JP2007233149A - Voice recognition device and voice recognition program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition device and a voice recognition program, for attaining voice recognition with high accuracy. <P>SOLUTION: The voice recognition device, for performing voice recognition from input voice by one or more speaker clusters in which a sound feature is different from each other, comprises: a sound analysis means for converting the input voice to a sound feature amount; an identification means for identifying attributes associated with a speaker cluster from the sound feature amount which is obtained by the sound analysis means; and a continuous voice recognition means for performing continuous voice recognition based on a search network for correct word searching, which is created from a sound model and a language model of the speaker cluster which is established beforehand, and restriction information of attributes associated with the speaker cluster to the input voice. <P>COPYRIGHT: (C)2007,JPO&INPIT

Description

  The present invention relates to a voice recognition apparatus and a voice recognition program, and more particularly to a voice recognition apparatus and a voice recognition program for realizing highly accurate voice recognition.

  In speech recognition used for subtitle production and metadata production for broadcast programs, it is important to improve speech recognition performance in which male and female speakers are mixed.

Therefore, for the speech recognition of multiple speaker clusters with different acoustic features, the method using a unique acoustic model that does not depend on gender or the like, or speaker cluster attributes such as gender in the pre-processing of continuous speech recognition are used. Methods that are determined in advance are known (see, for example, Patent Document 1 and Non-Patent Document 1). As another method, a method is known in which an acoustic model of a plurality of speaker clusters such as men and women is operated in parallel on a word dictionary, but transition between word dictionaries of different speaker clusters within one utterance is not allowed. (For example, see Patent Document 2 and Non-Patent Document 2).
JP 2003-99083 A F. Kubala, et al. "The 1996 BBN Byblos HUB-4 Transcription System", DARPA Speech Recognition Works, pp. 90-93, 1997. JP 2005-345772 A Hiroshi Yamamoto et al., “Structure and Control Method of Speech Recognition Part in Japanese-English Speech Translation System“ ATR-MATRIX ””, Proceedings of the Acoustical Society of Japan, 2-Q-21, pp. 161-162, 1998.3.

  Among the conventional speech recognition methods described above, in the case of a method using a single acoustic model that does not depend on gender, etc., a global unique acoustic model is learned in advance by ignoring all speaker clusters of the learning speech. Since speech recognition is performed using only this acoustic model, it is not necessary to consider speaker clusters at the time of recognition and can be realized very easily. However, since the recognition rate is generally low compared to a case where speaker clusters such as a gender-dependent acoustic model are considered, there is a practical problem.

  In addition, in the case of the method for predetermining the speaker cluster attributes such as gender in the above-described continuous speech recognition pre-processing, a method based on likelihood comparison of acoustic models expressing the acoustic characteristics of the entire speaker cluster, There is a method based on phoneme recognition using an acoustic model of a speaker cluster.

  Here, in the former case, the speaker cluster attribute is determined in the speech section at the beginning of each utterance. Since the speaker cluster attribute is assumed throughout one utterance, both male and female voices are mixed in one utterance. If an incorrect speaker cluster attribute is assumed due to the influence of speech such as easy conversations or background noise, the recognition rate is lowered.

  In the latter case, since the final phoneme recognition result and speaker cluster attribute are determined after all of the input speech of one utterance is obtained, speech recognition cannot be started immediately after the start of speaking, and live broadcasting is performed. There are practical problems in applications that require online processing and a small time delay, such as the production of captions for programs.

  Furthermore, in the case of the technique in which the acoustic models of a plurality of speaker clusters such as men and women described above are operated in parallel on the word dictionary but the transition between the word dictionaries of different speaker clusters within one utterance is not allowed, The speaker cluster is considered only at the beginning of each utterance, and speaker clusters that do not match the input speech are gradually excluded from the search target as the speech recognition process proceeds. There is a problem that the recognition rate is lowered in the case of voices such as conversations that can easily be mixed.

  The present invention has been made in view of the above-described problems, and an object thereof is to provide a speech recognition device and a speech recognition program for realizing speech recognition with high accuracy.

  In order to solve the above problems, the present invention employs means for solving the problems having the following characteristics.

