JP2014164261A - Information processor and information processing method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To appropriately reject vocalization whose only pronunciation correction parts are different, and whose other parts are coincident or similar in pronunciation correction learning.SOLUTION: A grammar creation part 107 creates first to third phoneme sequence lists each including at least one phoneme sequence and including mutually different phoneme sequences. A voice recognition part 103 voice-recognizes which of the first to third phoneme sequence lists includes the phoneme sequence to which voice data to be input are equivalent. The second phoneme sequence list includes the phoneme sequence in which the predetermined subsequence of the phoneme sequence included in the first phoneme sequence list is replaced with a different phoneme subsequence. The third phoneme sequence list includes the phoneme sequence in which the replaced subsequence in the second phoneme sequence list is replaced with a further different phoneme subsequence.

Description

  The present invention relates to information processing for evaluating the accuracy of pronunciation of a user.

  In foreign language learning, pronunciation correction learning, which learns utterance with accurate pronunciation, is one of the important exercises. However, it is difficult for beginners to evaluate their pronunciation by themselves, and it is necessary to ask a specialized language teacher. For example, Patent Document 1 proposes a method of replacing this with a mechanical evaluation using a voice recognition technique.

  The method disclosed in Patent Document 1 performs two-word speech recognition with a word in which a pronounced pronunciation of a word is replaced with a target phoneme, and evaluates the pronunciation based on whether or not the correct pronunciation is recognized.

  Here, not only valid utterances are necessarily input to the speech recognition apparatus, but there is a possibility that user fillers, irrelevant utterances, and sudden noises may be input. Presenting the evaluation result for such an irrelevant sound causes confusion for the user, and therefore it is necessary to appropriately reject the irrelevant sound.

  As a technique for rejecting irrelevant sounds other than words to be recognized, a technique using a garbage model that uniformly expresses all phonemes is known (see Non-Patent Document 1). However, the method that simply uses the garbage model has the following problems.

  First, since the user does not know what to say, it is necessary to determine whether or not the voice itself matches the garbage. However, only the pronunciation correction part is greatly different, and when correcting the same or similar words (for example, “late” and “rate”) in other parts, the user can enter “date”, “gate”, “Et”, etc. There is a possibility of pronunciation. Since such words are evaluated with matching except for the correction part at the time of speech recognition, there is a high possibility of matching well with “late” or “rate” rather than garbage.

  In pronunciation correction learning of "late" and "rate", the evaluation of pronunciation for "late" and "rate" without rejecting "date", "gate" or "eto" etc., which is different from the word whose pronunciation should be corrected Displaying will cause distrust to the user. In other words, when evaluating the accuracy of pronunciation of a specific part of a word, it is desirable to reject a pronunciation that matches only the part not to be evaluated.

JP 2004-53652 A

"Unnecessary Word Processing Method in Spoken Sentences Using Garbage HMM" IEICE Transactions, A, Fundamentals / Boundaries, J77-A (2), pp. 215-222, February 25, 1994

  An object of the present invention is to appropriately reject utterances that differ only in the pronunciation correction part and match or are similar in other parts in the pronunciation correction learning.

  The present invention has the following configuration as one means for achieving the above object.

  The information processing according to the present invention creates first to third phoneme sequence lists each including at least one phoneme sequence and different phoneme sequences. Speech recognition is performed as to whether the phoneme sequence list corresponds to the phoneme sequence, and the second phoneme sequence list replaces a predetermined subsequence of the phoneme sequence included in the first phoneme sequence list with a subsequence of a different phoneme The phoneme sequence is included, and the third phoneme sequence list includes a phoneme sequence in which the replaced subsequence in the second phoneme sequence list is further replaced with a subsequence of a different phoneme.

  According to the present invention, in pronunciation correction learning, only the pronunciation correction part is different, and utterances that match or are similar to other parts can be appropriately rejected. By this rejection, it is possible to avoid the display of an inappropriate evaluation result and prompt the user to input voice again.

1 is a block diagram illustrating a configuration example of an information processing apparatus according to a first embodiment. The figure showing the transition of the display screen of the display part at the time of using a pronunciation correction learning function. The figure explaining the structural example of dictionary item DB and problem DB. 6 is a flowchart for explaining pronunciation learning correction processing according to the first embodiment. 6 is a flowchart for explaining pronunciation learning correction processing according to the first embodiment. FIG. 3 is a block diagram illustrating a configuration example of an information processing apparatus according to a second embodiment. 9 is a flowchart for explaining pronunciation learning correction processing according to the second embodiment.

