JPH08241096A - Speech recognition method - Google Patents

Abstract

PURPOSE: To discard non-grammatical utterance during the course of a search process. CONSTITUTION: A partial hypothesis is formed by additionally connecting and branching phonemes in accordance with the grammer 41 of a tree structure and a score function gi(t) is determined by trellis calculation while collecting the partial hypothesis i, the corresponding HMM and input speeches. The max. value in the score of the nongrammatically formed partial hypothesis is simultaneously obtained as a reference score function go(t). The respective max. values of the respective differences between the forward heuristic function gΛ(t) and gi(t) and go(t) are respectively determined as evaluation values Si, So and the search is progressed by discarding the partial hypothesis in which Si-So is below the threshold.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a corresponding acoustic model for a number of partial hypotheses, which can be connected according to the control of a given grammar, of a phonetic unit such as a phoneme, a syllable, a half syllable, and a word. The present invention relates to a voice recognition method for collating an input voice and searching for a candidate close to an input voice.

[0002]

2. Description of the Related Art FIG. 3A shows a procedure of speech recognition processing using a phoneme as a unit of recognition. The input voice 11 is the analysis processing unit 12
Thus, the vector data of the feature parameter is converted into a time series, and the search processing unit 13 performs matching with the phoneme model 15 while using the constraint condition of the grammar 16. And
The phoneme sequence having the highest evaluation value is output as the recognition result 14.

As signal processing in the analysis processing section 12,
Often used is Linear Predictive Analysis (Linear P
(Reductive Coding, LPC), and the characteristic parameters are LPC cepstrum, LPC delta cepstrum, mel cepstrum,
There is logarithmic power. Probability as the phoneme model 15
Hidden Markov Model (Hidden Markov Model, modeled based on statistical theory)
(M method) is the mainstream. Details of this HMM are disclosed, for example, in "Speech Recognition by Probabilistic Model" by Seiichi Nakagawa, edited by The Institute of Electronics, Information and Communication Engineers.

The search processing unit 13 evaluates the plausibility of the partial hypotheses, which are phoneme sequences that are allowed to be connected by grammar, with the input speech, and expands the phonemes to the partial hypotheses one by one. While proceeding with the search. Here, the partial hypothesis represents a phoneme string connected according to the restriction of the order of phonemes shown in the grammar, and
Extension of a phoneme to a partial hypothesis means that one more phoneme is connected to the phoneme sequence of the partial hypothesis according to the grammar.

For each partial hypothesis, 1. Phoneme sequence, 2. 2. A score function that is a result of matching with an acoustic model by trellis calculation or the like; Three pieces of information, that is, an evaluation value indicating the plausibility of the partial hypothesis with respect to the input voice, are stored. When the identification number of the partial hypothesis is i and the time is t, the score function is expressed as g _i (t). In the search processing unit 13,
First, the first phoneme allowed by the grammar is expanded to a partial hypothesis, the HMM corresponding to the phoneme is collated with the vector data time series (input speech) of the analyzed characteristic parameter, and each partial hypothesis i Score function g _i (t) at time t
Ask for. The trellis method and the Viterbi method are available as the matching method with the HMM. For more details, see the Institute of Electronics, Information and Communication Engineers, edited by Seiichi Nakagawa, "Speech recognition by probabilistic model".
Is disclosed in. The evaluation value of the partial hypothesis i is obtained from the score function g _i (t) by the method described later, and the phoneme sequence, the score function g _i (t), and the evaluation value are recorded for this partial hypothesis. And every time the phoneme is expanded thereafter,
The search process proceeds while obtaining the evaluation value for the partial hypothesis. In addition, when two or more types of phonemes can be expanded from the phoneme string of the partial hypothesis due to grammatical constraints, the original partial hypothesis is duplicated by the number of kinds of phonemes that can be expanded, and each phoneme is expanded. And calculate evaluation values for them. For the partial hypothesis in which the phoneme cannot be extended by the grammar, the phoneme expansion is terminated as the hypothesis that the phoneme sequence is accepted as the grammar. When the phonemes cannot be expanded for all the partial hypotheses, it means that all phoneme strings allowed as grammar are matched with the input speech.
The search process 13 ends. Recognition result 1 of the hypothetical phoneme sequence with the highest evaluation value at that time or the corresponding word or sentence
Output as 4.