  The invention described in claim 1 is a speech recognition apparatus that performs speech recognition by one or a plurality of speaker clusters having different acoustic characteristics from an input speech, and an acoustic analysis unit that converts the input speech into an acoustic feature amount; Speaker cluster attribute identifying means for identifying speaker cluster attributes from acoustic features obtained by the acoustic analysis means, and a search for correct word search generated from the acoustic model and language model of the speaker cluster set in advance It has a continuous speech recognition means for performing continuous speech recognition based on the network and constraint information of the speaker cluster attribute for the input speech.

  According to the first aspect of the present invention, voice recognition can be realized with high accuracy.

  The invention described in claim 2 has an acoustic model composed of one or a plurality of speaker clusters expressing acoustic features of speech, and a language model expressing transitions between preset words, The present invention is characterized by comprising network expansion means for expanding an acoustic model of a speaker cluster into a search network according to the language model and a word dictionary preset for each speaker cluster.

  According to the second aspect of the present invention, a highly accurate search network can be generated according to the content of the input voice. Moreover, it is possible to realize speech recognition with high accuracy for the input speech using this search network.

  According to a third aspect of the present invention, the network expansion means follows a transition from the recognition start state of the utterance start edge to the word dictionary start edge of all speaker clusters, and a language model between word dictionaries of the same speaker cluster. Transition from the word dictionary end of each speaker cluster to the beginning of the word dictionary of a different speaker cluster according to changes in the speaker cluster attribute, transition from the word dictionary end of each speaker cluster to the recognition end state at the end of speech And a search network that enables at least one of the transitions from the recognition end state to the recognition start state for the next utterance.

  According to the third aspect of the present invention, it is possible to improve the accuracy of the search network by performing each state transition.

  The invention described in claim 4 is characterized in that the speaker cluster attribute identification means outputs the time information when the speaker cluster changes and / or the attribute information of the speaker cluster after the change with respect to the input speech. Features.

  According to the fourth aspect of the present invention, since the change portion of the speaker cluster can be grasped before the continuous speech recognition, the speech recognition can be executed with higher accuracy.

  In the invention described in claim 5, when the speaker cluster attribute changes, the continuous speech recognition means changes the speaker cluster after the change from the word dictionary end of the speaker cluster before the change in the search network. It is possible to make a transition to the beginning of the word dictionary, and to continuously search the word dictionary of the speaker cluster before the change.

  According to the fifth aspect of the present invention, voice recognition can be performed with high accuracy for each voice of a plurality of speaker clusters having different acoustic characteristics. Thereby, even when voices of a plurality of speaker clusters are mixed during one utterance, high-accuracy voice recognition can be realized with a smaller calculation amount and a smaller delay time than in the past.

  The invention described in claim 6 is a speech recognition program for causing a computer to execute speech recognition processing for performing speech recognition by one or a plurality of speaker clusters having different acoustic characteristics from input speech. Acoustic analysis processing for converting into feature amounts, speaker cluster attribute identification processing for identifying speaker cluster attributes from the acoustic feature amounts obtained by the acoustic analysis processing, and acoustic models and language models of the speaker clusters set in advance Based on the search network for the correct word search generated from the above and the restriction information of the speaker cluster attribute for the input speech, the computer is caused to execute continuous speech recognition processing for performing continuous speech recognition.

  According to the sixth aspect of the present invention, voice recognition can be realized with high accuracy. In addition, voice recognition processing can be easily realized by installing an execution program in a computer.

  According to the present invention, speech recognition can be realized with high accuracy.

<Outline of the present invention>
The present invention enables transition between words of a plurality of speaker clusters in one utterance while using an acoustic model of one or a plurality of speaker clusters. For example, male and female voices are easily mixed in one utterance. For speech such as conversations, high-accuracy speech recognition is realized with a smaller amount of computation and a smaller delay time than in the past.

  Specifically, a search network is developed in parallel for acoustic models of a plurality of speaker clusters such as men and women, elderly people, adults and children according to a language model and a word dictionary. For example, when the speaker cluster attribute of the input speech changes from A to B, it is possible to make a transition from the end of the word cluster of the speaker cluster A to the start of the word dictionary of the speaker cluster B in the search network. By continuously searching the word dictionary of the cluster A, flexible parallel speech recognition of a plurality of speaker clusters mixed during one utterance is realized.

