JP2004177551A - Unknown speech detecting device for voice recognition and voice recognition device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an unknown speech detecting device for voice recognition that can detect an unknown speech with high accuracy by using not only a single decision scale, but also a plurality of decision scales in combination. <P>SOLUTION: In addition to collation using word patterns of a vocabulary registered in a recognition dictionary, the unknown speech detecting device for voice recognition performs collation, using subword connection models, in which chain probability of subwords is taken into consideration in the stage of voice recognition to find a plurality of unknown speech detection scales from collation results, and takes those results into consideration to decide an unknown speech. <P>COPYRIGHT: (C)2004,JPO

Description

[0001]
TECHNICAL FIELD OF THE INVENTION
The present invention relates to a speech recognition method in a speech recognition device.
[0002]
[Prior art]
2. Description of the Related Art Conventionally, a speech recognition apparatus defines an acceptable speech recognition vocabulary, searches for a recognition vocabulary most similar to an input speech, and outputs the same as a recognition result. Therefore, even if the user makes an utterance outside the speech recognition vocabulary, the most similar vocabulary is selected from the speech recognition vocabulary, and the recognition result is incorrect. Therefore, it is necessary to determine whether the utterance of the user is a word included in the speech recognition vocabulary, a word other than the utterance, or a stagnant word, and reject the unknown utterance.
[0003]
A method of rejecting such unknown utterances is to form a model of each word in the recognized vocabulary by connecting HMMs (one of models representing voice patterns) in units shorter than words called subwords. A word that gives the maximum likelihood is searched for, and this likelihood is set as the recognition likelihood (for example, see Non-Patent Document 1). Also, the maximum likelihood for the input speech is obtained from the model obtained by connecting the arbitrary subwords HMM corresponding to the arbitrary kana sequence, and this is set as the reference likelihood. By comparing the recognition likelihood obtained in this way with the reference likelihood, unknown utterances are detected and rejected.
[0004]
However, in such a method, in calculating the reference likelihood, there is no restriction on the concatenation of arbitrary subwords HMM, and the likelihood for a non-Japanese series is often selected as the maximum likelihood. Such comparison between the reference likelihood and the recognition likelihood did not provide a sufficient effect of rejecting unknown utterances. Further, since the connection of all subword HMMs is compared, a large resource is required in terms of the amount of processing calculation. In order to solve such a problem, for example, in the speech recognition method, both the rejection accuracy of unknown utterances and the processing amount are improved by introducing the affinity of connection between subwords HMM as a transition probability (for example, Patent Document 1).
[0005]
[Patent Document 1]
JP-A-10-171489
[Non-patent document 1]
Watanabe et al., "Rejection of unknown utterance by likelihood correction using syllable recognition", IEICE Transactions, Vol. J75-D-II, No. 12 (1992)
[0006]
[Problems to be solved by the invention]
However, the conventional method described above has the following problems.
[0007]
That is, the model based on the concatenation of the sub-word HMMs can be regarded as a model that recognizes the input speech as a kana sequence. Even if the transition probability is introduced as a constraint on the concatenation between the sub-word HMMs, the model is generated. The sequence is still hard to match well with that of the input speech. That is, the likelihood of reference obtained by such a model is not sufficiently accurate, and the effect of rejecting unknown utterances is not sufficient.
[0008]
[Means for Solving the Problems]
In order to achieve the above object, an unknown utterance detection device for speech recognition according to the first aspect of the present invention includes a speech analysis unit that analyzes input speech and converts it into a sequence of feature parameters, and a recognition dictionary that defines a vocabulary to be recognized. Storage means;
A speech model storage unit that models a standard pattern of speech, and a vocabulary model defined in the recognition dictionary are constructed using the model stored by the speech model storage unit, and words to be compared with the input speech Level matching means, a subword transition probability storage means for defining a transition probability between subwords, and a speech model stored by the speech model storage means, taking into account the subword transition probabilities stored by the subword transition probability storage means. Sub-word level matching means for concatenating and collating with the input voice;
An unknown utterance scale calculation unit that calculates a plurality of unknown utterance scales from the word level matching unit and the subword level matching unit; and an unknown utterance scale based on the plurality of scales calculated by the unknown utterance scale calculation unit. Unknown utterance determining means for performing determination.