  Hereinafter, information processing according to an embodiment of the present invention will be described in detail with reference to the drawings. The information processing apparatus described below is an electronic dictionary function, for example, a portable electronic dictionary apparatus, a portable terminal apparatus such as a smartphone in which an electronic dictionary application (AP) is installed, a tablet terminal apparatus, or the like.

[Device configuration]
The block diagram of FIG. 1 shows a configuration example of the information processing apparatus of the first embodiment.

  The audio input unit 101 of the information processing apparatus includes a microphone, an analog-digital converter (ADC), a microprocessing unit (MPU), and the like. The voice input unit 101 inputs voice data obtained by converting the voice collected by the microphone into an electrical signal by the ADC to the information processing apparatus.

  The audio recording unit 102 includes a memory such as a RAM, an MPU, and the like, and records audio data input from the audio input unit 101. The voice recognition unit 103 is configured by an MPU or the like, and recognizes voice data as a predetermined word. The words recognized by the speech recognition unit 103 are determined according to a given grammar, and have a function of selecting a predetermined number of words determined as having high accuracy in the vocabulary list as candidate words.

  The speech recognition unit 103 performs speech recognition using a hidden Markov model (HMM) having a mel frequency cepstrum coefficient (MFCC) as a feature quantity. There is one HMM for each combination of one phoneme and information about the phoneme environment before and after it. This combination is called a triphone, and a triphone with a phoneme P in which a phoneme L is continuous in the front and a phoneme R is continuous in the back is expressed as “L-P + R”. At the beginning and end of the word, the triphone is constructed assuming that the silent phoneme “sil” is connected.

  Further, the speech recognition unit 103 holds one garbage model (hereinafter, GBG) which is an HMM learned from all phonemes. GBG is a model that represents unspecified phonemes and noise, and does not represent a specific phoneme. The GBG is not only statically held, but the GBG may be updated to adapt to the usage environment using ambient noise collected by the voice input unit 101 or the like.

  The display unit 104 includes a liquid crystal panel, an MPU, and the like, and displays information to be presented to the user. The displayed information includes dictionary items (hereinafter referred to as contents) such as the word candidate selected by the speech recognition unit 103 and the meaning of the word corresponding to the selected word. The display screen is not limited to the liquid crystal panel, and may be a cathode ray tube or an organic electroluminescence (EL) panel, for example.

  The operation unit 105 includes a touch panel, an MPU, and the like arranged on the display screen of the display unit 104. The user refers to the display on the display unit 104 and operates the touch panel of the operation unit 105 to input a user instruction to the information processing apparatus. That is, a user interface (UI) is provided by a combination of the display unit 104 and the operation unit 105.

  The user instruction includes a start of utterance of a word to be recognized, an instruction to select one from a plurality of word candidates, and the like. The device for inputting the user instruction is not limited to the touch panel, and may be a switch or a button, for example.

  The problem selection unit 106 is configured by an MPU or the like, and selects a problem for the user to correct pronunciation according to an operation on the operation unit 105. The problem consists of one word to be pronounced by the user (a correct word to be described later) and one or more words (a reference word to be described later) that are confused with the pronunciation of the correct word (difficult to distinguish). The problem is selected from problems stored in a problem database (DB) 111 described later.

  The grammar creation unit 107 is configured by an MPU or the like, and creates a speech recognition grammar based on the problem selected by the problem selection unit 106. The created grammar is used by the speech recognition unit 103.

  The phoneme clustering unit 108 is configured by an MPU or the like, and performs clustering of similar phonemes. The model integration unit 109 is configured by an MPU or the like, integrates a plurality of phoneme acoustic models, and newly creates an acoustic model.

  The dictionary item DB 110 includes a nonvolatile memory such as a ROM, a flash memory, and the like, and stores dictionary items such as word meanings provided by the information processing apparatus. A dictionary item is associated with a headword, and a user can select a headword to be examined and browse the corresponding dictionary item. The problem DB 111 is composed of a nonvolatile memory such as a ROM, a flash memory, or the like, and stores problems for the problem selection unit 106 to select.

[Usage form]
First, a mode when the user uses the pronunciation correction learning function of the information processing apparatus will be described. FIG. 2 shows the transition of the display screen of the display unit 104 when using the pronunciation correction learning function.