  Hereinafter, preferred embodiments of a speech recognition apparatus and a speech recognition program according to the present invention having the above-described features will be described in detail with reference to the drawings.

<Voice recognition device: device configuration>
FIG. 1 is a diagram showing a configuration example of a speech recognition apparatus according to the present invention. The speech recognition apparatus 10 shown in FIG. 1 is configured to include a network expansion unit 11, an acoustic analysis unit 12, a speaker cluster attribute identification unit 13, and a continuous speech recognition unit 14.

  The network expansion means 11 expands the search network 23 based on the word dictionary of each speaker cluster using the acoustic model 21 and the language model 22 of one or a plurality of speaker clusters, and continues the expanded search network 23. Output to the voice recognition means 14.

  Here, the speaker cluster acoustic model 21 is, for example, when the number of speaker clusters is 2 (A, B), the speaker cluster A is male, the speaker cluster B is female, or the speaker cluster A is adult, The speaker cluster B can be set arbitrarily, such as a child, the speaker cluster A can be a wideband voice, and the speaker cluster B can be a narrowband voice. In the present invention, the present invention is not limited to this. For example, the number of speaker clusters may be three or more by combining the above-described speaker clusters, or may be singular.

  The unit of the acoustic model can be arbitrarily set such as phonemes, syllables, environment dependence, environment independence, and the like. For example, a stochastic statistical model represented by a hidden Markov model (for example, Seiichi Nakagawa, “according to probability model”). Speech recognition ", The Institute of Electronics, Information and Communication Engineers, 1988) can be used. Moreover, the language model 22 can arbitrarily set, for example, an N-gram model in units of words. Details of the search network 23 in the network expansion means 11 will be described later.

  The acoustic analysis unit 12 converts the input speech 24 to be a speech recognition target into an acoustic feature amount 25 and outputs the converted acoustic feature amount 25 to the speaker cluster attribute identification unit 13 and the continuous speech recognition unit 14. Here, the acoustic feature quantity 25 has the same configuration as the acoustic feature quantity used for learning the acoustic model 21 of each speaker cluster. For example, a cepstrum representing frequency characteristics, short-time power, dynamic feature quantities thereof, and the like. It can be.

  Further, the speaker cluster attribute identification unit 13 obtains the speaker cluster attribute for each section based on time information and the like in the input voice online and in real time. Further, the speaker cluster attribute identification unit 13 displays speaker cluster change information such as time information when the speaker cluster changes with respect to the input speech 24 and / or attribute information of the speaker cluster after the change. 26 is output to the continuous speech recognition means 14. Thereby, since the change part of a speaker cluster can be grasped | ascertained before the continuous speech recognition, speech recognition can be performed with higher accuracy. Details of the speaker cluster attribute identification method in the speaker cluster attribute identification unit 13 will be described later.

  The continuous speech recognition means 14 receives the search network 23, the acoustic feature value 25, and the speaker cluster attribute 26, and the likelihood for the acoustic feature value 25 by an acoustic model linked to the phoneme of each word on the search network 23. Are calculated sequentially.

  The continuous speech recognition means 14 recognizes a character string of a word most suitable for the utterance content of the input speech 24 by performing linguistic scoring when the search network 23 transitions according to the language model 22. The result 27 is output.

  Here, the continuous speech recognition means 14 starts calculating the likelihood at every word start of the word dictionary of every speaker cluster at the start of the utterance. At this time, if the likelihood calculation proceeds with respect to more input speech, the continuous speech recognition means 14 gradually decreases the likelihood of the acoustic model of the words of the speaker cluster that does not match the input speech. Exclude.

  However, if the continuous speech recognition means 14 recognizes that a change has occurred in the speaker cluster within one utterance by the speaker cluster attribute 26, the continuous speech recognition means 14 allows the transition to the beginning of the word dictionary of the changed speaker cluster. All the search paths for speaker clusters remaining in the search target are also retained. Therefore, voice recognition of a plurality of speaker clusters in one utterance can be realized with high accuracy.

  In other words, in the present invention, when the speaker cluster attribute 26 is designated by the speaker cluster attribute 26, the search for other speaker clusters is not stopped, leaving only the designated word dictionary. A search target is a word dictionary of both a new speaker cluster having a certain number and a speaker cluster that is already being searched. For this reason, it is possible to consider the possibility of mixing a plurality of speaker clusters in one utterance with more detailed continuous speech recognition than speaker cluster identification while suppressing an increase in the amount of calculation.