[0009]
According to the above configuration, in the determination of the unknown utterance, it is possible to determine the possibility that the input voice is an unknown utterance from a plurality of viewpoints, and it is possible to exhibit high unknown utterance detection performance.
[0010]
Further, in the unknown utterance detection device according to the first invention, in the unknown utterance scale calculating means, the likelihood of a word obtained by the word level matching means and a subword chain likelihood obtained by the subword level matching means. It is desirable to include a value calculated based on the difference between the two.
[0011]
According to the above configuration, an effect of correcting the likelihood of the word based on the likelihood of the subword chain is obtained, and a high unknown utterance detection performance is obtained.
[0012]
Further, in the unknown utterance detection device according to the first invention, the unknown utterance scale calculating means obtains the acoustic feature of the word model of the first candidate obtained by the word level matching means and the subword level matching means. It is desirable to include a value calculated based on the similarity between the acoustic features of the obtained subword chain model.
[0013]
According to the above configuration, it is possible to determine an unknown utterance by focusing on the similarity of the acoustic features of the two models, and high unknown utterance detection performance is obtained.
[0014]
In the unknown utterance detection device according to the first aspect of the present invention, the unknown utterance scale calculating means includes a difference between the likelihood of the word of the first candidate and the likelihood of the word of the lower candidate obtained by the word level matching means. It is desirable to include a value calculated based on
[0015]
According to the above configuration, at the time of recognition of an unknown utterance, the word level matching unit can model a feature that an erroneous candidate is similar and can be obtained with a high likelihood, and high unknown utterance detection performance can be obtained.
[0016]
Further, in the unknown utterance detection device according to the first invention, in the unknown utterance scale calculating means, the acoustic feature of the first candidate word obtained by the word level matching means and the acoustic feature of the lower candidate word are obtained. It is desirable to include a value calculated based on the similarity between the two.
[0017]
According to the above configuration, it is possible to determine an unknown utterance by focusing on the acoustic similarity between the models of the candidate word, and a high unknown utterance detection performance can be obtained.
[0018]
Further, the speech recognition apparatus of the second invention is a speech recognition apparatus for recognizing an input speech by collating with a model corresponding to a vocabulary registered in a recognition dictionary, wherein the unknown speech detection apparatus And outputs the recognition result in consideration of the output result of the unknown utterance detection device.
[0019]
According to the above configuration, the speech recognition apparatus does not always output any one of the vocabularies in the recognition dictionary for any input speech, and the speech content is not included in the recognition dictionary. For example, this can be transmitted to the user, and it is possible to provide a user-friendly interface among various voice recognition interfaces including the voice recognition device.
[0020]
BEST MODE FOR CARRYING OUT THE INVENTION
Hereinafter, embodiments of the present invention will be described with reference to the drawings.
[0021]
(Embodiment 1)
FIG. 1 shows a block diagram of an unknown utterance detection device according to the present embodiment. In FIG. 1, reference numeral 1 denotes an acoustic analysis unit that A / D converts an input voice and converts the input voice into a time series of feature parameters. Reference numeral 2 denotes a speech piece pattern storage unit that stores speech pieces of standard speech used for matching with the feature parameters of the input speech.
[0022]
Here, the speech piece is a VC pattern in which the second half of a vowel section of a voice is connected to the first half of a consonant section subsequent thereto, and a CV in which the second half of a consonant section is connected to the first half of a subsequent vowel section. It means a set of patterns. However, in addition to this, the speech piece is a set of phonemes that are almost equivalent to each letter of the alphabet when Japanese is written in Roman characters, and the mora of a character that is almost equivalent to each letter when it is written in Japanese. It may be a set, a set of subwords meaning a chain of a plurality of moras, or a mixed set of these sets.
