JP4259100B2 - Unknown speech detection device for speech recognition and speech recognition device - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a speech recognition method in a speech recognition apparatus.
[0002]
[Prior art]
Conventionally, in a speech recognition device, an acceptable speech recognition vocabulary is defined, and a recognition vocabulary most similar to the input speech is searched for and output as a recognition result. Therefore, even if the user utters outside the speech recognition vocabulary, the most similar vocabulary is selected from the speech recognition vocabulary, so the recognition result is incorrect. For this reason, it is necessary to determine whether a user's utterance is a word included in the speech recognition vocabulary, a word other than that, or an utterance, and a function of rejecting these unknown utterances.
[0003]
The method for rejecting such unknown utterances is to connect the HMMs (one of the models expressing speech patterns) of units shorter than words called subwords to construct a model for each word in the recognition vocabulary, and to input speech A word that gives the maximum likelihood is searched for, and this likelihood is set as the recognition likelihood (see Non-Patent Document 1, for example). Further, the maximum likelihood for the input speech is obtained from the model based on the concatenation of arbitrary subword HMMs corresponding to the arbitrary kana sequence, and this is used as the reference likelihood. The unknown utterance is detected and rejected by comparing the recognition likelihood and the reference likelihood obtained in this way.
[0004]
However, in such a method, there is no restriction on concatenation of arbitrary subword HMMs in the calculation of the reference likelihood, and the likelihood for a non-Japanese sequence is often selected as the maximum likelihood, and as a result, In comparison between the reference likelihood and the recognition likelihood, a sufficient unknown speech rejection effect could not be obtained. Moreover, in order to compare the concatenation of all subword HMMs, a large amount of resources was required in terms of processing complexity. For such problems, for example, in speech recognition methods, the accuracy of rejection of unknown utterances and the amount of processing are improved by introducing the affinity of connection between subword HMMs as a transition probability (for example, (See Patent Document 1).
[0005]
[Patent Document 1]
Japanese Patent Laid-Open No. 10-171489 [Non-Patent Document 1]
Watanabe et al., "Rejection of unknown utterances 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 as described above has the following problems.
[0007]
In other words, a model based on concatenation of subword HMMs can be regarded as a model that recognizes input speech as a kana sequence, but even if a transition probability is introduced as a constraining constraint between subword HMMs, this model may be generated. It is still difficult to say that the sequence is sufficiently consistent with that of the input speech. That is, the reference likelihood 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, the unknown speech detection apparatus 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 series of feature parameters, and a recognition dictionary that defines a recognition target vocabulary. Storage means;
A speech model storage means that models a standard pattern of speech and a vocabulary model defined in the recognition dictionary using the model stored by the speech model storage means, and a word that is collated with the input speech Level matching means, subword transition probability storage means for defining transition probabilities between subwords, and speech models 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 connecting and collating with input speech;
Based on the unknown utterance scale calculation means for calculating a plurality of unknown utterance scales from the word level matching section and the subword level matching section, and based on the plurality of scales calculated by the unknown utterance scale calculation section, And an unknown utterance judging means for judging.
[0009]
According to the above configuration, in the determination of an unknown utterance, it is possible to determine the possibility that the input speech is an unknown utterance from a plurality of viewpoints, and it is possible to show high unknown utterance detection performance.
[0010]
Further, the unknown utterance detection device according to the first invention is characterized in that, in the unknown utterance scale calculation means, the word likelihood obtained by the word level matching means and the 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, the word likelihood correction effect by the subword chain likelihood is obtained, and high unknown utterance detection performance is obtained.
[0012]
Further, the unknown utterance detection device according to the first aspect of the invention is characterized in that the unknown utterance scale calculation means obtains the acoustic features of the first candidate word model 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 characteristics of the generated subword chain model.
[0013]
According to the above configuration, it is possible to determine an unknown utterance focusing on the similarity between the acoustic features of the two models, and high unknown utterance detection performance can be obtained.
[0014]
The unknown utterance detection device according to the first aspect of the present invention is the unknown utterance scale calculation unit, wherein the difference between the likelihood of the first candidate word obtained by the word level matching unit and the likelihood of the lower candidate word is obtained. 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 wrong candidates are similar and can be obtained with a likelihood, and high unknown utterance detection performance can be obtained.