  The above-described configuration of the speech recognition apparatus 10 enables transition between words of a plurality of speaker clusters during one utterance while using a high-accuracy acoustic model of a plurality of speaker clusters, and allows men and women during one utterance. Even in speech such as conversations that are likely to be mixed, it is possible to realize highly accurate speech recognition with a smaller amount of computation and less delay time than in the past.

  In the speech recognition apparatus 10 described above, the search network 23 is developed from the acoustic model 21 and the language model 22 of the speaker cluster by the network expansion unit 11. However, the present invention is not limited to this. The search network 23 may be stored in the continuous speech recognition means 14 or other storage means (not shown).

<Search network 23>
Here, the expansion content of the search network 23 in the network expansion means 11 is demonstrated. FIG. 2 is a diagram illustrating an example of a search network when the number of speaker clusters is two. FIG. 3 is a diagram showing an example of the internal structure of the word dictionary of each speaker cluster.

  The search network 23 output from the network expansion means 11 is linked to the recognition start state 31, the word dictionary 32 linked to the speaker cluster A acoustic model, and the speaker cluster B acoustic model as shown in FIG. The word dictionary 33 and the recognition end state 34 can be configured.

  Here, the word dictionary to be used is, for example, a network in which phonetic symbols and the like are expanded for all words in the recognized vocabulary, and the correct word search at the time of continuous speech recognition is input by an acoustic model corresponding to each phoneme. Likelihood calculation is performed on the acoustic feature quantity of speech.

  In addition, the word dictionary can be configured by a tree structure dictionary that shares the phonemes at the beginning of each word, or a linear dictionary in which the phonemes of each word are independent for each word. For example, as shown in FIG. 3, the word dictionary in the case of a tree structure dictionary develops phoneme unit phonetic symbols of each word on the network in a tree structure (for example, the word “red (/ a / k / a /)”). And the word “autumn (/ a / k / i /) share phonemes / a / and / k /), and each phoneme is linked to a corresponding acoustic model.

  Transition from the recognition start state 31 to the first phoneme of all words in all the speaker clusters 32 and 33 can be made without restriction immediately after the start of recognition of the utterance start edge (arrow * 1 shown in FIG. 2). Further, from the end of the word dictionary 32 of the speaker cluster A, a transition can be made in accordance with the language model 22 to the start of the word dictionary 32 of the same speaker cluster A in order to recognize subsequent words (shown in FIG. 2). Arrow * 2).

  In addition, a transition can be made from the end of the word dictionary 32 of the speaker cluster A to the start of the word dictionary 33 of a different speaker cluster B according to the change of the speaker cluster attribute (arrow * 3 shown in FIG. 2).

  Similarly, from the end of the word dictionary 33 of the speaker cluster B, a transition can be made in accordance with the language model 22 to the start of the word dictionary 33 of the same speaker cluster B in order to recognize subsequent words (see FIG. Arrow 2 shown in FIG. Further, from the end of the word dictionary 33 of the speaker cluster B, a transition can be made to the start of the word dictionary 32 of a different speaker cluster A according to the change of the speaker cluster attribute (arrow * 3 shown in FIG. 2). .

  At the end of the utterance, it is possible to transition from the end of the word dictionary 32, 33 of each speaker cluster to the recognition end state 34 (arrow * 4 shown in FIG. 2). Furthermore, from the recognition end state 34, a transition can be made to the recognition start state 31 in order to recognize the next utterance (arrow * 5 shown in FIG. 2). Note that the start and end of the utterance described above can be detected quickly and with high accuracy using the accumulated phoneme likelihood.

  As described above, the network expansion means 11 makes a transition from the recognition start state of the utterance start edge to the word dictionary start edge of all speaker clusters as described above, a transition according to the language model between the word dictionaries of the same speaker cluster, Transition from the word dictionary end of each speaker cluster to the beginning of the word dictionary of a different speaker cluster in response to changes in speaker cluster attributes, transition from the word dictionary end of each speaker cluster to the recognition end state at the end of speech, and By configuring a search network that enables at least one of the transitions from the recognition end state to the recognition start state for the next utterance, each state transition is performed to improve the accuracy of the search network. The search network 23 can be used to realize highly accurate speech recognition.