[0023]
Reference numeral 3 in FIG. 1 denotes a word dictionary storage unit in which rules for synthesizing the word pattern of the speech recognition vocabulary by connecting the speech segments are stored. Reference numeral 4 denotes a word matching unit that compares an input voice expressed in a time series of feature parameters with the synthesized word pattern and obtains a likelihood corresponding to the similarity for each word.
[0024]
Reference numeral 5 denotes a transition probability storage unit that stores transition probabilities that represent the naturalness of the connection as continuous values when the voice segments are arbitrarily combined. In the present embodiment, a 2 gram probability of a phoneme is used as the transition probability. The 2 gram probability of a phoneme means a probability P (y | x) that a phoneme y is connected after a preceding phoneme x, and is obtained in advance using a large number of Japanese text data or the like. However, the transition probability may be a 2 gram probability of a mora, a 2 gram probability of a subword, or a 2 gram probability of a mixture of these, or a 3 gram probability other than the 2 gram probability.
[0025]
Reference numeral 6 in FIG. 1 is obtained by calculating the likelihood of a pattern formed by arbitrarily combining the above-mentioned speech piece patterns and the input speech expressed as a time series of the characteristic parameters in consideration of the above-mentioned transition probabilities. This is a speech sequence matching unit that uses the maximum likelihood as a reference likelihood.
[0026]
7 is a difference between the likelihood of the word that obtained the highest value (first candidate) and the word that obtained the next highest value (second candidate) among the likelihoods of each word calculated by the word matching unit. Is normalized by the word length.
[0027]
No. 8 is a candidate phoneme sequence similarity calculation unit that calculates the distance between the phoneme sequence of the first candidate and the phoneme sequence of the second candidate in order to obtain the acoustic similarity between the first candidate and the second candidate. is there.
[0028]
Reference numeral 9 denotes a candidate / speech sequence score difference calculation unit that calculates a difference between the likelihood of the first-place candidate and the reference likelihood calculated by the speech sequence matching unit by normalizing the difference with the word length.
[0029]
Reference numeral 10 denotes a candidate / speech sequence / phoneme sequence similarity calculation unit that calculates the acoustic similarity between the first candidate and the sequence determined as the optimal sequence by the speech sequence matching unit as the distance between the phoneme sequences. is there.
[0030]
Reference numeral 11 denotes each value obtained by the above-described inter-candidate score difference calculation unit, the candidate / phoneme sequence similarity calculation unit, the candidate / speech sequence score difference calculation unit, and the candidate / speech sequence / phoneme sequence similarity calculation unit. Overall, the unknown utterance determination unit determines whether or not the input voice is an unknown utterance.
[0031]
In the present embodiment, the above-mentioned four scales are used as the scales used in the unknown utterance determination unit. In addition, the likelihood of each word candidate and its distribution, and the locality in the word section are also used. It is also possible to use a measure such as the amount of change in the score or the duration information of the phonemes constituting the word. Further, as a method of determining an unknown utterance based on a plurality of scales, in the present embodiment, a linear discriminant obtained in advance using a large number of recognition result cases is used. However, besides this, it is also effective to use a so-called learning machine such as a neural network, a decision tree, or an SVM (support vector machine).
[0032]
Next, the processing operation of unknown utterance detection in the present embodiment will be described. The input voice is first analyzed by the voice analysis unit after A / D conversion, and is converted into LPC vectors every 10 ms. The LPC vector is a parameter that indicates the spectral envelope of the short-time spectrum of the voice, and is used as a parameter that well represents the phonological features of the voice. In a normal speech recognition method, a time series of LPC vectors obtained at regular intervals from an input speech is used as a feature vector of the input speech and matched with a previously obtained word model, and is called a likelihood for each word. Find a score.
[0033]
In the present embodiment, a word model is created using a speech piece pattern and a word dictionary. That is, the speech pattern stored in the speech pattern storage unit is connected to form a word pattern based on the speech pattern connection rule for synthesizing the word pattern stored in the word dictionary storage unit. FIG. 2 illustrates an image in which speech pattern called CV / VC pattern used in the present embodiment is connected to synthesize word pattern “Hachi no He”.