[0016]
Further, the unknown utterance detection device according to the first aspect of the invention is characterized in that, in the unknown utterance scale calculation means, the acoustic features of the first candidate word and the acoustic features of the lower candidate words obtained by the word level matching means. 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 unknown utterances focusing on acoustic similarity between models of candidate words, and high unknown utterance detection performance can be obtained.
[0018]
The speech recognition device according to the second invention is a speech recognition device that recognizes an input speech by collating with a model corresponding to a vocabulary registered in a recognition dictionary, wherein the unknown utterance detection device The recognition result is output in consideration of the output result of the unknown utterance detection device.
[0019]
According to the above configuration, the speech recognition device 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. Thus, it is possible to convey this to the user, and it is possible to provide a user-friendly interface in various voice recognition interfaces on which the voice recognition device is mounted.
[0020]
DETAILED DESCRIPTION OF THE INVENTION
Embodiments of the present invention will be described below with reference to the drawings.
[0021]
(Embodiment 1)
FIG. 1 shows a block diagram of an unknown utterance detection apparatus according to the present embodiment. In FIG. 1, reference numeral 1 denotes an acoustic analysis unit that performs A / D conversion on input speech and converts it into a time series of feature parameters. Reference numeral 2 denotes a speech segment pattern storage unit that stores speech segments of standard speech used for matching with feature parameters of input speech.
[0022]
Here, the speech segment is a VC pattern in which the latter half of the vowel section of speech is connected to the first half of the consonant section that follows it, and the CV in which the latter half of the consonant section is connected to the first half of the vowel section that follows it. It means a set of patterns. However, the speech piece is a set of phonemes that are roughly equivalent to one letter of the alphabet when Japanese is written in Roman letters, and a mora that is almost equivalent to one letter of one letter when Japanese is clearly marked. It may be a set, a set of subwords meaning a chain of a plurality of mora, or a mixed set of these sets.
[0023]
Reference numeral 3 in FIG. 1 denotes a word dictionary storage unit in which rules for combining the speech pieces to synthesize word patterns of the speech recognition vocabulary are stored. Reference numeral 4 denotes a word matching unit that compares an input speech 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 express the naturalness of connection as continuous values when speech pieces are arbitrarily combined. In this embodiment, a 2-gram probability of phonemes is used as the transition probability. The 2-gram probability of a phoneme means the probability P (y | x) that a phoneme y connects after the preceding phoneme x, and is obtained in advance using a large number of Japanese text data. However, the transition probability may be a 2-gram probability of mora, a 2-gram probability of subwords, or a 2-gram probability of a mixture of these, and may be a 3-gram probability in addition to the 2-gram probability.
[0025]
6 in FIG. 1 is obtained by calculating the likelihood of the input speech expressed as a time series of feature parameters in consideration of the transition probability, and a pattern formed by arbitrarily combining the speech segment patterns. It is a speech sequence matching unit that uses the maximum likelihood as a reference likelihood.
[0026]
7 is the difference between the likelihoods of the words with the highest value (first candidate) and the words with the second highest value (second candidate) among the likelihoods of each word calculated by the word matching unit. Is a score difference calculation unit between candidates that is calculated by normalizing with a word length.
[0027]
8 is a similarity calculation unit between candidate phoneme sequences 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 which normalizes and calculates the difference between the likelihood of the first candidate and the reference likelihood calculated by the speech sequence matching unit.
[0029]
10 is a candidate / speech sequence / phoneme sequence similarity calculation unit that calculates the acoustic similarity between the first candidate and the sequence determined as the optimum sequence by the speech sequence matching unit as a distance between the phoneme sequences. is there.
[0030]
11 shows the values obtained by the 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, this is an unknown utterance determination unit that determines whether or not the input speech is an unknown utterance.
[0031]
In the present embodiment, the above four scales are listed as scales used in the unknown utterance determination unit, but in addition to this, the likelihood of each word candidate itself and its distribution, and the locality within the word section It is also possible to use scales such as the amount of change in score and the duration information of phonemes constituting words. As a method for determining an unknown utterance based on a plurality of scales, a linear discriminant obtained in advance using a large number of recognition result cases is used in this embodiment. However, the use of so-called learning machines such as neural networks, decision trees, and SVM (support vector machines) is also effective.
