JP2010197644A - Speech recognition system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition system which does not require registering work for utilizing an enroll function before speech recognition, and which significantly improves a recognition rate of speech recognition without causing incorrect recognition and malfunction. <P>SOLUTION: A recognition computer 2 in the recognition system 1 includes: a speech information acquiring means 8 for detecting speech V which is output by a speaker S, and for acquiring speech information 17; an utterance tendency specifying means 9 for analyzing and specifying utterance tendency regarding the speech V, based on the speech information 17; a dictionary group storage means 10 for storing a reference dictionary SD and a plurality of utterance tendency dictionaries X1 etc.; a dictionary selection means 11 for selecting one utterance tendency dictionary X1 etc. which matches utterance tendency; a reference collating means 12 for collating a vocabulary by using the reference dictionary SD; an utterance tendency collating means 13 for collating the vocabulary by using the utterance tendency dictionary X1 etc.; and a vocabulary output means 14 for outputting the vocabulary regarding the recognized speech V. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a speech recognition system, and in particular, can be used for man-machine interface technology related to speech input used in medical education simulators, etc., and improves the recognition rate of speech uttered by a speaker. The present invention relates to a speech recognition system that can be made to operate.

  2. Description of the Related Art Conventionally, development of speech recognition technology that recognizes speech by receiving speech uttered by a speaker with a speech input device such as a microphone, acquiring this as speech information, and analyzing the speech information has been in progress. . This technology makes it possible to operate a computer or the like only by voice input without using an operation input device such as a keyboard or a mouse, and it is necessary to limit the movement of elderly people unaccustomed to computer operation and limbs. Even a caregiver can easily operate a computer or the like. In addition, by adopting voice input and voice recognition technology for car navigation systems, etc., it is possible to perform destination setting and start of guidance by voice without taking the driver's hand off the steering wheel. Can be increased. As described above, the technology related to speech input and speech recognition is used in a wide range of fields, and is expected to be used in new technical fields.

  In the technology related to voice recognition, if the voice uttered by the speaker cannot be accurately recognized, it may lead to malfunctions such as malfunction or error of the device. Therefore, in the speech recognition technology, how accurately the acquired speech information can be recognized without misrecognizing becomes a particularly important issue in terms of practical use. Here, when recognition by speech input is performed, a recognition dictionary (database) in which phonemes, frequency characteristics, and the like are registered in advance is used for each vocabulary included in the speech. At this time, since the products for speech recognition on the market are intended for a wide range of unspecified users, they are uttered with accents and pronunciations of standard pronunciation tendencies (so-called “standard words” and “common words”). It is set so that the vocabulary can be properly recognized.

  Therefore, a speaker who has a habit of speaking very quickly or conversely, a speaker with extremely different voice quality (high or low tone), or a speaker whose accent or pronunciation differs from the standard word due to local speaking In the speech recognition technology (speech recognition software) set as a standard word, the above-described recognition rate may be significantly reduced. For this reason, stable speech recognition cannot be performed for the above-described speakers, and the text creation work or the like may be complicated compared to input using a keyboard or the like.

  Therefore, in order to improve the recognition rate even for speaker utterances other than the standard pronunciation tendency, a specific speaker's voice is registered in advance, analyzed, and made into a database. Yes. More specifically, it is accompanied by a function (so-called “enroll function”) that drastically improves the recognition rate by allowing a speaker to read a predetermined sentence aloud in advance and acquiring and registering it as voice information. Voice recognition software has been developed. In this case, an acoustic model based on a speaker's pronunciation tendency is constructed by a speech recognition device such as a computer on which speech recognition software is operated, and when the speaker utters, a specific acoustic model constructed individually is The recognition process is implemented using this. As a result, the recognition rate is maintained at a certain level or more, and is improved to the extent that there is no practical problem.

  However, the voice recognition technology that employs the enrollment function described above may have the following problems. That is, in these voice recognition systems, the registration work in the previous stage for improving the recognition rate may be very complicated. That is, it is necessary to analyze the pronunciation tendency of a specific speaker based on statistical processing, and it is necessary to force a task of reading a huge amount of sentences aloud at the time of registration. For example, in the case of commercially available voice recognition software, when registering to use the enrollment function, the speaker reads a plurality of sentences specified by the software, and the contents are analyzed by the voice recognition software and registered as data It was necessary to execute repeatedly. Therefore, at least 30 minutes or more may be required to finally complete registration. Therefore, there are many cases where it is useless to spend time on the complicated registration work as described above, except for a specialized operator who constantly uses a system using voice recognition technology, and the enrollment function itself is fully utilized. In many cases it was not. As a result, since the speech recognition process is performed without using the enrollment function, the software may be used at a low recognition rate.