[0034]
In the speech piece pattern, parameters indicating a standard distribution (assuming a normal distribution) of LPC vectors of each speech piece are stored in time series. In recent years, a transition network called HMM (Hidden Markov Model) is often used as a model for speech recognition. Even when the HMM model is used, the HMM model representing the speech pattern is stored in the speech pattern storage unit 2, and the rule regarding the transition between the HMM models is defined in the word dictionary storage unit 3, whereby the HMM of the word is defined. It is possible to build a model.
[0035]
The feature parameter time series of the input speech is compared with the word pattern in the word matching unit 4 and the likelihood for all the words defined in the word dictionary storage unit 3 or the upper candidate words remaining in the beam of a certain likelihood. Is calculated and sorted in descending order of likelihood. FIG. 3 shows an output example of words sorted in the order of likelihood.
[0036]
In parallel with this, the speech sequence matching section 6 also performs matching of an arbitrary sequence of speech segments. In this method, speech segments are connected freely under certain restrictions, and a speech segment sequence closest to the input speech and its likelihood are calculated. At this time, if no restriction is added to the connection of the voice segments, the calculation result becomes a non-Japanese sequence, and the likelihood cannot be said to be a sufficiently significant value. Therefore, in the process of searching for the optimal speech segment sequence, the phoneme 2 gram probability stored in the transition probability storage unit 3 is used as the cost of speech segment selection and connection. For the phoneme 2 gram probability, a large number of Japanese texts of the recognition task and the same task are converted into a phoneme sequence, and the one calculated based on this is used.
[0037]
In FIG. 4, as an example of the phoneme 2 gram probability, a probability that five vowels / a /, / i /, / u /, / e /, / o / follow the preceding phoneme / k /, respectively, is illustrated. In the case of this example, it is shown that the vowel that tends to follow the consonant / k / is / a /, and then / i /. The speech sequence matching unit 6 obtains a value obtained by weighting the likelihood of the concatenated speech pattern and the likelihood of transition obtained by the logarithmic sum of the phoneme 2 gram probabilities, and adopts the sequence having the highest value.
[0038]
FIG. 5 shows an example of calculating the transition likelihood for the sequence / kobajasi / from the phoneme 2gram probability. Also, in FIG. 6, as an example showing the effect of introducing the transition likelihood, the speech sequence output from the speech sequence matching unit and the likelihood for the input speech "Kobayashi" are shown. As shown in this figure, when the transition likelihood is not used, a sequence that is far apart from “kobayashi” of / pobaeasii / obtains a more similar / obajasi / higher pattern likelihood. By using the total likelihood in consideration of the likelihood, / obajasi / closer to "kobayashi" is selected.
[0039]
As described above, the recognition likelihood for each word defined in the word dictionary storage unit, the reference likelihood, and the speech sequence at that time are obtained. Next, based on these, various scales for determining unknown utterances are calculated. Do.
[0040]
First, the inter-candidate score difference calculator 7 calculates the highest likelihood (the likelihood of the first candidate) and the next highest likelihood (the likelihood of the second candidate) among the likelihoods obtained by the word matching unit. Calculate the value obtained by dividing the likelihood difference by the word length. For example, in the case of the result shown in FIG. 3, the difference between the likelihood of “Kobayashi” and “Hayashi” is normalized by the word length to obtain 6.9.
[0041]
The candidate phoneme sequence similarity calculator 8 calculates the similarity between the phoneme sequence of the first candidate and the phoneme sequence of the second candidate. Here, as the similarity between the sequences, a value obtained by normalizing the editing distance by the sum of the sequence lengths of the two sequences is used. The editing distance is defined as a minimum when the cost required for one element replacement (replacement), one element deletion (dropout), and one element addition (insertion) is set to 1 when one series is edited and converted to the other series. Means the total cost when editing with the cost of FIG. 7 exemplifies a method of obtaining an edit distance for two phoneme sequences / uenoeki / and / jenokii /. According to such a method, the candidate phoneme sequence similarity calculation unit calculates the edit distance 3 of “Kobayashi (/ kobayashi /)” and “Hayashi (/ hajasi /)” for each phoneme sequence in the case of the result shown in FIG. 3, for example. Outputs the value of 0.21 divided by the sum of the lengths. In this embodiment, the value based on the edit distance is used as the inter-sequence similarity. However, it is also effective to use the inter-sequence distance in consideration of the acoustic similarity between phonemes. is there.