[0032]
Next, an unknown utterance detection processing operation in the present embodiment will be described. The input speech is first analyzed by the speech analysis unit after A / D conversion, and converted into LPC vectors every 10 milliseconds. The LPC vector is a parameter that means a spectral envelope of a short-time spectrum of speech, and is used as a parameter that well represents the phonological characteristics of speech. In normal speech recognition methods, the time series of LPC vectors obtained from input speech at regular intervals is used as the feature vector of input speech and is matched with a word model obtained in advance, and is called the likelihood for each word. Find the score.
[0033]
In the present embodiment, a word model is created using a voice fragment pattern and a word dictionary. That is, based on the speech segment concatenation rule for synthesizing the word patterns stored in the word dictionary storage unit, the speech pattern stored in the speech segment pattern storage unit is connected to construct a word pattern. FIG. 2 shows an image of synthesizing the word pattern “Hachinohe” by concatenating speech segment patterns called CV / VC patterns used in the present embodiment.
[0034]
In the speech segment pattern, parameters indicating the standard distribution (assuming normal distribution) of the LPC vectors of each speech segment 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 expressing the speech segment pattern is stored in the speech segment pattern storage unit 2, and the rules for the transition between the HMM models are defined in the word dictionary storage unit 3, thereby 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 with a certain likelihood. Are calculated and sorted in descending order of likelihood. FIG. 3 shows an output example of words sorted in order of likelihood.
[0036]
In parallel with this, the speech sequence matching unit 6 also matches an arbitrary sequence of speech segments. In this method, speech segments are freely concatenated under certain restrictions, and a speech segment sequence closest to the input speech and its likelihood are calculated. At this time, if no restrictions are applied to the connection of the speech pieces, the calculation result is a non-Japanese sequence, and the likelihood is not sufficiently meaningful. Therefore, in the search process of the optimum speech segment sequence, the phoneme 2gram probability stored in the transition probability storage unit 3 is used as the speech segment selection and connection cost. For phoneme 2gram probabilities, the recognition task and a large amount of Japanese text in the same task are converted into phoneme sequences and used based on this.
[0037]
In FIG. 4, as an example of the phoneme 2gram probability, the probability that five vowels / a /, / i /, / u /, / e /, / o / respectively follow the preceding phoneme / k / is illustrated. In this example, it is shown that the vowel that is likely to follow after the consonant / k / is / a / and then / i /. The speech sequence matching unit 6 obtains a value obtained by adding the likelihood by the concatenated speech fragment pattern and the transition likelihood obtained by the logarithmic sum of the phoneme 2gram probabilities by weighting, and adopts the sequence having the highest value.
[0038]
Figure 5 shows an example of finding the transition likelihood for the sequence / kobajasi / from the phoneme 2gram probability. In addition, in FIG. 6, as an example showing the effect of introducing the transition likelihood, the speech sequence output by 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, the series that is far away from the `` Kobayashi '' called / pobaeasii / has a higher pattern likelihood than the similar / obajasi / By using the total likelihood considering 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. Based on this, calculation of various measures for determining unknown utterances is performed. Do.
[0040]
First, the inter-candidate score difference calculation unit 7 has the highest likelihood (the likelihood of the first candidate) and the second highest likelihood (the likelihood of the second candidate) among the likelihoods obtained by the word matching unit. Calculate the likelihood difference divided by the word length. For example, in the case of the result shown in FIG. 3, the difference in likelihood between “Kobayashi” and “Hayashi” is normalized by the word length to obtain 6.9.
[0041]
The candidate phoneme sequence similarity calculation unit 8 calculates the similarity of the first candidate phoneme sequence and the second candidate phoneme sequence. Here, the similarity between sequences uses a value obtained by normalizing the edit distance by the sum of the sequence lengths of the two sequences. The edit distance is the minimum when one element is edited and converted to the other, with the cost required for one element replacement (replacement), one element deletion (dropout), and one element addition (insertion) set to 1, respectively. This means the total cost when editing at the cost of. FIG. 7 exemplifies how to obtain the edit distance for two phoneme sequences / uenoeki / and / jenokii /. In accordance with such a method, the similarity calculation unit between candidate phoneme sequences, for example, in the case of the result as shown in FIG. 3, sets the edit distance 3 of “kobayashi (/ kobajasi /)” and “hayashi (/ hajasi /)” to each phoneme sequence. The value 0.21 divided by the sum of lengths 14 is output. In this embodiment, the value based on the edit distance is used as the similarity between sequences. However, it is also effective to use a distance between sequences in consideration of 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 example shown in FIGS. 3 and 6, 0.87 is obtained by normalizing the difference between the recognition likelihood 2055 of the first candidate and the reference likelihood 2014 by the word time length.