  In addition, the speech recognition analyzed by the speech recognition technology is the frequency characteristics of the speech uttered by the speaker, the strength of the sound, etc., and the pronunciation tendency is not always constant. In other words, depending on the emotion of the speaker (feeling emotional), even if the speaker utters voices with accents that are different from the normal pronunciation tendency, etc. There were times when I was uttering voice. Therefore, when registering the enrollment function, even if the speaker is registered in a standard language in a calm state, the pronunciation tendency may change due to emotions, etc. when the actual voice input is performed, and the enrollment function may not be fully demonstrated. there were. As a result, the recognition rate may be lowered.

  Therefore, in view of the above circumstances, the present invention does not require registration work for utilizing the enrollment function before speech recognition, identifies an appropriate pronunciation tendency dictionary according to each situation, and recognizes speech recognition. An object of the present invention is to provide a speech recognition system capable of dramatically improving the rate.

  In order to solve the above-described problem, the speech recognition system of the present invention includes: “a speech information acquisition unit that detects a speech of a speaker and acquires speech information; and the level of the speech based on the acquired speech information, Analyzing and specifying the speaker's pronunciation tendency including frequency characteristics, accent, and pitch, pronunciation tendency identifying means for identifying, standard dictionary constructed based on standardized standard pronunciation tendency, and identification different from the standard pronunciation tendency A dictionary group storage means for storing a plurality of pronunciation tendency dictionaries each constructed based on a pronunciation tendency, and matches or resembles the pronunciation tendency of the voice specified from the plurality of stored pronunciation tendency dictionaries A dictionary selection unit that selects one pronunciation tendency dictionary, a standard collation unit that collates and recognizes a vocabulary included in the speech using the speech information and the standard dictionary, and the standard The vocabulary determined to be unrecognized by the collating means is collated by using the pronunciation tendency dictionary selected by the dictionary selecting means, and recognized pronunciation tendency collating means, the standard collating means and the pronunciation tendency collating means And a vocabulary output means for outputting the vocabulary related to the speech recognized by at least one of the voices ”.

  Here, the pronunciation tendency relates to the voice uttered by the speaker and can be distinguished from other voices such as the pitch of the voice, the frequency distribution of the acoustic characteristics, the accent / intonation, and the pitch. In addition, in order to specify the pronunciation tendency, it may include discrimination based on voice quality, tone of voice, and emotion (feeling of emotion) when the speaker speaks. Such pronunciation tendency may be quantified by frequency distribution or the like.

  On the other hand, the standard dictionary stored in the dictionary group storage means is defined based on the standard pronunciation tendency, that is, the vocabulary accent and average voice quality and pitch in the case of a standard word (common word). Therefore, it is possible to perform speech recognition for a wide range of people without specifying a speaker. On the other hand, the pronunciation tendency dictionary includes a plurality of types that deviate from the standard pronunciation tendency of the standard dictionary and show different pronunciation tendency depending on the speaker. More specifically, it can be set individually according to various types such as the level of voice, accent differences due to regional differences such as utterances and dialects, voice speed differences, and age groups. That is, a pronunciation tendency dictionary is constructed according to the classified pronunciation tendency differences. A dictionary group is constituted by these plural dictionaries and is stored in the dictionary group storage means.

  Therefore, according to the voice recognition system of the present invention, the pronunciation tendency of the speaker is specified based on the acquired voice information, and a pronunciation tendency dictionary that matches or resembles the specified pronunciation tendency is selected from the dictionary group storage means. Selected. When performing speech recognition processing, first, vocabulary is collated using a standard dictionary based on standard pronunciation tendency, and recognition processing is performed. After that, the vocabulary that has not been recognized in the standard dictionary is collated using the selected pronunciation tendency dictionary and recognized. As a result, the collation result using the two dictionaries is output by the vocabulary output means. In other words, in the case of the speech recognition system of the present invention, since two-stage recognition processing is performed using two dictionaries, the speech recognition rate can be improved. In particular, the recognition rate can be dramatically improved because the recognition process is performed after grasping the habit of the speaker's utterance by the pronunciation tendency dictionary specified based on the pronunciation tendency.