[0006] As described above, a search method for extending phonemes of partial hypotheses so that the number of phonemes of all partial hypotheses (phoneme strings) in the search process is equal is called a horizontal search method. When the horizontal search method is actually performed, calculation is performed for partial hypotheses corresponding to all phoneme sequences permitted by the grammar, and a large number of partial hypotheses must be calculated, which requires a lot of processing time. And Therefore, in the process of expanding phonemes to partial hypotheses, a method is often adopted in which only partial hypotheses that are likely to be the final recognition result are left and other partial hypotheses are discarded. Specifically, the evaluation value of the partial hypothesis determines whether or not the partial hypothesis remains. As a determination method, a method of leaving a certain number of partial hypotheses in order from the one with the highest evaluation value of the partial hypothesis, a method of providing a threshold value of the evaluation value of the partial hypothesis, and leaving only the partial hypothesis higher than the threshold value,
A combination of both methods is used. In such a horizontal search method, a method of leaving only a partial hypothesis with a certain probability and discarding the other partial hypotheses under certain conditions to perform a search is called a beam search method.

Explaining the above concretely, for example, a search process using HMM is performed on a grammar expressed by a tree structure as shown in FIG. 3B. Assuming that the processing up to the fourth phoneme has been completed and the case where the fifth phoneme is expanded will be described, in FIG. 3B, the partial hypothesis expanded from the first phoneme # to the fourth phoneme is “# i k a ”,“ # i k i ”,“ #
i m i ”. here," It is assumed that "" is a symbol indicating a phoneme division, and phoneme # indicates silence.

The first phoneme starts with # and extends to the fourth phoneme.
One partial hypothesis, "# i k i ”in FIG.
As can be seen from B, there are three types of phonemes as the fifth phoneme.
k, o, and m can be expanded. Also, the first phoneme is from #
Another partial hypothesis that started and expanded to the 4th phoneme,
"# i k "a" is the second phoneme as the fifth phoneme.
m and n can be expanded. In addition, the partial hypothesis “# i m
i ”is completed in the 4th phoneme, and the phoneme is expanded.
Not.

Expected for each phoneme depth in tree structure grammar
The same sound is generated in the beam search that discards the missing partial hypothesis.
For partial hypotheses with prime numbers,
Only the partial hypotheses with good evaluation values are left under certain conditions.
Here, as a certain condition, the top two with the highest evaluation value are
Only the partial hypothesis remains. As mentioned above,
The partial hypothesis expanded to 5 phonemes is "# i k i
o ”,“ # i k i k ”,“ # i k i
m ”,“ # i k a m ”,“ # i k a n "
There are 5 types, and the evaluation value of each partial hypothesis is in this order.
If it is high, the top two partial hypotheses are "# i k
i o ”and“ # i k i k ”only expands the next phoneme
It remains as a partial hypothesis that can be extended and the other partial hypotheses are abolished.
Abandon.

In this way, phonemes are expanded to the partial hypothesis,
The partial hypotheses to be left are limited under certain conditions, the phonemes are further expanded to the remaining partial hypotheses, and the same processing is continued until the phonemes cannot be expanded for all the partial hypotheses. Then, all partial hypotheses whose phonemes cannot be expanded, that is, the evaluation values of the hypotheses are compared, and the hypothesis with the highest evaluation value is output as the recognition result.

As a method of obtaining the evaluation value of the partial hypothesis from the score function g _i (t) of the partial hypothesis i, a forward heuristic function g ^ (t) common to all partial hypotheses estimated forward from the beginning of speech is used. The difference between this and the score function g _i (t) of the partial hypothesis i is obtained, and the value corresponding to the maximum value of the difference with respect to the time t is used as the evaluation value S _i of the partial hypothesis i. (For details of this method, see, for example, “Yoshiaki Noda, Shigeki Sagayama,” A ^* using forward likelihood ^.
HMM-LR Speech Recognition by Beam Search "IEICE Technical Report, Speech, SP94-23,199
4 "and" Japanese Patent Application No. 6-133339, acoustic recognition method ".