[0042]
The candidate / speech sequence score difference calculation unit calculates a value obtained by normalizing the difference between the likelihood of the first candidate obtained by the word matching unit and the reference likelihood obtained by the speech sequence matching unit by the time length of the word. . For example, in the examples shown in FIGS. 3 and 6, the difference between the recognition likelihood 2055 of the first-place candidate and the reference likelihood 2014 is normalized to 0.87 to obtain a word time length of 0.87.
[0043]
In the candidate / speech sequence / phoneme sequence similarity calculation unit, the normalized phoneme sequence of the first candidate obtained by the word matching unit and the optimal phoneme sequence obtained by the speech sequence matching unit are normalized and edited between the sequences. Calculate the distance. For example, in the examples shown in FIGS. 3 and 6, the editing distance of the phoneme sequences / kobajasi / and / obajasi / is normalized by the sum of the sequence lengths to obtain 0.07.
[0044]
For the four scales obtained as described above, the unregistered word determination unit obtains an appropriately weighted sum of these scales, and determines the magnitude by a threshold to determine whether or not the utterance is an unregistered word. Perform That is, the determination is performed according to the equation shown in FIG. In FIG. 8, CM1 to CM4 denote the inter-candidate score difference, the candidate phoneme sequence similarity, the candidate / speech sequence score difference, and the candidate / speech sequence / phoneme sequence similarity, respectively. The weight for each scale, θ, means a threshold.
[0045]
The weight for each scale used here is obtained in advance by a statistical method. That is, for each of a large number of cases of registered word utterances and unregistered word utterances, the above four scales are obtained, and the relationship between the four scales and the registered word utterances and unregistered word utterances is analyzed by a linear discriminant method. To find the weight.
[0046]
(effect)
Next, the effect of the unknown utterance detection method based on the present embodiment will be experimentally shown in comparison with the conventional method.
[0047]
Generally, there are two types of errors in such a detection problem. That is, a detection omission error and a source error in which an error that should not be detected is detected. These two errors are in a trade-off relationship, and it is known that if one error is reduced, the other increases. Therefore, such a problem is compared by using two scales as shown in FIG. Here, a high unknown utterance recall means that there are few detection omission errors, and a high unknown utterance relevance means that there are few source errors. It is desirable that both are high.
[0048]
In the experiment described below, the change of the unknown utterance recall rate and the unknown utterance matching rate by each method is examined when the determination threshold is increased or decreased when there are 100 unknown names. The following three methods are used for comparison.
[0049]
(1) Unknown utterance judgment based only on the difference between the recognition likelihood and the reference likelihood
(There is no restriction on the connection of voice segments)
(2) Unknown utterance judgment based only on the difference between the recognition likelihood and the reference likelihood
(Introduce transition probability for concatenation of voice segments)
(3) Unknown utterance determination using the four scales described in the above embodiment together
FIG. 10 shows the result. In FIG. 10, the unknown utterance matching rate is plotted on the horizontal axis, and the unknown utterance recall rate is plotted on the vertical axis. The higher the two, the better, and the curve in the figure can be said to have better detection performance as it goes to the upper right. From this result, it can be seen that the method of determining the unknown utterance using the four scales together with the method of determining the unknown utterance has higher detection performance than the method of determining the unknown utterance using only the recognition likelihood and the reference likelihood as in the related art. Is shown.
[0050]
(Embodiment 2)
The present embodiment relates to a speech recognition device including the unknown utterance detection unit according to the first embodiment. The present embodiment has a function of providing an easier-to-use voice recognition interface function by returning an optimal response result by simultaneously using a detection result of an unknown utterance together with a conventional voice recognition result. .