  Further, the speech recognition system according to the present invention may include, in addition to the above configuration, “the pronunciation tendency specifying means includes earliest specifying means for specifying the pronunciation tendency based on the voice uttered by the speaker at the earliest timing”. The dictionary selecting means further comprises a fixed selecting means for fixing the pronunciation tendency dictionary corresponding to the speaker based on the pronunciation tendency specified by the earliest specifying means. It doesn't matter.

  Therefore, according to the speech recognition system of the present invention, the pronunciation tendency is specified by the voice uttered by the speaker at the earliest timing, and then the recognition process is performed with the pronunciation tendency dictionary fixed for the speaker. Is called. That is, it becomes possible to identify the speaker's pronunciation tendency by recognizing the first sentence. As a result, the identification of the pronunciation tendency is completed once, and no burden is imposed on the subsequent system.

  Furthermore, the speech recognition system of the present invention has, in addition to the above configuration, “the pronunciation tendency specifying means further includes sequential specifying means for sequentially specifying the pronunciation tendency every time the speaker utters the voice, The dictionary selecting means may comprise a dictionary reselecting means for reselecting the pronunciation tendency dictionary based on the pronunciation tendency sequentially specified.

  Therefore, according to the speech recognition system of the present invention, as described above, the pronunciation tendency dictionary and the pronunciation tendency dictionary are selected for each voice uttered, while the pronunciation tendency dictionary for the speaker is fixed. Is called. As a result, even if the speaker is the same speaker, his / her voice may change quickly or the tone of the voice may differ. In such a case, if the pronunciation tendency dictionary is specified first, the recognition rate by the pronunciation tendency dictionary may be reduced. Therefore, although a slight burden is imposed on the system, it is possible to prevent the recognition rate from being lowered by appropriately specifying the utterance tendency and selecting the dictionary.

  Further, the voice recognition system according to the present invention, in addition to the above configuration, “the voice information acquisition means acquires the voice information in a conversation format formed by mixing a plurality of voices uttered by the speaker. It does not matter.

  Therefore, the speech recognition system of the present invention acquires speech information related to conversational speech in which a plurality of speakers gather and have a conversation. Thereby, it is possible to identify a pronunciation tendency for each speaker participating in the conversation and perform a speech recognition process based on the pronunciation tendency.

  As an effect of the present invention, the recognition rate related to speech recognition can be improved by performing speech recognition processing in two stages: a standard dictionary and a pronunciation tendency dictionary. Furthermore, by specifying the pronunciation tendency every time the speaker utters the voice, the voice recognition process can be performed stably without being influenced by the emotion of the speaker.

It is explanatory drawing which shows schematic structure of the speech recognition system of this embodiment. It is a block diagram which shows the functional structure of the recognition computer in a speech recognition system. It is a flowchart which shows the flow of a process of a recognition computer. It is a flowchart which shows the flow of a process of a recognition computer.

  A speech recognition system 1 (hereinafter simply referred to as “recognition system 1”) according to an embodiment of the present invention will be described below with reference to FIGS. Here, FIG. 1 is an explanatory diagram showing a schematic configuration of the speech recognition system 1 of the present embodiment, and FIG. 2 is a block diagram showing a functional configuration of the recognition computer 2 in the speech recognition system 1, and FIGS. 4 is a flowchart showing a processing flow of the recognition computer.

  The recognition system 1 of the present embodiment is mainly configured by a recognition computer 2 that functions as a voice recognition device, as shown in FIGS. Here, as shown in FIGS. 1 and 2, the recognition computer 2 is mainly constructed using a commercially available personal computer, and a voice input device 3 such as a microphone that acquires the voice V uttered by the speaker S; An operation input device 4 such as a keyboard for inputting and operating various data, and a liquid crystal display 6 having an output screen 5 for outputting characters of recognition results are respectively connected to a computer main body 7 and mainly configured. ing.