  In FIG. 2A, reference numeral 201 denotes a housing of the information processing apparatus. Reference numeral 202 denotes a touch panel, and the user inputs a user instruction to the operation unit 105 by referring to information displayed on the display unit 104 and touching a button display. Reference numeral 203 denotes a microphone of the voice input unit 101.

  When the apparatus is turned on, the electronic dictionary AP is activated, etc., and the electronic dictionary function is enabled. Further, when the user instructs execution of the pronunciation correction learning function, the display unit 104 displays the problem list 204. Is displayed (Fig. 2 (A)). When the user touches the problem he / she wants to practice, the display unit 104 displays a word to be uttered and a message prompting the utterance (FIG. 2B).

  The user touches the “start” button 205 and then starts speaking. When the operation unit 105 detects a touch at a position corresponding to the “voice input” button 205, the voice recording unit 102 records voice data input from the voice input unit 101 in a memory.

  The voice recognition unit 103 determines the end of voice input using VAD (voice activity detection) or the like. If it is determined that the voice input has been completed, voice recognition is performed based on the voice data recorded in the memory of the voice recording unit 102, and the utterance is evaluated. The display unit 104 displays contents corresponding to the evaluation.

  FIG. 2C shows a display notifying that the user has pronounced correctly when it is evaluated that the utterance of the user is closest to the word to be pronounced (“late” in FIG. 2). FIG. 2D shows a display notifying that the user's pronunciation is incorrect when the user's utterance is evaluated to be closest to a similar word (for example, “rate”). FIG. 2 (E) is a display that prompts the user to re-utter (re-input voice data) when it is evaluated that the user's utterance is not close to either a word to be pronounced or a similar word.

[DB configuration]
A configuration example of the dictionary item DB 110 and the problem DB 111 will be described with reference to FIG. Here, an example is shown in which the dictionary item DB 110 and the problem DB 111 constitute an English-Japanese dictionary, but a dictionary or encyclopedia in other languages may be constituted.

  As shown in FIG. 3 (A), the dictionary item DB 110 has “item ID”, “item heading”, “pronunciation”, “content” as fields, and a plurality of records in which information of each field is recorded ( Field record).

  In the “item ID” field, an identification number (item ID) uniquely added to each record of the dictionary item DB 110 is recorded. In the “item heading” field, words (leading words) representing items of the item record are recorded. For example, in the case of an English-Japanese dictionary, the headword is an English word.

  In the “phonetic notation” field, a phoneme sequence indicating how to read the headword recorded in the “item heading” field is recorded. For example, in the case of an English-Japanese dictionary, it is a pronunciation symbol string of English words that are headwords. For example, as shown in the item record of item ID = 0 in FIG. 3A, there are cases where four phonetic notations delimited by the delimiter “/” are recorded in the “phonetic notation” field. That is, a plurality of phonetic notations may be recorded in one item record. Moreover, although the international phonetic symbol (IPA) is used for the phonetic notation, the phonetic notation may be recorded by other systems. The voice recognition unit 103 performs voice recognition with reference to the phonetic notation.

  Information related to the entry word of the item record is recorded in the “content” field. The information recorded in the “Content” field is the information to be browsed by the user. For example, in the case of English-Japanese dictionaries, Japanese text is mainly used to explain the meaning and examples of English words that are headwords. is there.

  As shown in FIG. 3B, the problem DB 111 is composed of a plurality of problem records, and a unique identification number (vocabulary list ID) is added to each problem record. The problem record is a word list including a correct answer flag, word notation (word spelling), phoneme series (phonetic symbol string), and correction location information.

  The correct answer flag is a flag indicating a word that should be pronounced correctly in the word list. A word with the correct answer flag set to TRUE is called a “correct word”, and a word with a false flag is called a “control word”. The question record includes at least one correct word and at least one control word.

  The correction location information is information indicating the position of the pronunciation to be corrected in the word. For each word notation and phoneme series, the pair of the start point and end point of the range to be corrected is held as a numerical value indicating the number of characters or the number of phonemes from the beginning. The phoneme corresponding to the phonetic symbol indicated by the correction location information in the phoneme sequence is called “corrected phoneme”. The correct phoneme of the correct word is called “correct phoneme”, and the correct phoneme of the contrast word is called “control phoneme”. The correct phoneme and the contrast phoneme are partial strings of a phoneme sequence including one or more phonemes, and are not limited to a single phoneme.