[0043]
In addition, the candidate / speech sequence / phoneme sequence similarity calculation unit normalizes editing between sequences of the first candidate phoneme sequence obtained by the word matching unit and the optimal phoneme sequence obtained by the speech sequence matching unit. Calculate the distance. For example, in the example 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 a sum that appropriately weights these scales, and determines whether it is an unregistered word utterance by determining the magnitude with a threshold value. To do. That is, the determination is performed according to the equation shown in FIG. In FIG. 8, CM1 to CM4 mean the score difference between candidates, the similarity between candidate phoneme sequences, the difference between candidates / speech sequence scores, the similarity between candidates / speech sequences / phoneme sequences, and w1 to w2 Weighting for each scale, θ means a threshold value.
[0045]
Also, the weighting for each scale used here is obtained in advance by a statistical method. In other words, for the many 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 discrimination method. Find the weight for.
[0046]
(effect)
Next, the effect of the unknown utterance detection method based on this embodiment will be experimentally shown in comparison with the conventional method.
[0047]
In general, there are two types of errors in such detection problems. That is, a detection omission error and a well-being error in which something 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, for such a problem, a comparison based on two scales as shown in FIG. 9 is performed. Here, a high unknown utterance recall rate means fewer detection errors, and a high unknown utterance match rate means fewer errors. Both of these are desirably high.
[0048]
In the experiment shown below, when there is an unknown person name of 100 words and the determination threshold value is increased or decreased, changes in the unknown utterance recall rate and the unknown utterance adaptation rate by each method are examined. The following three methods are compared.
[0049]
(1) Judgment of unknown utterance based only on difference between recognition likelihood and reference likelihood
(There are no restrictions on the connection of audio pieces)
(2) Judgment of unknown utterance based only on difference between recognition likelihood and reference likelihood
(Introducing transition probabilities for concatenation of audio fragments)
(3) FIG. 10 shows the result of the unknown utterance determination using the four scales described in the above embodiment together. In FIG. 10, the unknown utterance adaptation rate is taken on the horizontal axis, and the unknown utterance recall rate is taken on the vertical axis. The higher the both, the better. Therefore, it can be said that the curve in the figure shows better detection performance as it goes to the upper right. From this result, it can be seen that the method of determining unknown utterances using four measures together has higher detection performance than determining unknown utterances using only the recognition likelihood and the reference likelihood as in the prior art. Indicated.
[0050]
(Embodiment 2)
The present embodiment relates to a speech recognition apparatus in which the unknown utterance detection unit in the first embodiment is mounted. The present embodiment has a function of providing a voice recognition interface function that is easier for the user to use 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 apparatus on which the unknown utterance detection apparatus shown in FIG. 1 is mounted. A speech analysis unit 20, a speech segment 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 constitute an unknown utterance detection device, The candidate / phoneme sequence similarity calculation unit 27, the candidate / speech sequence score difference calculation unit 28, the candidate / speech sequence / phoneme sequence similarity calculation unit 29, and the unknown utterance determination unit 30 are the same as those in the first embodiment. Speech analysis unit 1, speech segment 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 The configuration is the same as that of 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 does not simply output the unknown utterance determination result as correct or incorrect, but outputs a continuous value indicating the likelihood of unknown utterance. 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. In addition, the structure by the recognition result output part 31 by the audio | voice analysis part 20, the audio | voice piece pattern storage part 21, the word dictionary storage part 22, and the word matching part 23 makes the structure similar to a normal voice recognition apparatus.
[0053]
In the present embodiment, the input speech obtains a result related to the known utterance or the unknown utterance regarding the utterance content from the unknown utterance detection unit 30 by the same steps as in the first embodiment. At the same time, a likelihood is given to each word obtained from the word matching unit 23, and a recognition result candidate is obtained from the result sorted according to the likelihood.
[0054]
The recognition result output unit 31 outputs an optimum response in consideration of the recognition result candidate and the result regarding the likelihood of unknown utterance obtained by the unknown utterance detection unit.