A specific example of how to obtain the evaluation value of this partial hypothesis and
And extended to the 4th phoneme, the partial hypothesis "# i k
Diagram of how to calculate evaluation value when phoneme o is extended to "i"
4 will be described. Fig. 4 shows matching between phoneme strings and input speech.
The score function obtained by performing the trellis calculation is
A three-dimensional diagram with three axes: elementary sequence, input speech, and score
Therefore, the curve 31 shows the partial hypothesis “# i k
i "score function, g_i4(T) at time t₁Sono
Score value g_i4(T₁) Indicates that the input voice is at time t₁until
That this partial hypothesis (phoneme sequence) was uttered in the shortest time
It is a score that indicates the plausibility of assuming
t₂Score in_i4(T₂) Indicates that the input voice is at time t₂Well
And suppose that this partial hypothesis was uttered in the longest time
It is a score that shows the plausibility of when
t₁, T₂And the duration of the phoneme o, the time t₃Decided
And the input voice will be phoneme by that time in that section.
Column "# i k i "o" is uttered respectively
Curve 3 is what made the plausibility (score) of when
2, that is, the curve 32 is the partial hypothesis “#” of the input voice. i
k i Score function g for "o"_i5(T).
That is, the partial hypothesis "# i k The score function 31 of “i” is
Is calculated as the initial value of the likelihood at each time.
Then, the trellis calculation is used to accumulate the scores of phoneme o at each time.
Then, "# i k i Find the score function 32 of "o"
Meru.

The trellis calculation is an HMM showing an acoustic model.
And the vector time-series data of the characteristic parameters obtained by analyzing the input speech, and the probability calculation of the vector time-series data is performed for all transitions of the HMM that reach the final state of the HMM at time t. The probability value at t can be obtained. Here, the log value of the probability value is used as the score (likelihood).

Next, in order to obtain the evaluation value of the partial hypothesis, a forward-looking statement obtained by common grammar (without grammatical restrictions and allowing expansion to any phoneme) common to each partial hypothesis estimated from the beginning of the speech When the heuristic function g ^ (t) is obtained and is subtracted from the score function g _i (t) of this partial hypothesis as in the following formula (1), the maximum value S _i is obtained.
S _i indicates the plausibility of the partial hypothesis i,
By setting this as the evaluation value of the partial hypothesis i, the evaluation value of the partial hypothesis that is normalized with respect to time can be obtained.

S _i = max {g _i (t) −g ^ (t)} (1) max is the maximum in {} for each t Note that when a search is performed without grammar, an evaluation value close to the correct answer is obtained. However, since the number of partial hypotheses is remarkably large and the number of evaluation values that are almost the same is large and selection becomes difficult, as described above, the search is performed under the constraint of grammar.

[0016]

In the voice recognition, by reducing the search processing amount, the recognition processing time is shortened and the voice recognition in actual use becomes easy to use. Further, by reducing the amount of search processing, it becomes possible to practically operate voice recognition even on a computer having a low processing capacity. In order to reduce the amount of search processing, it is necessary to discard partial hypotheses that are unlikely in the search process and reduce the number of partial hypotheses to be expanded. However, in the conventional beam search that holds a certain number of partial hypotheses with high evaluation values, there is a partial hypothesis with a small evaluation value among partial hypotheses that hold a certain number of hypotheses, that is, a partial hypothesis that cannot be a plausible recognition result. However, the partial hypothesis is not discarded, and wasteful processing is performed. Further, in the beam search in which a threshold is set and a partial hypothesis whose evaluation value is higher than the threshold is retained, partial hypotheses with a small evaluation value are discarded, but in general, the evaluation value is the number of recognized vocabularies, speakers, Since it is greatly affected by the input speech length, it is difficult to set a threshold value that can effectively discard the partial hypothesis without dropping the correct partial hypothesis.

That is, the evaluation value calculated by the conventional method is effective for comparing the partial hypotheses, but is greatly affected by the number of recognition vocabularies, the speaker, and the input speech length, and therefore its absolute value It is difficult to evaluate the partial hypothesis using the value itself.

[0018]

According to the present invention, the utterance content of the input speech is calculated in the search process, that is, for each depth of a phonetic unit (phoneme, syllable, semi-syllable, word, etc.) in the tree structure grammar. Is estimated as a reference evaluation value by assuming that the answer is correct, and the evaluation value obtained by performing a collation with an acoustic model by concatenating speech units under the constraint of the conventional grammar is the reference value. The evaluation value is normalized, and the partial hypothesis whose normalized evaluation value is less than or equal to the threshold value is discarded.

By the above normalization, the influence of the number of recognized vocabulary, the speaker, the input speech length, etc. is removed from the evaluation value of the partial hypothesis,
In the search process, you can surely discard the partial hypothesis that has no prospect,
By increasing the search efficiency and using this normalized evaluation value, the search processing amount can be reduced.