  The dictionary item DB 110 and the problem DB 111 are created in advance and stored in the non-volatile memory of the information processing apparatus or incorporated in the data for installing the electronic dictionary AP. However, the problem records may not be created in advance, but may be collected as required by collecting item records in the dictionary item DB 110 that meet the given conditions, and may be created and stored in a memory such as a RAM. If a problem record is created as necessary, the storage capacity of the nonvolatile memory can be suppressed, but the processing load when the electronic dictionary function is executed increases.

[Information processing]
The pronunciation learning correction process according to the first embodiment will be described with reference to the flowcharts of FIGS. The pronunciation learning correction process is executed by the MPU of the information processing device, which executes the electronic dictionary function stored in the non-volatile memory of the information processing device shown in FIG. 1 and the electronic dictionary AP installed in the non-volatile memory. It is realized by doing. In other words, the MPU functions as a control unit that controls the operation and processing of the configuration shown in FIG.

  When the pronunciation learning correction function is enabled, the MPU displays the problem record stored in the problem DB 111 on the display unit 104 as the problem list 204 as shown in FIG. 2 (A) (S401), and handles the problem. Wait for the position to be touched (S402). In the following, for the sake of simplification, “the position corresponding to XXX on the touch panel is touched” is expressed as “XX is touched”.

  When a problem is touched, the MPU performs phoneme clustering according to the selected problem by the phoneme clustering unit 108, assuming that the problem is selected (S403).

● Clustering (S403)
Clustering is performed for the following triphones.
・ Triphone (correct phoneme triphone) with the phoneme environment of the correct and correct phonemes of the selected problem in the correct word,
A triphone with a phoneme environment before and after the control phoneme of the selected problem in the control word (control phoneme triphone),
A triphone in which the central phoneme of the above triphone is replaced with the entire phoneme used in the speech recognition unit 103.

  Here, each phoneme is mapped to a point on the Euclidean space based on the HMM corresponding to each triphone used by the speech recognition unit 103, and clustering is performed in the Euclidean space. Each HMM has s states, and each state has an N-dimensional single Gaussian distribution with no covariance. For example, s = 3 and N = 26.

The mean m and variance σ of the Gaussian distribution of the i-th state of the HMM of phoneme X are respectively expressed by the following equations.
┌ ┐
│m _Xi1 │
│ ： │
_{m Xi = │m Xij │ ... (} 1)
│ ： │
│m _XiN │
└ ┘
┌ ┐
│σ _Xi1 │
│ ： │
_{σ Xi = │σ Xij │ ... (} 2)
│ ： │
│σ _XiN │
└ ┘

At this time, the phoneme X is mapped to a point in the s × N-dimensional Euclidean space expressed by the following equation. Here, the dimension compression may be performed using a technique such as principal component analysis (PCA).
┌ ┐
│m _X11 / σ _X11 │
│ ： │
│m _X1N / σ _X1N │
│m _X21 / σ _X21 │
p _X = │: │… (3)
│m _X2N / σ _X2N │
│ ： │
│m _Xs1 / σ _Xs1 │
│ ： │
│m _XsN / σ _XsN │
└ ┘

Here, the following distance is introduced into the Euclidean space. The distance between point P and point Q is determined as follows:
d _C (P, Q) = min [{w (c1) d (P, c1) + w (c2) d (Q, c2) | c1, c2∈C} ∪ {d (P, Q)}]… (Four)
Where C is the set of points mapped to the correct phoneme triphone and the contrast phoneme triphone,
d is the normal Euclidean distance,
w is a function from C to a non-negative real number.

  For example, the function w is a function that gives 0.5 for the correct phoneme triphone and 1 for the reference phoneme triphone. The function w serves as a weight for making it easier for a phoneme close to the correct phoneme to enter the same cluster than a phoneme close to the reference phoneme. Note that the method of giving weights is not limited to this, and the correct answer phoneme and the reference phoneme are not uniform, and different weights may be given to each element.

That is, the distance d _C is a distance in which all elements of the set C are identified and allowed to pass through the elements of the set C. With the distance d _C, performs clustering by K-means clustering method.