[0055]
That is, the recognition result output unit rejects all the results obtained as recognition result candidates when the probability of unknown utterance is high, and outputs a result indicating that the recognition is rejected. When the probability of unknown utterance is moderately high, one or more candidates are output from the top among the recognition result candidates, and a signal indicating that the result is not sufficiently reliable is also given. Further, when the likelihood of unknown utterance is sufficiently low, one or more top candidates are output from the recognition result candidates.
[0056]
With the configuration as described above, for example, when the speech recognition apparatus according to the present embodiment is mounted on a television receiver and program selection is performed using the speech input interface, the following effects can be obtained. In other words, if a user utters a word that is not registered in the recognition dictionary for speech recognition, such as the name of a program that has not been broadcast or the name of a broadcasting station that cannot be received in the past, it would simply be a recognition error in the past. And gave users distrust that they didn't know what to say.
[0057]
However, when the user utters such an unknown word, the speech recognition apparatus according to the present embodiment can notify the user that such a program name or broadcast station name does not exist. Become. Also, even if the recognition result is ambiguous, in the past, it may have been possible to continue the process while being ambiguous and switch the video to a program that the user does not want, but the recognition result is ambiguous according to this embodiment. It is possible to perform processing such as switching the program after notifying the user and presenting the confirmation means, and it is possible to efficiently avoid problems due to recognition errors that tend to 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 an unknown utterance detection method in a speech recognition apparatus,
By using not only a single determination scale but also a plurality of determination scales, there is an effect of detecting an unknown utterance with high accuracy.
[0060]
In addition, the second invention has the effect of providing a speech recognition interface that is more usable for the user by outputting the output of the recognition result in consideration of the 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 apparatus according to a first embodiment of the present invention. FIG. 2 is a diagram showing an example of constructing a word pattern from speech pieces in the embodiment. FIG. 4 is a diagram showing an example of output of a list of words and likelihoods output by a word matching unit. FIG. 4 is a diagram showing an example of phoneme 2gram probabilities in the embodiment. FIG. 6 is a diagram showing an example of a phoneme transition likelihood in a word calculated in FIG. 6. FIG. 6 is a diagram showing an effect of introducing a transition likelihood in calculation of a reference likelihood in the embodiment. FIG. FIG. 8 is a diagram showing a method for obtaining an edit distance between sequences. FIG. 8 is an expression diagram showing rules for determining unknown utterances from a scale related to four unknown utterances in the embodiment. FIG. 9 is a conventional method in the embodiment. The effect compared with FIG. 10 is a diagram showing an experimental result showing an effect compared with the conventional method in the embodiment. FIG. 11 is a diagram showing an evaluation scale for evaluating the speech recognition apparatus in the second embodiment of the present invention. Block diagram [Explanation of symbols]
1,20 Speech analysis unit
2,21 Voice segment pattern storage
3,22 Word dictionary storage
4,23 Word matching part
5,24 Transition probability storage
6,25 Speech sequence matching section
7,26 Candidate score difference calculator
8,27 Candidate / phoneme sequence similarity calculator
9,28 Candidate / voice sequence score difference calculator
10,29 Candidate / voice sequence / phoneme sequence similarity calculator
11,30 Unknown utterance determination unit
31 Recognition result output section

Claims (3)

Speech analysis means for analyzing the input speech and converting it into a series of feature parameters;
A recognition dictionary storage means for defining a recognition target vocabulary;
A voice model storage means that models a standard pattern of voice;
Vocabulary model defined in the recognition dictionary, constructed using the model stored by the speech model storage section, and word-level matching means for matching the input speech,
Subword transition probability storage means for defining transition probabilities between subwords;
The speech model stored by the speech model storage section, the linked consideration the stored sub-word transition probabilities by sub-word transition probabilities storing means, and a sub-word level matching means for matching the input speech,
(1) and likelihood of words obtained by the word-level matching means, first and unknown speech measure is a calculated value based on the difference between the time the Yu subword chain obtained by the sub-word level matching means,
(2) Calculated based on the similarity between the acoustic characteristics of the first candidate word obtained by the word level matching means and the acoustic characteristics of the subword chain obtained by the subword level matching means. A second unknown utterance scale that is a value;
(3) a third unknown utterance scale that is a value calculated based on the likelihood of the first candidate word obtained by the word level matching means and the likelihood of the lower candidate word;
(4) A fourth unknown value which is a value calculated based on the similarity between the acoustic feature of the first candidate word obtained by the word level matching means and the acoustic feature of the lower candidate word Utterance scale,