  The computer main body 7 includes a voice recognition function SR for analyzing and recognizing the voice V and the recognized voice V in the voice V using various dictionaries (standard dictionary SD, pronunciation tendency dictionary X1, etc.). A vocabulary collation function VC for collating vocabulary is established, and various functions can be exhibited. Furthermore, the computer main body 7 has an interface and control mechanism for transmitting and receiving signals to and from the devices 3, 4, 6 and the like described above, a communication function that enables connection to a network environment such as the Internet, and a recognition system. Arithmetic processing including a storage means 19 (dictionary group storage means 10 and the like) such as a hard disk with built-in recognition system software (not shown) for functioning as 1, and a CPU for performing various processes based on the recognition system software It is comprised including a part etc. Since the configurations and functions of these personal computers are already well known, description thereof is omitted here.

  Here, as a functional configuration, the recognition computer 2 detects the voice V uttered by the speaker S via the voice input device 3, converts the voice signal related to the voice V into an electrical signal, and converts this into voice information. The voice information acquisition means 8 acquired as 17 and the pronunciation tendency relating to the voice V based on the acquired voice information 17 are analyzed, and the pronunciation tendency specifying means 9 for specifying this, and the pronunciation of a standard word or the like in advance Register each vocabulary based on the standardized standard pronunciation tendency, register each vocabulary according to the voice quality, accent, pitch, and accent, etc. A plurality of pronunciation tendency dictionaries X1, X2, X3. . . , Xn, a dictionary group (dictionary group) that is converted into electronic data and stored as a database, and voices obtained from a plurality of pronunciation tendency dictionaries X1 stored in the dictionary group storage unit 10 The dictionary selection means 11 for selecting the pronunciation tendency dictionary X1 or the like that matches or most closely matches various parameters such as the pronunciation tendency specified by the information 17 and the frequency characteristics, and the acquired voice information 17 and the dictionary group storage means 11 Using the standard dictionary SD, the standard collation unit 12 that collates and recognizes the vocabulary contained in the voice V and the vocabulary (unrecognized vocabulary) that is not recognized by the standard collation unit 12 is selected based on the pronunciation tendency. The pronunciation tendency collation means 13 that collates and recognizes using the pronunciation tendency dictionary X1 etc. and the voice V recognized by the standard dictionary SD and the pronunciation tendency dictionary X1 etc. And lexical output means 14 for outputting the vocabulary respectively are composed mainly provided.

  Here, in the recognition system 1 of the present embodiment, it is possible to adopt two methods for specifying the pronunciation tendency and subsequently selecting the pronunciation tendency dictionary X1 and the like, and it can be designated by the selection of the speaker S. It has become. More specifically, as a partial function of the pronunciation tendency specifying means 9, the earliest specifying means for specifying the pronunciation tendency based on the voice V uttered by the speaker S at the first timing in the recognition system 1 of the present embodiment. As a partial function of the dictionary selecting means 11, the pronunciation tendency dictionary X1 corresponding to the speaker S is fixed based on the pronunciation tendency specified by the earliest specifying means 15, and the subsequent voice V is also fixed. As a selection dictionary fixing method using fixed selection means 16 that performs recognition processing of voice V using the same pronunciation tendency dictionary X1 and a partial function of pronunciation tendency specifying means 9, every time a speaker utters voice V In addition, as a function of the sequential specifying means 17 for sequentially specifying the pronunciation tendency and the dictionary selecting means 11, the pronunciation tendency dictionary X1 and the like are reselected based on the pronunciation tendency sequentially specified by the sequential specifying means 17. Are two methods of selecting a dictionary change method comprising the that dictionary reselection unit 18.

  In the case of the selection dictionary fixing method, the pronunciation tendency is specified and the pronunciation tendency dictionary X1 is selected based on the earliest voice V for the speaker S, and the process of the speech recognition processing in the recognition system 1 is the pronunciation tendency dictionary. X1 etc. are always fixed. Therefore, since the process related to the pronunciation tendency specification and the selection of the pronunciation tendency dictionary X1 is completed at once, the subsequent voice recognition process can be performed promptly. For this reason, the recognition system 1 during voice recognition has an excellent advantage that it does not impose an excessive burden. However, there is a case where the pronunciation tendency of the same speaker S is deflected due to a change in the emotion of the speaker S, a quick mouth, or a loud voice. In that case, the pronunciation tendency dictionary X1 once fixed may make incorrect recognition or vocabulary collation impossible, which may reduce the recognition rate.