Since the distance d _C between elements of the set C is 0, all elements of the set C are included in the same cluster. A cluster including the set C is referred to as a “correct answer cluster”, and an element of the set obtained by removing the elements of the set C from the correct answer cluster is referred to as an “excluded phoneme”. There may be no excluded phonemes.

  This clustering is a means for determining the similarity between phonemes. That is, the triphone HMMs classified into the same cluster are determined to be similar speech expressions.

The number of states s and the number of state dimensions N of the HMM are not limited to s = 3 and N = 26, and the method of mapping to the metric space is not limited to the above. Each state of the HMM is not limited to an N-dimensional single Gaussian distribution having no covariance, but may be a Gaussian distribution having a covariance, a mixed Gaussian distribution, or another type of distribution such as a Poisson distribution. The distance d _C used here is a legitimate distance that satisfies the distance axiom, but this is not necessarily the case. For example, a pseudo distance such as symmetrized Kullback-Divergence can be used. .

  Also, instead of correcting the distance for the correct phoneme and the reference phoneme by the same view of the set C, for example, a figure surrounded by the elements of the set C as vertices (if there are only two elements of the set C, these The shortest distance to the line segment connecting the two) may be used.

  When the phoneme clustering is completed, the MPU integrates the acoustic model for the clusters other than the correct cluster among the clusters obtained by the clustering in step S403 by the model integration unit 109 (S404).

● Acoustic model integration (S404)
The integration of the acoustic model is simply performed by having the HMM corresponding to one triphone in the cluster represent the cluster. In step S403, the triphone HMM closest to the center of gravity of the cluster determined by the K-means method (at a distance d _C ) is regarded as the acoustic model of the cluster.

  A symbol that uniquely identifies the cluster is assigned to the integrated HMM. This symbol is regarded as one phoneme and is called “integrated phoneme” of the cluster. The integrated phoneme is the same as a triphone that represents the phoneme environment before and after the triphone.

  Thus, an object is to reduce the number of HMMs to be calculated by the speech recognition unit 103 and the amount of calculation of the cluster. Here, create an HMM in which each phoneme HMM is connected in parallel, or create an HMM using a state as a mixed GMM multiplied by a weight for each state of each phoneme HMM, An acoustic model may be integrated.

  When the integration of the acoustic models is completed, the MPU creates a list of phonemes from which the phonemes (correct phonemes), reference phonemes, and excluded phonemes of the correct clusters are excluded from all phonemes by the grammar creation unit 107 (S405). The phonemes included in this list are called “rejecting phonemes”. GBG (garbage model) is also regarded as one of the phonemes and added to the rejecting phonemes.

  Since the excluded phoneme is a phoneme close to the correct phoneme / contrast phoneme, there is a possibility that a user who is in the process of correcting pronunciation will utter the excluded phoneme with the intention of uttering the correct phoneme or the reference phoneme. Since those close to the phonemes to be corrected are not rejected, they are excluded here.

  Next, the MPU creates a list of phoneme sequences in which the correct phoneme in question is replaced with a rejecting phoneme, and all other phonemes are replaced with integrated phonemes of the cluster to which the MPU belongs by the grammar creation unit 107 ( S406). Then, in addition to the phoneme series included in the created list, a grammar including correct words, contrast words, and GBG as words is created (S407). Note that phonemes belonging to the correct cluster are not replaced.

  The words and GBGs of the phoneme series list created at the beginning are words for determining that they do not match either the correct word or the reference word, and are referred to as “reject words” here. In other words, a word obtained by replacing the phoneme to be evaluated with all other possible phonemes is added as a rejection word to the speech recognition grammar.

  Each word in the grammar is expressed as a triphone sequence, and speech recognition is performed by the corresponding HMM. Here, for GBG and integrated phonemes, triphones that are not determined by the corresponding HMM may appear in the grammar. In this case, the model integration unit 109 creates and substitutes an HMM in which triphone phonemes having the same front and back environment as the necessary triphone are integrated using phonemes included in the cluster of integrated phonemes. Similarly, in the case of GBG, HMMs are integrated and replaced using triphones of all phonemes.

  Next, the MPU waits for the user to touch the “voice input” button 205 on the touch panel (operation unit 105) (S408). When the “voice input” button 205 is touched, the voice input unit 101 starts voice input by the voice input unit 101 and voice data recording by the voice recording unit 102 (S409), and performs voice input by performing VAD or the like. Is determined to end (S410). For example, voice and non-voice HMMs are prepared, and the likelihood of each model is compared in VAD. If the likelihood of the voice HMM is high, it is determined as voice, and if the likelihood in the non-voice HMM is high, it is determined as non-voice. Then, the end of voice input is determined. Note that the VAD is not limited to this, and may adopt a method of determining the sound when the energy of the sound signal exceeds a specific threshold.