[0020]

Embodiments of the present invention will be described below. The input speech is analyzed in the same manner as in the past, and vector data of time-series feature parameters is obtained. As the search processing, a phoneme is used as a unit of speech extended to the partial hypothesis, a phoneme-synchronized beam search in which the number of phonemes in each partial hypothesis is constant, and an acoustic model is an HMM.
In this case, an embodiment to which the present invention is applied will be described with reference to FIG. The phoneme extension processing unit 42 using the constraint condition of the grammar 41
The phoneme is expanded to the partial hypothesis i by, and the trellis calculation processing unit 43
Matches the HMM corresponding to the phoneme sequence with the input voice. The evaluation value calculation processing unit 47 obtains the evaluation value S _i of the partial hypothesis i from the obtained score function g _i (t) of the partial hypothesis i. In the conventional method, a beam search is performed in which a certain number of partial hypotheses having a high evaluation value S _i of the partial hypothesis _i are held and then discarded. However, in the present invention, the score function calculation processing unit 4
Calculated by the score function g ₀ (t) the method described below for the reference evaluation value in 5, we obtain the reference evaluation value S _O in the evaluation value calculation processing unit 48 in the same manner as described above. Then 'seek, the difference S _i the difference evaluation value S _i and the reference evaluation value S ₀ of the partial hypotheses i (partial hypotheses i normalized evaluation value S _i) of' what is large, as expected with no partial hypotheses Discard and proceed with the search.

Explaining specifically with reference to the example of FIG. 3B, the fourth tone
The partial hypothesis obtained by extending the phoneme from the elementary partial hypothesis is "# i
k i o ”,“ # i k i k ”,“ # i
k i m ”,“ # i k a m ”,“ # i k
a There are five types of "n", and each partial hypothesis is a partial hypothesis.
i and the evaluation value of the partial hypothesis i is S_iAnd the standard evaluation value S
₀Then, the normalized evaluation of the partial hypothesis i is performed by the following equation (2).
Value S_i′ Is obtained.

S_i′ = S_i-S₀ (2) When the input voice is actually generated as "Ikioi"
Then, "# i k i The partial hypothesis of "o" is the most correct answer
The evaluation value becomes high soon. Also,"# i k a m ”
As for the partial hypothesis that is far from the correct answer, its evaluation value
Is getting smaller. The reference evaluation value depends on the content of the input voice.
Estimated evaluation value assuming that the answer is correct.
For example, since it was obtained by non-grammar, grammatical
There are no restrictions, and all combinations of acoustic models are allowed
The same phoneme sequence as the content of the input speech
Or, the phoneme sequence of the phoneme sequence
The combination should have the highest evaluation value.
This standard evaluation value is "# i k i part of "o"
It is close to the theory's evaluation value. Therefore, the normalized evaluation value
S_iThe value of ′ is not close to 0 for the partial hypothesis that is close to the correct answer.
Therefore, the partial hypothesis far from the correct answer has a large negative value.
Normalized evaluation value S_iThis tendency of ′ is S₀And S _iAre both the same
S because it is made from one input voice₀And S_iSpeakers including
Characteristics are subtracted from the normalized evaluation value and removed.
Hard to depend on. S for the same reason_iSaid tendency of input sound
It does not depend on the voice length. Also, keep the number of partial hypotheses constant.
In the beam search,
It is necessary to change the number of theories, but the evaluation value itself is recognized
Normalized evaluation value because it does not change even if the number of vocabularies changes
S_i′ Has little influence on the number of recognized vocabularies.

When a partial hypothesis having a low normalized evaluation value S _i ′ is discarded in the beam search, a threshold value L is set and S _i ′ <L.
However, if L is a constant number or a value depending on the time length of the partial hypothesis, for example, if the time length of the partial hypothesis is long, a large negative value is set in the above example. Good. Equation (1) is used as the calculation method in the evaluation value calculation processing units 47 and 48 in FIG.
When (t) is equal to the score function g ₀ (t) for the reference evaluation value, the normalized evaluation value S _i ′ can be obtained using the following equation (3). g _i (t) is a score function of the partial hypothesis i, and g ₀ (t) is a score function for the reference evaluation value. According to the equation (3), the amount of calculation processing for the normalized evaluation value S _i ′ can be significantly reduced.