<Speaker cluster attribute identification method>
Next, a speaker cluster attribute identification method in the speaker cluster attribute identification unit 13 will be described. As a speaker cluster attribute identification method, for example, when the number of speaker clusters is 2 (male and female), phoneme recognition of males and females is performed in parallel by a sequential determination process (for example, JP-A-2001-92496). Subword levels of multiple speaker clusters (eg phonemes, syllables, triphones, etc.) with little delay time by using a method such as gender parallel phoneme recognition that outputs the better speech recognition result from the results (scores) determined and confirmed ) Speech recognition can be performed to identify which speaker cluster belongs to which speaker cluster.

  In addition, gender-parallel phoneme recognition is possible, which allows for transition between men and women, quickly detects the beginning and end of utterances using cumulative phoneme likelihood, and identifies speaker cluster attributes based on the results There is a technique. Here, the above-mentioned content is demonstrated using figures. In the following example, an example is shown in which phoneme recognition using a gender-dependent gender-dependent acoustic model is used as an example of a cluster attribute, and speech segment detection with a small time delay from input speech is executed.

<Speaking section detection>
FIG. 4 is a diagram illustrating an example of a gender-parallel phoneme recognition network. As shown in FIG. 4, gender-parallel phoneme recognition is possible, which allows transition between men and women, and uses the cumulative phoneme likelihood to quickly detect the beginning and end of an utterance.

Specifically, a phoneme environment-independent acoustic model learned from male speaker speech in various acoustic environments, using a cepstrum, short-time power, and dynamic features of the feature vectors of the input speech as well as a woman who learned in the same way A phoneme network as shown in FIG. 4 is constructed from the acoustic model using phoneme bigrams. Here, the non-voice model for each _g ∈ {0 (male), 1 (female)} is sill, and the other phoneme models are ph _{g, i} .

Here, the feature vector sequence of the input speech from the start detection start time τ of the utterance to the current time t is x _τ ^t , and the maximum likelihood phoneme sequence and the cumulative phoneme logarithmic likelihood of the non-speech model at the start end are respectively shown below ( 1) and (2).

ここで、発話始端では最尤音素列の累積尤度の対数値Ｌ_１と、始端の非スピーチ音響モデルの累積尤度の対数値Ｌ_２との差が一定の閾値θ_{ｓｔａｒｔ}を超えた時、すなわち（Ｌ_１−Ｌ_２）＞θ_{ｓｔａｒｔ}となる時、最尤音素列始端の非音声モデルの終端から一定の時間長ｔ_{ｓｔａｒｔ}遡った時刻を発話始端とする。

Here, the logarithmic value L ₁ of the cumulative likelihood of the maximum likelihood sequence of phonemes in speech start, when the difference between the logarithmic value L ₂ of the cumulative likelihood of the beginning of non-speech acoustic model exceeds a certain threshold theta _start, That is, when (L ₁ −L ₂ )> θ _start , a time that is a predetermined time length t _start from the end of the non-voice model at the _start of the maximum likelihood phoneme sequence is set as the _start of speech.

Note that in order to absorb a long non-speech until the utterance start end is detected, phoneme recognition is reset and the start end detection start time τ is updated when the start end detection condition is not satisfied continuously for the time length t _idle . Further, during phoneme recognition, transition between male and female phoneme models is permitted, and the penalty (weight) at that time is added to the cumulative phoneme log likelihood.

  Next, the cumulative phoneme logarithmic likelihood of the terminal non-speech model and the maximum likelihood phoneme sequence is represented by the following equations (3) and (4), respectively.