  Step S409 is repeated until the voice input is completed. When the voice input is completed, the MPU performs voice recognition of the recorded voice data by the voice recognition unit 103 (S411), and the recognition result is the correct word, the reference word, and the rejection word described in the grammar created in step S407. It is determined which of the words is close (S412).

  Next, if the determined word is the correct word, the MPU displays a notification that the user pronounced correctly on the display unit 104 (FIG. 2 (C)) (S413). If it is a contrast word, the user's utterance is recognized as a contrast word and a hint notification for correctly speaking is displayed on the display unit 104 (FIG. 2 (D)) (S414) The process is terminated.

  If the MPU determines that the word is a rejection word, the MPU displays a message prompting the user to speak again (FIG. 2 (E)) (S415), and the process returns to step S408. In other words, pronunciations that are significantly different from either the correct word or the contrast word are rejected, and an inappropriate result display is avoided, and the user is prompted to input the utterance again.

  In this way, a first phoneme sequence list composed of phoneme sequences of correct words is created. A second phoneme sequence list of reference words is created that includes a phoneme sequence in which a predetermined subsequence of a phoneme sequence included in the first phoneme sequence list is replaced with a subsequence of a different phoneme. A third phoneme sequence list of rejection words is generated, including a phoneme sequence in which the replaced subsequence in the second phoneme sequence list is further replaced with a different phoneme subsequence. Then, it is determined which phoneme sequence of the first to third phoneme sequence lists corresponds to the speech data recognition result. Thus, only the pronunciation correction part is different, and words that are the same or similar in other parts are appropriately rejected, display of inappropriate results can be avoided, and the user can be prompted to input voice again.

  The information processing according to the second embodiment of the present invention will be described below. Note that the same reference numerals in the second embodiment denote the same parts as in the first embodiment, and a detailed description thereof will be omitted.

  In the first embodiment, an example in which a grammar is dynamically created when the pronunciation correction learning function is executed has been described. However, when the number of phonemes is large, or when a problem with a large number of correct words and reference words is selected, the processing load may increase and the processing speed may decrease. In consideration of such a situation, Embodiment 2 in which a grammar is created in advance for each problem and the created grammar is stored will be described.

  A block diagram of FIG. 6 shows a configuration example of the information processing apparatus of the second embodiment. 1 is that the grammar creation unit 107, the phoneme clustering unit 108, and the model integration unit 109 are replaced with a grammar DB 112 and a grammar selection unit 113.

  The grammar DB 112 stores a grammar used by the speech recognition unit 103 for recognition. The grammar is created in advance for each problem in the same procedure as the method described in the first embodiment. The grammar selection unit 113 is configured by an MPU or the like, and searches and acquires a grammar corresponding to the problem selected by the user from the grammar stored in the grammar DB 112.

  The pronunciation learning correction process according to the second embodiment will be described with reference to the flowchart of FIG. The difference from the process shown in FIG. 4 is that the process from steps S403 to S407 is replaced with a process (S421) in which the grammar selection unit 113 selects and acquires a grammar corresponding to the selected problem from the grammar DB 112. Other processes are substantially the same as the processes in FIGS. There is no difference in the grammar selected and acquired by the speech recognition unit 103 in step S412.

  Thus, if the grammar is created in advance, the processing load when the pronunciation correction learning function is executed can be reduced.

[Other Examples]
The present invention can also be realized by executing the following processing. That is, software (program) that realizes the functions of the above-described embodiments is supplied to a system or apparatus via a network or various storage media, and a computer (or CPU, MPU, etc.) of the system or apparatus reads the program. It is a process to be executed.