[0024] _{S i '= max {g i} (t) -g 0 (t)} (3) max score function g ₀ for a reference evaluation value S _O those in {} for each t is maximum ( The method of obtaining t) is shown below. <Score function calculation method 1 for reference evaluation value> Each phoneme HMM usually has about three states, and in each state, there is an output probability density distribution of the weighted sum of a plurality of probability density functions. Here, the characteristic parameter of the input speech at each time is given to all output probability density distributions, the highest output probability density value is selected, and the maximum likelihood for each time, which is the logarithm thereof, is obtained. The cumulative value of this maximum likelihood over time is calculated, and this is used as the score function for the reference evaluation value. Oτ
Is a characteristic parameter at time τ, and p _j (Oτ) is an output probability density value obtained by giving the characteristic parameter to the output probability density distribution j, g ₀ (t) is given by the equation (4).

G ₀ (t) = Σ maxp _j (Oτ) (4) Σ is from τ = 0 to t, max is the maximum value for all _j in p _j (Oτ), usually from one HMM to another HMM
The transition from the end state of one HMM to another HM
The transition is made under the condition that the transition to the initial state of M is made, but this score function eliminates the transition condition and the constraint of grammar, and from any state of any HMM to any of any HMM. The score function when the Viterbi calculation is performed with the transition probability being 1 and the transition probability being 1 is shown. As the search progresses, most of p _j (Oτ) becomes
Since the trellis calculation is performed in the search process, the result can be used and the amount of calculation is small.

<Calculation Method 2 of Score Function for Reference Evaluation Value> In the calculation method 1, the maximum value of the output probability density values obtained from the output probability density distributions of all the states of all HMMs is obtained. In this calculation method 2, the maximum value of the output probability density values for all the output probability density distributions that have been calculated by trellis calculation up to the present in the course of the search processing is obtained. For example, as shown in FIG. 2, when the power density distributions p ₁ , p ₂ , p ₃ ... Of each state of each HMM are plotted on the vertical axis and the time t is plotted on the horizontal axis, in the example of FIG. In this example, the silent length for which the output probability density value is predicted for the output density distribution of each state of the HMM is calculated from time 0 to time 3 (the area in which this calculated value is filled is indicated by 51), and the shortest silence The output probability density value for the output density distribution of each state of the HMM of the next phoneme i is calculated from the time 2 after the end time 1 of 1 to the time 4 after the end time 3 of the longest silence. An area in which the calculated value is filled is indicated by 52. Similarly, the output probability density value of each HMM state of the phoneme k is calculated as a region 53 in FIG. Although such calculation is advanced by the search, each time 0, 1, in FIG.
The maximum value of the calculated output probability density values in 2, ... Is obtained. This maximum value is sequentially added to obtain g ₀ (t).
In this way, since the score function g ₀ (t) is calculated from the output probability density distribution which is constrained by the grammar in the search process, a score function closer to the actual grammar can be obtained. Moreover, since only the output probability density value already calculated in the trellis calculation is used, the calculation for the score function g ₀ (t) is hardly necessary. Even in such a calculation method, it is considered that the output probability density value of the part not restricted by the grammar is smaller than that obtained by the trellis calculation, and the number g ₀ (t) is correctly estimated.

<Score function calculation method 3 for reference evaluation value> As described in the explanation of the horizontal search method, phonemes are expanded to partial hypotheses, and a score function is obtained by performing collation such as trellis calculation. . In this case, the phoneme is expanded in a grammar that can expand any phoneme to each partial hypothesis, that is, without grammar, and at each time of the score function obtained by matching the corresponding acoustic model with the input speech. The maximum value of is the score function for the reference evaluation value. In this case, the transition constraint of the HMM is left. This method has a stronger grammatical force than the above two methods, and by using this, a highly accurate normalized evaluation value S _i ′ can be obtained. The amount of calculation also increases.

<How to calculate score function for reference evaluation value
Method 4> In the calculation method 3 of the score function for the reference evaluation value
And not a grammar that can expand any phoneme,
A grammar that allows only the phoneme array structure peculiar to Japanese
Perform the calculation and use the obtained score function as a positive heuristic.
Let it be a tick function. Allow Japanese phoneme array structure
The phoneme sequence is, for example, "o m o sh i r o
i "and" s u t o r a i k u "
In general, the restriction that consonants do not come after
are doing. "S t r ai Phoneme chain "k"
Satisfies the phoneme array structure in English, but Japanese sounds
It is not a prime array structure.