このとき、終端が非スピーチ音響モデルとなる最尤音素列のうち、最大の累積尤度の対数値Ｌ_３と、同話者クラスタのスピーチ音響モデルを終端とする最尤サブワード列の累積尤度の対数値Ｌ_４との差が一定の閾値θ_ｅｎｄを越えた状態で、更に時間長ｔ_ｅｎｄ１を継続して超えた場合、すなわちｔ_ｅｎｄ１継続して（Ｌ_３−Ｌ_４）＞θ_ｅｎｄが一定の時間長ｔ_ｅｎｄ１継続して満たされた場合、現時刻ｔから時間長ｔ_ｅｎｄ２（ｔ_ｅｎｄ２＜ｔ_ｅｎｄ１）遡った時刻を発話終端とする。

At this time, among the maximum likelihood phoneme sequences whose end is a non-speech acoustic model, the maximum likelihood likelihood logarithm L ₃ and the maximum likelihood subword sequence cumulative likelihood that ends with the speech acoustic model of the speaker cluster in a state in which the difference between the logarithmic value _{L 4} of exceeds a certain threshold theta _{end the,} if it exceeds by further continuing the time length _{t end1,} i.e. _{t end1} continued _{(L 3 -L 4)> θ} end is When the predetermined time length t _end1 is satisfied continuously, the time that is time duration t _end2 (t _end2 <t _end1 ) from the current time t is set as the speech end point.

As a result, it is possible to detect the time of the utterance start and the end of utterance and / or the cluster attribute of the section. Note that t _start and t _end2 described above can be arbitrarily set to 0 or more. Of course, in addition to this, an utterance interval detection method such as a method using voice power can be used.

  Based on the time information and attribute information of the utterance start end and utterance end obtained in this way, the above-described speaker cluster attribute 26 for the acoustic feature amount 25 can be identified. Here, FIG. 5 is a diagram illustrating an example of speaker cluster attributes.

For example, the speaker cluster attribute, as shown in FIG. 5, the column X ₁ of the acoustic feature amount X of the input _speech, ..., with respect to X _T, if the speaker cluster of men and women, from X ₁ to X _t Is made up of speaker cluster attribute information such as male, X _{t + 1} to X _k being female, and X _{k + 1} to X _T being male, and time information on each acoustic feature.

  Note that the speaker cluster attribute described above is not limited to men and women in the present invention. For example, elderly people, adults, and children may be attributed. Attribute division may be performed.

  This makes it possible to transition between words in multiple speaker clusters during one utterance while using a high-accuracy acoustic model of multiple speaker clusters. For speech such as the above, it is possible to realize highly accurate speech recognition with a smaller amount of computation and a smaller delay time than in the past.

<Execution program>
Here, the above-described speech recognition apparatus 10 can perform the speech recognition processing according to the present invention using the above-described dedicated device configuration or the like, but an execution program (speech) that can cause a computer to execute the processing in each configuration. For example, the speech recognition processing according to the present invention can be realized by installing the program into a general-purpose personal computer, server, or the like.

<Hardware configuration>
Here, an example of a hardware configuration of a computer capable of executing speech recognition processing according to the present invention will be described using the drawings. FIG. 6 is a diagram illustrating an example of a hardware configuration capable of realizing the speech recognition process according to the present invention.

  6 includes an input device 41, an output device 42, a drive device 43, an auxiliary storage device 44, a memory device 45, a CPU (Central Processing Unit) 46 for performing various controls, and a network connection device. 47, which are connected to each other by a system bus B.

  The input device 41 has a pointing device such as a keyboard and mouse operated by the user and a voice input device, and inputs various operation signals and voice signals such as a program execution instruction from the user. The output device 42 has a display for displaying various windows and data necessary for operating the computer main body for performing the processing according to the present invention, a speaker for outputting sound, and the like. And results can be displayed or output as audio.

  Here, in the present invention, the execution program installed in the computer main body is provided by, for example, the recording medium 48 such as a CD-ROM. The recording medium 48 on which the program is recorded can be set in the drive device 43, and the execution program included in the recording medium 48 is installed in the auxiliary storage device 44 from the recording medium 48 via the drive device 43.

  Further, the drive device 43 can record the execution program according to the present invention on the recording medium 48. Thereby, using the recording medium 48, it can be easily installed in a plurality of other computers, and voice recognition processing can be easily realized.

  The auxiliary storage device 44 is a storage means such as a hard disk, and can store an execution program according to the present invention, a control program provided in a computer, and the like, and can perform input / output as necessary. The auxiliary storage device 44 stores the speaker cluster acoustic model 21 and language model 22, search network 23, input speech 24, acoustic feature 25, speaker cluster attribute 26, speech recognition result 27, and the like. It can also be used as storage means.