Claims (17)

Creating means for creating first to third phoneme sequence lists each including at least one phoneme sequence and different phoneme sequences;
Recognizing means for recognizing which of the first to third phoneme series lists the input voice data corresponds to the phoneme series included in the list,
The second phoneme sequence list includes a phoneme sequence obtained by replacing a predetermined partial sequence of a phoneme sequence included in the first phoneme sequence list with a partial sequence of different phonemes,
The information processing apparatus includes a phoneme sequence in which the third phoneme sequence list is obtained by replacing the replaced subsequence in the second phoneme sequence list with a subsequence of a different phoneme.   2. The information processing apparatus according to claim 1, further comprising user interface means for a user to select a word for inputting the voice data and to display an evaluation based on the result of the voice recognition.   3. The information processing apparatus according to claim 2, wherein the user interface means performs a display prompting re-input of the speech data when the speech recognition result indicates a phoneme sequence included in a third phoneme sequence list. The first phoneme sequence list includes phoneme sequences corresponding to the word selected by the user;
The second phoneme sequence list includes phoneme sequences corresponding to words that are difficult to distinguish between the selected word and pronunciation,
4. The phoneme sequence according to claim 1, wherein the third phoneme sequence list includes a phoneme sequence different from a phoneme sequence corresponding to the selected word and a phoneme sequence corresponding to the word whose pronunciation is difficult to distinguish. Information processing apparatus described in 1.   4. The phoneme sequence list according to any one of claims 1 to 3, wherein the third phoneme sequence list includes the entire phoneme sequence in which at least the entire phoneme recognized by the recognizing unit is replaced with a single subsequence. Information processing apparatus described in the section. The recognition means recognizes noise;
4. The creation unit according to claim 1, wherein the third phoneme sequence list includes a phoneme sequence in which the recognized noise is regarded as a phoneme and a partial sequence is replaced. Information processing device. The recognition means uses, for each phoneme, an acoustic model for each phoneme that is continuous before and after,
7. The information processing according to claim 6, further comprising: an acoustic model creating unit that creates an acoustic model when the phonemes regarded as the noise are consecutively used before and after using the acoustic model for each of the phonemes. apparatus. Furthermore, a determination means for determining the similarity between phonemes,
4. The information processing apparatus according to claim 1, further comprising: exclusion means for excluding a phoneme sequence having similar substrings before and after the replacement from the third phoneme sequence list.   9. The information processing apparatus according to claim 8, wherein the determination unit clusters all phonemes recognized by the recognition unit and determines that phonemes included in the same cluster are similar.   10. The information processing apparatus according to claim 9, wherein the determination unit performs clustering on points where the phonemes are mapped to a predetermined metric space.   The determination means is a point mapping phonemes belonging to the replaced subsequence in the first phoneme sequence list, and a point mapping phonemes belonging to the replaced subsequence in the second phoneme sequence list; 11. The information processing apparatus according to claim 10, wherein a distance obtained by assuming that the two are the same is used for the clustering.   The determination means includes mapping a phoneme belonging to the replaced subsequence in the first phoneme sequence list and mapping a phoneme belonging to the replaced subsequence in the second phoneme sequence list. 11. The information processing apparatus according to claim 10, wherein distances obtained by multiplying different weights are used for the clustering.   The metric space is a Euclidean space, and the determination means includes a centroid of at least one point mapping a phoneme belonging to the replaced subsequence in the first phoneme sequence list, and a second phoneme sequence list. 11. The information processing apparatus according to claim 10, wherein a distance of a line segment connecting a centroid of at least one point to which a phoneme belonging to the replaced subsequence is mapped is used for the clustering. Furthermore, from the acoustic model of a plurality of phonemes, comprising an integration means for creating a new acoustic model representing the plurality of phonemes,
The integration unit creates a new acoustic model from the acoustic models of phonemes included in the cluster clustered by the determination unit,
14. The information processing apparatus according to claim 8, wherein the recognizing unit substitutes the new acoustic model for a phoneme included in the cluster. An information processing method for an information processing apparatus having a creation unit and a recognition unit,
The creating means creates first to third phoneme sequence lists each including at least one phoneme sequence and different phoneme sequences;
The recognizing means recognizes the input speech data as to which of the first to third phoneme sequence lists corresponds to the phoneme sequence included;
The second phoneme sequence list includes a phoneme sequence obtained by replacing a predetermined partial sequence of a phoneme sequence included in the first phoneme sequence list with a partial sequence of different phonemes,
The information processing method including the phoneme sequence in which the third phoneme sequence list further replaces the replaced subsequence in the second phoneme sequence list with a subsequence of a different phoneme.   15. A program for causing a computer to function as each unit of the information processing apparatus according to any one of claims 1 to 14.   A computer-readable recording medium in which the program according to claim 16 is stored.

Images