  On the other hand, in the case of the selection dictionary variation method, the pronunciation tendency is specified for each utterance, and the process of repeatedly selecting the pronunciation tendency dictionary X1 and the like is performed. Can respond. Further, even when a plurality of speakers S are gathered and recognize the speech V such as a conference format in which conversations are performed at a time, it is not necessary to identify each speaker S, and the pronunciation tendency of the speech V itself is increased. Since the recognition process may be performed based on this, it has an advantage in this respect compared to the selected dictionary fixing method. However, since the pronunciation tendency is specified and the pronunciation tendency dictionary X1 is selected for each utterance, there is a possibility of imposing an excessive burden on the recognition system 1. In particular, when a plurality of speakers S speak at the same time, it is necessary to perform processing such as selection for each utterance (speech V), which may eventually delay the output of the vocabulary. Therefore, the speaker S can arbitrarily select which method to use in accordance with the surrounding situation in which speech recognition is performed.

  Next, an example of the flow of processing related to speech recognition by the recognition system 1 of the present embodiment will be described based on FIGS. 3 and 4. First, before performing speech recognition, an input of a selection instruction related to the selection method such as the pronunciation tendency dictionary X1 is received (step S1). The input of this selection instruction is determined by the speaker S in consideration of the merits and demerits described above. When selection of the selection method by the speaker S is determined and an input of such a selection instruction is accepted (step S1), a flag F is set so that the process can be identified for each selection method (step S2). Specifically, in the recognition system 1 of the present embodiment, F = 1 is given when the selection dictionary fixing method is instructed, and F ≠ 1 when the selection dictionary variation method is instructed. Is performed (step S2). The flag F can be used for determination in a later step.

  Thereafter, the speaker S performs a detection process as to whether or not the voice V that is the target of voice recognition is being emitted (step S3). At this time, when the voice V for voice recognition is detected via the voice input device 3 by the speech of the speaker S (YES in step S3), the recognition computer 2 of the recognition system 1 passes through the voice input device 3. This is input, the voice signal of voice V is converted into an electrical signal based on frequency characteristics, waveform information, etc., and this is acquired as voice information 17 (step S4). Here, since the process of converting the voice V into an electrical signal and acquiring the voice information 17 as electronic data is a well-known technique, the description thereof is omitted here. As a result, the voice information 17 is stored in the storage unit 19 of the recognition computer 2. On the other hand, when the voice V is not detected by the voice input device 3 (NO in step S3), the process of step S3 is continuously repeated, the voice V is detected, and the process waits until the voice information 17 is acquired.

  Then, the pronunciation tendency of the voice V (speaker S) is analyzed and specified from the acquired voice information 17 (step S5). Here, the pronunciation tendency is specified by characteristics of the voice V obtained by analyzing the voice information 17 (voice quality of the speaker S, level of the voice V, frequency characteristics of the voice V, frequency distribution, voiceprint, pronunciation, accent, Intonation, pitch of voice V and utterance speed). That is, the pronunciation tendency varies depending on each speaker S. Therefore, the speakers S can be individually identified by the difference in the waveform of the voice V (so-called “voiceprint”). Further, the pronunciation tendency may vary depending on the emotion of the speaker S when speaking. Here, an acoustic model SM for specifying the pronunciation tendency of the speaker S is stored in the storage means 19 in advance, and the characteristics of the voice V detected by using this are compared to specify the pronunciation tendency. Will be. The specified pronunciation tendency may be stored in the storage unit 19 as electronic data (not shown). Further, after the selection of the pronunciation tendency dictionary X1, which will be described later, is basically unnecessary, it may be temporarily stored in a memory or the like and deleted after selection.

  After that, the pronunciation specified by the acoustic model SM is selected from a plurality of pronunciation tendency dictionaries X1 having different pronunciation tendency stored in the dictionary group storage means 10 of the storage means 19 as a database. One pronunciation tendency dictionary X1 or the like that matches or is most similar to the tendency is selected (step S6). In this case, the degree of approximation of the pronunciation tendency is shown in numerical form, and a process of selecting the most similar according to the value of the degree of approximation is performed.