  The CPU 46 performs various calculations and data input / output with each hardware component based on a control program such as an OS (Operating System) and an execution program read from the auxiliary storage device 44 and stored in the memory device 45. Each processing in the voice recognition processing can be realized by controlling the processing of the entire computer. Various information necessary during the execution of the program can be acquired from the auxiliary storage device 44 and can also be stored.

  The network connection device 47 obtains an execution program from another terminal connected to the communication network or executes the program by connecting to a communication network such as a telephone line or a LAN (Local Area Network) cable. The execution result obtained in this way or the execution program in the present invention can be provided to other terminals or the like.

  With the hardware configuration described above, the above-described voice recognition processing can be realized at a low cost without requiring a special device configuration. In addition, voice recognition processing can be easily realized by installing a program.

<Voice recognition processing procedure>
Next, a voice recognition processing procedure using the execution program (voice recognition program) in the present invention will be described with reference to a flowchart. FIG. 7 is a flowchart illustrating an example of a voice recognition processing procedure.

  In FIG. 7, first, immediately after the start of the program, a search network is developed using acoustic models and language models of a plurality of speaker clusters (S01). In addition, the process so far may be processed previously as pre-processing.

  Next, it is determined whether or not there is a voice input (S02). When a voice is input (YES in S02), a short interval of, for example, about 25 milliseconds necessary for calculating the acoustic feature amount for one frame. Are digitally input (S03) and acoustic analysis is performed (S04). Next, the speaker cluster attribute of each section of the input speech is identified (S05).

  Here, in the speaker cluster attribute identified in S05, it is determined whether or not the attribute has changed (S06). If the attribute has changed (YES in S06), the word of another speaker cluster that has changed is changed. Transition to the dictionary is permitted (S07). Further, when there is no attribute change (NO in S06), or after the processing of S07 is completed, the likelihood is sequentially calculated with the acoustic model of each word on the search network (S08). In the process of S08, the likelihood is calculated in consideration of the score by the language model.

  Here, it is determined whether or not the voice recognition process is to be ended (S09). If the voice recognition process is not to be ended (NO in S09), the process returns to S02 and thereafter the same process is continued and the continuous voice recognition process is executed. . In addition, when there is no voice input in the process of S02 (NO in S02), that is, when the utterance is finished, or when the voice recognition process is finished in the process of S09 (YES in S09), the word string of the recognition result Is output (S10), and the process ends. Of course, it is also possible to sequentially output a word string as a result of recognition before the end of utterance by the above-described sequential determination process.

  As described above, voice recognition can be performed quickly and with high accuracy by voice recognition processing using a voice recognition program. In addition, voice recognition processing can be easily realized by installing a program.

<Comparison example of speech recognition results>
Here, a comparative example of speech recognition results between the conventional method and the method of the present invention will be described. As an example, as a result of speech recognition such as interviews of news programs with mixed male and female voices, the word recognition rate when using the only acoustic model that does not depend on gender, which is the conventional method, is 12.2% (input speech The ratio of the recognition processing time to the time length of the above = the recognition processing real time ratio 0.81 times). In addition, when the male and female acoustic models are operated in parallel, the word error recognition rate when the transition between words of male and female during one utterance is not allowed is 11.9% (recognition processing real time ratio 0.93 times) Similarly, the word misrecognition rate when the transition between words of men and women is always allowed without restriction was 11.3% (recognition processing real time ratio 1.28 times).

  On the other hand, the word misrecognition rate when the transition between the male and female words according to the present invention is permitted / not permitted according to the gender attribute constraint is 11.1%, and the recognition processing real time ratio is 0. The effect of the present invention was shown especially in recognition rate and processing time.

  In other words, in recognition of gender speech, the gender-dependent acoustic model is operated in parallel with gender rather than gender-independent acoustic models, and the transition between men and women during utterance is allowed, and the misrecognition rate is lower. The error recognition rate and the recognition time were improved due to the restriction of.

  As described above, according to the present invention, speech recognition can be realized with high accuracy. Specifically, the present invention expands an acoustic model of a plurality of speaker clusters into a search network in parallel according to a language model and a word dictionary, and uses the restriction of speaker cluster attributes of the input speech during one utterance. By enabling the transition between words of different speaker clusters at the same time, even when voices of multiple speaker clusters are mixed during one utterance, high-accuracy speech with less computation and less delay time than before Recognition can be realized.