[0051]
FIG. 11 is a block diagram of a speech recognition device equipped with the unknown utterance detection device shown in FIG. A speech analysis unit 20, a speech piece pattern storage unit 21, a word dictionary storage unit 22, a word matching unit 23, a transition probability storage unit 24, a speech sequence matching unit 25, a candidate score difference calculation unit 26, which constitutes the unknown utterance detection device, The candidate / phoneme sequence similarity calculator 27, the candidate / speech sequence score difference calculator 28, the candidate / speech sequence / phoneme sequence similarity calculator 29, and the unknown utterance determiner 30 are the same as those in the first embodiment. Speech analysis unit 1, speech piece pattern storage unit 2, word dictionary storage unit 3, word matching unit 4, transition probability storage unit 5, speech sequence matching unit 6, candidate score difference calculation unit 7, candidate / phoneme sequence similarity calculation It has the same configuration as the unit 8, the candidate / speech sequence score difference calculation unit 9, the candidate / speech sequence / phoneme sequence similarity calculation unit 10, and the unknown utterance determination unit 11.
[0052]
However, the unknown utterance detection unit 30 outputs a continuous value indicating the likelihood of unknown utterance, instead of simply outputting the unknown utterance determination result as correct or not. Further, the present embodiment includes a recognition result output unit 31 that outputs a recognition result in consideration of the results of both the word matching unit 23 and the unknown utterance detection unit 30. The configuration of the speech analysis unit 20, the speech piece pattern storage unit 21, the word dictionary storage unit 22, and the recognition result output unit 31 by the word matching unit 23 is the same as that of a normal speech recognition device.
[0053]
In the present embodiment, for the input voice, the result of the unknown utterance likelihood or the unknown utterance likelihood of the utterance content is obtained from the unknown utterance detection unit 30 by the same steps as those in the first embodiment. At the same time, likelihood is given to each word obtained from the word matching unit 23, and a recognition result candidate is obtained from a result sorted by the magnitude of likelihood.
[0054]
The recognition result output unit 31 outputs an optimal response in consideration of the recognition result candidate and the result regarding the unknown utterance likelihood obtained by the unknown utterance detection unit.
[0055]
That is, if the likelihood of unknown utterance is high, the recognition result output unit rejects all the results obtained as recognition result candidates and outputs a result indicating that the result is rejected. If the likelihood of unknown utterance is moderately high, one or more candidates from the top of the recognition result candidates are output, and a signal indicating that the result is not sufficiently reliable is also given. Further, if the likelihood of unknown utterance is sufficiently low, one or more top candidates among the recognition result candidates are output.
[0056]
With the above-described configuration, for example, the following effects can be obtained when the voice recognition device of the present embodiment is mounted on a television receiver and program selection is performed by the voice input interface. That is, conventionally, when a user utters a word that is not registered in the recognition dictionary for speech recognition, such as a program name that is not being broadcast or a name of a broadcast station that cannot be received, if a user utters a word, the conventional method simply recognizes a recognition error. And gave the user distrust of not knowing what to say.
[0057]
However, when the user utters such an unknown word, the voice recognition device of the present embodiment can inform the user that no such program name or broadcast station name exists. Become. Also, in the case where the recognition result is ambiguous, in the related art, the processing may be continued in an ambiguous manner and the video may be switched to a program not desired by the user. Thus, the user can be notified of the fact, and the confirmation means can be presented, and then the program can be switched. Thus, it is possible to efficiently avoid problems caused by recognition errors that often occur in voice recognition.
[0058]
The same effect can be applied not only to a voice recognition device in a television receiver but also to a destination search function in a car navigation system, an automatic telephone number guidance system by voice, and the like.
[0059]
【The invention's effect】
As described above, the first invention of the present invention is a technique for detecting an unknown utterance in a speech recognition device.
Using not only a single judgment scale but also a plurality of judgment scales has an effect of detecting an unknown utterance with high accuracy.
[0060]
Further, the second invention has an effect of providing a user-friendly voice recognition interface by outputting an output of a recognition result in consideration of a result of the unknown utterance detection device according to the first invention. Have.