  Next, using the standard dictionary SD of the dictionary group storage means 10, the vocabulary contained in the acquired speech information 17 is collated, and speech recognition processing is performed (step S7). Here, the voice V of the voice information 17 is composed of a plurality of vocabularies, and each voice vocabulary is collated with a vocabulary registered in advance in the standard dictionary SD by using a voice recognition technique, whereby the voice V is recorded. Can be recognized. When recognition of all vocabularies constituting the voice V by the standard dictionary SD is completed (YES in step S8), that is, all the voices V uttered by the speaker S are pronounced with a standard pronunciation tendency, and the standard dictionary SD Is recognized, the recognized vocabulary is output as character information to the output screen 5 of the liquid crystal display 6 (step S9). Thereby, the voice V is recognized as voice and converted into character information. Such character information may be stored in the storage means 19 (not shown).

  Here, when the recognition of all vocabularies is not completed by the standard dictionary SD (NO in step S8), that is, when a part (or all) of the standard dictionary SD is not collated and cannot be recognized, Using the selected pronunciation tendency dictionary X1 or the like, the vocabulary of the voice V that has not been recognized by the standard dictionary SD is collated (step S10). Then, when the recognition of all vocabularies constituting the speech V is completed by the selected pronunciation tendency dictionary X1 or the like (in this case, the result of collation and recognition by the previously used standard dictionary SD is included) (YES in step S11) ), The process proceeds to step S9, and the recognized vocabulary is output as character information. On the other hand, if the recognition of all vocabularies has not been completed (NO in step S11), character information including an unrecognized part is output to the output screen 5 (step S12). Similarly, the character information may be stored in the storage unit 19.

  Thereafter, the recognition computer 2 detects whether a new voice V is emitted from the speaker S (step S13). Here, when the voice V is detected via the voice input device 3 (YES in step S13), the value of the flag F previously specified is determined (step S14). That is, in the case of F = 1 (YES in step S14), that is, in the case of the selected dictionary fixing method in which the voice recognition process is performed for the new voice V using the previously selected pronunciation tendency dictionary X1 or the like, The voice information 17 for the correct voice V is acquired (step S15), and the process proceeds to step 7. Thereby, the voice recognition process using the standard dictionary SD and the fixed pronunciation tendency dictionary X1 is performed on the voice information 17, and the processes from step S7 to step S13 are repeated.

  On the other hand, if F ≠ 1 (NO in step S14), that is, if the pronunciation tendency dictionary X1 or the like is selected for each detected voice V, the process proceeds to step S4, and the pronunciation tendency based on the voice V is changed. The voice recognition processing using the specification, selection of the pronunciation tendency dictionary X1, etc., the standard dictionary SD, the reselected pronunciation tendency dictionary X1, etc. is performed, and the processing from step S4 to step S13 is repeated.

  Thereby, when the speech by the speaker S is continuously performed, the recognition computer 2 detects the respective voices V, and uses the two types of dictionaries SD, X1, etc., and performs the voice V Can be converted into character information and output. Furthermore, since the selection method of the pronunciation tendency dictionary X1 and the like can be selected before the recognition starts, the optimum one can be selected according to the voice V that is the target of voice recognition, and the recognition rate can be improved.

  Here, when a new voice V is not emitted from the speaker S and is not detected by the voice input device 3 (NO in step S13), the processes in steps S14 and S15 are canceled. Thereafter, the presence / absence of an instruction to terminate the system is detected (step S16). If the instruction is detected (YES in step S16), the system is terminated (step S17). On the other hand, when the instruction is not detected (NO in step S16), the process returns to step S1, and the speech recognition process is continued.

  The recognition system 1 of the present embodiment stores a pronunciation tendency dictionary X1 or the like that is previously constructed according to a plurality of pronunciation trends, and an optimal dictionary is selected based on the pronunciation tendency of the speaker S. Furthermore, by performing collation processing using the standard dictionary SD constructed according to the standard pronunciation tendency as the first processing, it is possible to reduce vocabulary collation processing using the pronunciation tendency dictionary X1 and the like. That is, the standard dictionary SD is common in recognizing sounds related to general utterances, and even if it is a voice having a different pronunciation tendency, it can be applied to most of them. It is formed so that vocabulary can be recognized. Therefore, first, most of the vocabulary in the voice V is recognized by the processing using the standard dictionary SD, and the pronunciation tendency dictionary X1 is used only for the remaining vocabulary (unrecognized vocabulary) that could not be recognized by the standard dictionary SD. Is done. Thereby, the recognition rate concerning the voice recognition of the voice V becomes stable. Although it is possible to apply the pronunciation tendency dictionary X1 etc. suddenly without using the standard dictionary SD, there is a possibility that the recognition of the part having the standard pronunciation tendency is inferior, and the voice V is completely converted. Therefore, it may be difficult to output as character information.