Means for calculating an unknown utterance scale for calculating
To the unknown utterance measure calculating section respectively calculated the four unknown utterance measure, the value obtained by adding the multiplied by the weights determined by the statistical approach, based on whether it meets a predetermined threshold value, the unknown speech An unknown utterance detection apparatus for speech recognition, comprising: an unknown utterance determination means for performing determination. Speech analysis means for analyzing the input speech and converting it into a series of feature parameters;
A recognition dictionary storage means for defining a recognition target vocabulary;
A voice model storage means that models a standard pattern of voice;
A vocabulary model defined in the recognition dictionary is constructed using the model stored by the speech model storage means, and a word level matching means for matching with the input speech;
Subword transition probability storage means for defining transition probabilities between subwords;
Subword level matching means for connecting the speech model stored by the speech model storage means in consideration of the subword transition probability stored by the subword transition probability storage means, and performing collation with the input speech;
(1) and likelihood of words obtained by the word-level matching means, first and unknown speech measure is a calculated value based on the difference between the time the Yu subword chain obtained by the sub-word level matching means,
(2) Calculated based on the similarity between the acoustic characteristics of the first candidate word obtained by the word level matching means and the acoustic characteristics of the subword chain obtained by the subword level matching means. A second unknown utterance scale that is a value;
(3) a third unknown utterance scale that is a value calculated based on the likelihood of the first candidate word obtained by the word level matching means and the likelihood of the lower candidate word;
(4) A fourth unknown value which is a value calculated based on the similarity between the acoustic feature of the first candidate word obtained by the word level matching means and the acoustic feature of the lower candidate word Utterance scale,
Means for calculating an unknown utterance scale for calculating
Based on whether or not a value obtained by adding a weight obtained by a statistical method to each of the four unknown utterance scales calculated by the unknown utterance scale calculation unit satisfies a predetermined threshold, Unknown utterance judging means for outputting a continuous value indicating
Based on the output result of the unknown utterance determination means,
When the probability of unknown utterance is high, reject the result obtained as a recognition result candidate, and output a result indicating that it is rejected,
The unknown time speech likeliness is medium, the recognition result and outputs at least the upper one candidate among the candidates, and outputs a result, which means that the output of the at least upper one candidate is not sufficiently reliable,
An unknown utterance detection apparatus for speech recognition, comprising: a recognition result output unit that outputs at least one upper candidate among recognition result candidates when the likelihood of unknown utterance is low. Speech analysis means for analyzing the input speech and converting it into a series of feature parameters;
  A recognition dictionary storage means for defining a recognition target vocabulary;
  A voice model storage means that models a standard pattern of voice;
  A vocabulary model defined in the recognition dictionary is constructed using the model stored by the speech model storage means, and a word level matching means for matching with the input speech;
  Subword transition probability storage means for defining transition probabilities between subwords;
  Subword level matching means for connecting the speech model stored by the speech model storage means in consideration of the subword transition probability stored by the subword transition probability storage means, and performing collation with the input speech;
(1) A first unknown utterance scale that is a value calculated based on a difference between a word likelihood obtained by the word level matching means and a subword chain likelihood obtained by the subword level matching means;
(2) Obtained by the word level matching means 1 A second unknown utterance scale that is a value calculated based on the similarity between the acoustic characteristics of the candidate words and the acoustic characteristics of the subword chain obtained by the subword level matching means;
(3) Obtained by the word level matching means 1 A third unknown utterance scale that is a value calculated based on the difference between the likelihood of the rank candidate word and the likelihood of the lower candidate word;
(Four) Obtained by the word level matching means 1 Of the fourth unknown utterance scale, which is a value calculated based on the similarity between the acoustic characteristics of the rank candidate words and the acoustic characteristics of the lower candidate words, the first unknown utterance An unknown utterance scale calculating means for calculating a scale and at least one unknown utterance scale from the second unknown utterance scale to the fourth unknown utterance scale;
  For each unknown utterance scale calculated by the unknown utterance scale calculation unit, a value obtained by adding a weight obtained by a statistical method satisfies a predetermined threshold value, and an unknown utterance is determined. An unknown utterance detection device for speech recognition, comprising: an unknown utterance determination means for performing speech recognition.

Images