  As a result, for example, a plurality of acoustic features of a speaker's voice are assumed, such as men and women, elderly people, adults and children, etc. The present invention can be applied in a situation where voices with special features are mixed in one utterance, or in voice recognition that requires online processing such as caption production of a live broadcast program and a small time delay.

  In addition, the present invention is applied to technologies in various fields using speech recognition and language processing, such as subtitle production for broadcast programs, voice dialogue systems, voice word processors, automatic creation of meeting minutes, and control of devices by voice. be able to.

  The preferred embodiments of the present invention have been described in detail above, but the present invention is not limited to such specific embodiments, and various modifications, within the scope of the gist of the present invention described in the claims, It can be changed.

It is a figure which shows the example of 1 structure of the speech recognition apparatus in this invention. It is a figure which shows an example of the search network in case the number of speaker clusters is two. It is a figure which shows an example of the internal structure of the word dictionary of each speaker cluster. It is a figure which shows an example of the network of gender parallel phoneme recognition. It is a figure which shows an example of a speaker cluster attribute. It is a figure which shows an example of the hardware constitutions which can implement | achieve the speech recognition process in this invention. It is a flowchart which shows an example of a speech recognition process sequence.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Speech recognition apparatus 11 Network expansion means 12 Acoustic analysis means 13 Speaker cluster attribute identification means 14 Continuous speech recognition means 21 Acoustic model 22 Language model 23 Search network 24 Input speech 25 Acoustic feature 26 Speaker cluster attribute 27 Speech recognition result 31 Recognition start state 32, 33 Word dictionary 34 Recognition end state 41 Input device 42 Output device 43 Drive device 44 Auxiliary storage device 45 Memory device 46 CPU
47 Network connection device 48 Recording medium

Claims (6)

In a speech recognition apparatus that performs speech recognition by one or a plurality of speaker clusters having different acoustic characteristics from input speech,
Acoustic analysis means for converting the input speech into acoustic features;
Speaker cluster attribute identifying means for identifying a speaker cluster attribute from an acoustic feature obtained by the acoustic analysis means;
Continuous speech recognition based on a search network for searching for correct words generated from the acoustic model and language model of the speaker cluster set in advance, and constraint information of the speaker cluster attribute for the input speech A speech recognition apparatus comprising speech recognition means. An acoustic model composed of one or a plurality of speaker clusters expressing the acoustic features of speech, and a language model expressing transitions between preset words,
The speech recognition apparatus according to claim 1, further comprising a network expansion unit that expands an acoustic model of the speaker cluster to a search network according to the language model and a word dictionary set in advance for each speaker cluster. . The network deployment means includes
Transition from the recognition start state of the utterance start edge to the word dictionary start edge of all speaker clusters, transition according to the language model between word dictionaries of the same speaker cluster, change of each speaker cluster according to change of speaker cluster attribute Transition from the end of the word dictionary to the beginning of the word dictionary of a different speaker cluster, transition from the end of the word dictionary to the recognition end state of each speaker cluster at the end of speech, and from the recognition end state to the recognition start state for the next utterance The speech recognition apparatus according to claim 2, comprising a search network that enables at least one of the transitions. The speaker cluster attribute identification means includes:
4. The speech recognition apparatus according to claim 1, wherein time information when a speaker cluster changes and / or attribute information of the speaker cluster after the change is output with respect to the input speech. 5. . The continuous speech recognition means includes
When the speaker cluster attribute changes, it enables transition from the word dictionary end of the speaker cluster before the change in the search network to the word dictionary start of the speaker cluster after the change, and the speaker before the change 5. The speech recognition apparatus according to claim 2, wherein the cluster word dictionary is continuously searched. In a speech recognition program for causing a computer to perform speech recognition processing for performing speech recognition by one or a plurality of speaker clusters having different acoustic characteristics from input speech,
An acoustic analysis process for converting the input speech into acoustic features;
Speaker cluster attribute identification processing for identifying speaker cluster attributes from acoustic features obtained by the acoustic analysis processing;
Continuous speech recognition based on a search network for searching for correct words generated from the acoustic model and language model of the speaker cluster set in advance, and constraint information of the speaker cluster attribute for the input speech A speech recognition program for causing a computer to execute speech recognition processing.

Images