating step for recognizing speech by collating the pronunciation unit with the feature vector;
Including
The pronunciation unit conversion step includes:
A character determination step (410) for determining whether each of the names is made up of Chinese alphabet characters or Roman alphabet characters by an alphabet identifier (305) ;
A Chinese sound determination step (420) for determining whether or not the character of the Roman alphabet is a Chinese sound by means of a sound identification device (325);
A Chinese character / sound converter (315) converts the characters of the Chinese alphabet and the characters of the Roman alphabet, which is the Chinese pinto, into an ordered series of pronunciation units. A conversion step (430);
An English character / sound conversion step (435) for converting the characters of the Roman alphabet determined by the English character / sound converter (330) to be not the Chinese pimp into the pronunciation unit;
Including
The feature vector conversion step includes:
A spoken word receiving step (505) for receiving the spoken word by means of a microphone (120) ;
A conversion step (510) for converting said spoken word into a feature vector;
A speech recognition method comprising: The matching step further includes:
Comparing the feature vector with the pronunciation unit with reference to a Gaussian mixture parameter;
Analyzing the feature vector by following a backtracking pointer at the end of the search using a beam search Viterbi algorithm;
The speech recognition method according to claim 1, further comprising: The voice recognition system is a personal electronic device, and the name is part of an electronic device address book or contact list stored in the voice recognition system (100) in association with a personal phone number or other contact information. The speech recognition method according to claim 1 , wherein the personal electronic device operates when the feature vector and the pronunciation unit match in the comparison step . A speech recognition system (100) for recognizing names in Chinese and English, wherein the speech recognition system (100)
At least one of a keypad (606) and touch screen (605) to which non-speech character text is input, wherein the character text represents a plurality of names;
An alphabet classifier (305) for determining whether each of the names is made up of Chinese alphabet characters or Roman alphabet characters;
A pinyin discriminator (325) for determining whether or not the Roman alphabetic character is a Chinese pinyin;
A Chinese character / sound converter (315) for converting the characters of the Chinese alphabet and the characters of the Roman alphabet, which is the Chinese pimp, into an ordered series of pronunciation units;
An English character / sound converter (330) for converting the characters of the Roman alphabet determined not to be a Chinese pin sound into the pronunciation unit;
A microphone (120) into which the spoken language, which is the spoken word, is input;
A feature extractor (125) for converting the spoken word into a feature vector;
A speech recognition system (100), comprising: an automatic speech recognition engine (130) for collating the feature vector with the pronunciation unit. The automatic speech recognition engine (130) decodes the feature vector by comparing the feature vector with the pronunciation unit with reference to a Gaussian mixture parameter, and uses a beam search Viterbi algorithm at the end of the search. The speech recognition system (100) of claim 4, wherein the feature vector is analyzed by following a backtracking pointer . The automatic speech recognition engine (130) further includes
A public vocabulary dictionary (110) storing the pronunciation units;
A hidden Markov model set (115) including the Gaussian mixture parameters representing selected speech patterns of Chinese and English respectively;
Including
The selected Chinese voice pattern depends on the context of the pronunciation unit,
The speech recognition system (100) according to claim 4 or 5, wherein the English selected speech pattern does not depend on the context of the pronunciation units. The voice recognition system (600) according to any one of claims 4 to 6 , wherein the voice recognition system is a personal electronic device such as a mobile phone or a portable information terminal.

Images