[Brief description of the drawings]
FIG. 1 is a block diagram of an unknown utterance detection device according to a first embodiment of the present invention.
FIG. 2 is a diagram showing an example of constructing a word pattern from a speech piece in the embodiment.
FIG. 3 is a diagram showing an output example of a list of words and likelihoods output by a word matching unit according to the embodiment;
FIG. 4 is a diagram showing an example of a phoneme 2 gram probability in the embodiment.
FIG. 5 is a diagram showing an example of phoneme transition likelihood in a word calculated based on a phoneme 2 gram probability in the embodiment;
FIG. 6 is a diagram showing an effect of introducing a transition likelihood in calculation of a reference likelihood in the embodiment.
FIG. 7 is a view showing a method for obtaining an editing distance between streams in the embodiment.
FIG. 8 is a diagram of an expression showing a rule for determining an unknown utterance from a scale regarding four unknown utterances in the embodiment.
FIG. 9 is a diagram of an expression showing an evaluation scale for showing an effect compared with the conventional method in the embodiment.
FIG. 10 is a view showing an experimental result showing an effect in the same embodiment as compared with the conventional method.
FIG. 11 is a block diagram of a speech recognition device according to a second embodiment of the present invention.
[Explanation of symbols]
1,20 Voice analysis unit
2,21 Speech piece pattern storage
3,22 Word dictionary storage
4,23 Word matching unit
5, 24 transition probability storage unit
6,25 Voice sequence matching unit
7,26 Candidate score difference calculator
8,27 candidate / phoneme sequence similarity calculator
9,28 Candidate / speech sequence score difference calculator
10,29 Candidate / speech sequence / phoneme sequence similarity calculator
11,30 Unknown utterance determination unit
31 Recognition result output unit

Claims (6)

Voice analysis means for analyzing the input voice and converting it into a sequence of feature parameters;
A recognition dictionary storage unit for defining a vocabulary to be recognized,
Voice model storage means modeling a standard voice pattern;
A word level matching means for constructing a vocabulary model defined in the recognition dictionary by using the model stored by the voice model storage means, and performing matching with an input voice;
Subword transition probability storage means for defining transition probabilities between subwords,
A sub-word level matching unit that connects the voice model stored by the voice model storage unit in consideration of the sub-word transition probability stored by the sub-word transition probability storage unit and performs collation with the input voice;
From the word level matching unit and the sub-word level matching unit, unknown utterance scale calculation unit means for calculating a plurality of unknown utterance scales,
An unknown utterance detection device for speech recognition, comprising: an unknown utterance determination unit that determines an unknown utterance based on the plurality of scales calculated by the unknown utterance scale calculation unit. 2. The unknown utterance detection device according to claim 1, wherein the unknown utterance scale calculating means calculates a difference between a likelihood of a word obtained by the word level matching means and a subword chain likelihood obtained by the subword level matching means. An unknown utterance detection device characterized by including a value calculated based on the utterance. 2. The unknown utterance detection device according to claim 1, wherein the unknown utterance scale calculating means includes an acoustic feature of the word of the first candidate obtained by the word level matching means and a subword obtained by the subword level matching means. An unknown utterance detection device including a value calculated based on a similarity between two acoustic features of a chain. 2. The unknown utterance detection device according to claim 1, wherein the unknown utterance scale calculating means is based on a difference between the likelihood of the word of the first candidate and the likelihood of the word of the lower candidate obtained by the word level matching means. An unknown utterance detection device characterized by including a value calculated by the above. 2. The unknown utterance detection device according to claim 1, wherein the unknown utterance scale calculation unit calculates an acoustic feature of the first-rank candidate word obtained by the word-level matching unit and an acoustic feature of a lower-rank candidate word, An unknown utterance detection device comprising a value calculated based on a similarity between the two. A speech recognition apparatus for recognizing an input speech by collating with a model corresponding to a vocabulary registered in a recognition dictionary, wherein the unknown speech detection apparatus according to any one of claims 1 to 5 is provided. A speech recognition device mounted on a tower and outputting a recognition result in consideration of an output result of the unknown utterance detection device.

Images