  The present invention has been described with reference to preferred embodiments. However, the present invention is not limited to these embodiments, and various modifications can be made without departing from the spirit of the present invention as described below. And design changes are possible.

  That is, in the recognition system 1 of the present embodiment, the one that recognizes the voice V of one speaker S is shown. However, the present invention is not limited to this. ) May be targeted. In this case, if the pronunciation tendency of each speaker S is specified in advance and the speaker S can be specified, a selection dictionary fixing method for fixing the above-described pronunciation tendency dictionary X1 or the like may be adopted. it can. However, since there is a possibility that many speakers S speak at a time, the utterances may overlap each other, and the possibility of erroneous recognition increases. For this reason, it is considered preferable to specify a pronunciation tendency for each individual utterance and adopt a selection dictionary variation method that varies the pronunciation tendency dictionary X1 and the like accordingly.

  Furthermore, in the recognition system 1 of the present embodiment, the one having a process for selecting a method related to selection of the pronunciation tendency dictionary X1 or the like is shown, but the present invention is not limited to this, and is limited to one of the methods. But of course it does n’t matter. That is, when the target of speech recognition is fixed and the variation in the pronunciation tendency of the speaker S is not expected, a selection dictionary fixing method may be adopted to perform quick and stable speech recognition.

1 Recognition system (voice recognition system)
2 recognition computer 3 voice input device 5 output screen 6 liquid crystal display 7 computer main body 8 voice information acquisition means 9 pronunciation tendency identification means 10 dictionary group storage means 11 dictionary selection means 12 standard collation means 13 pronunciation tendency collation means 14 vocabulary output means 15 Point identification means 16 Fixed selection means 17 Sequential identification means 18 Dictionary reselection means S Speaker SD Standard dictionary V Speech X1, X2, X3. . . Xn Pronunciation Trend Dictionary

Claims (4)

Voice information acquisition means for detecting the voice of the speaker and acquiring voice information;
Based on the acquired voice information, the pronunciation tendency specifying means for analyzing and specifying the speaker's pronunciation tendency including the height, frequency characteristics, accent, and pitch of the voice;
A dictionary group storage means for storing a standard dictionary constructed based on a standardized pronunciation tendency and a plurality of pronunciation tendency dictionaries each constructed based on a specific pronunciation tendency different from the standard pronunciation tendency;
A dictionary selection means for selecting one of the pronunciation tendency dictionaries that matches or resembles the pronunciation tendency of the specified voice from among the plurality of pronunciation tendency dictionaries stored;
Standard collating means for collating and recognizing vocabulary contained in the voice using the voice information and the standard dictionary;
Using the pronunciation tendency dictionary selected by the dictionary selecting means to collate and recognize the vocabulary determined to be unrecognized by the standard collating means;
A speech recognition system comprising: vocabulary output means for outputting the vocabulary related to the speech recognized by at least one of the standard collating means and the pronunciation tendency collating means. The pronunciation tendency specifying means is:
It further has a first point specifying means for specifying the pronunciation tendency based on the voice uttered by the speaker at the earliest timing,
The dictionary selecting means includes
The speech recognition system according to claim 1, further comprising: a fixed selection unit that fixes the pronunciation tendency dictionary corresponding to the speaker based on the pronunciation tendency specified by the earliest specifying unit. . The pronunciation tendency specifying means is:
Each time the speaker utters the voice, the speaker further includes sequential specifying means for sequentially specifying the pronunciation tendency,
The dictionary selecting means includes
The speech recognition system according to claim 1, further comprising a dictionary reselecting unit that reselects the pronunciation tendency dictionary based on the pronunciation tendency specified sequentially. The voice information acquisition means includes
The voice recognition system according to any one of claims 1 to 3, wherein the voice information in a conversation format formed by mixing a plurality of the voices uttered by the speaker is acquired. .

Images