G ₀ in calculation method 3 and calculation method 4
It can be said that the grammar for expanding the phonemes when calculating (t) includes the grammar for creating the partial hypothesis for the search. <Calculation Method 5 of Score Function for Reference Evaluation Value> In most cases, the final correct partial hypothesis has a larger score function than the other partial hypotheses. Therefore, the score functions g ₁ (t) of all partial hypotheses calculated in the search process,
The maximum value of g ₂ (t), g ₃ (t), ...
₀ (t). Expressed as an expression, it is as follows.

G ₀ (t) = max g _i (t) (5) max is the largest of all _i in g _i (t) This calculation method requires almost all the calculation amount for g ₀ (t). do not do. <Score Function Calculation Method 6 for Reference Evaluation Value> The calculation of the score function g ₀ (t) for obtaining the reference evaluation value S _O does not need to identify phonemes, and it is sufficient if the score can be obtained. Therefore, it is not necessary to use the HMM for each phoneme, and the acoustic model 1 for recognition as shown by the dotted line in FIG.
An acoustic model 46 different from 5 may be used. As the acoustic model 46, for example, even if one or several acoustic models are provided, by providing a large number of states, an acoustic phenomenon including a recognition target. It may be configured to express all, and this one acoustic model is repeatedly used,
Alternatively, in the case of several acoustic models, g ₀ (t) may be obtained by arbitrarily selecting and concatenating these acoustic models and matching them with the input voice to find a plausible one.

Description of Partially Modified In the above, a forward-looking heuristic function was obtained in order to obtain an evaluation value. For example, “Minami et al.“ Large word continuous speech recognition algorithm for number guidance ”, IEICE Transactions A Vol.J77-A, No. 2, pp. 190-
197.1994 ”, an estimated likelihood function h ^ (t) common to all hypotheses estimated backward from the end of speech is obtained, and this is calculated as a score function g _i (t).
To the evaluation value S _i . Furthermore, the present invention is applicable not only to speech recognition in units of phonemes, but also to recognition in units of syllables, semi-syllables, words, and the like.

Experimental examples are shown below. Phoneme balance 216
In the word recognition targeting the odd-numbered 108 words, the speech data of the odd-numbered 108 words as the vocabulary and the even-numbered 108 words as the words outside the vocabulary were given and evaluated. During the search, the recognition rate for word recognition in the vocabulary is the overall recognition rate as the value for evaluating the performance of discarding, and the rate of misrecognition in the word recognition in the vocabulary is “missing recognition result”, and the rate outside the vocabulary. The rate at which the recognition result was not rejected in the word recognition was defined as the false acceptance rate, and the average of the false rejection rate and the false acceptance rate was defined as the false positive rate. That is, it is considered that the discarding performance is good when the false determination rate can be suppressed to a low level while maintaining the recognition rate.

The above evaluation was carried out by changing the strength of disposal. For this, as the threshold L for rejecting the partial hypothesis, θ · t proportional to the time t was used, and the strength of discard was changed by changing the value of θ. The larger the value of θ, the stronger the disposal. As voice data, four speakers of MAU, MHT, FAF, and FSU in the voice database of ATR were used for evaluation. As an experimental system, HMM
-Using the LR voice recognition server. However, as the acoustic model, an environment-independent mixed continuous distribution HMM for unspecified speakers with the number of states 3 and the number of mixture distributions 4 and the number of phoneme models 54, which was learned from 9600 sentences of the ASJ continuous speech database was used. In this experiment, the chain of arbitrary phoneme combinations was used as the hypothesis for the reference evaluation value, and its likelihood function was used as the forward heuristic function.

FIG. 5 shows changes in the recognition performance and the discard performance when the strength of the dynamic discard in the case of the speaker MHT is changed.
The recognition processing time and the number of times of matching in the figure show the values normalized using the respective values in the full search. As you can see from the figure,
For example, as can be seen from around θ = 0, there is an effect of discarding while maintaining the recognition rate. Moreover, the number of times of matching is suppressed, and it can be seen that unnecessary partial hypotheses are rejected. However, since the vocabulary was small in this word recognition experiment, the amount of calculation for obtaining the heuristic function was relatively large, and the overall recognition processing time was almost the same as when performing a full search. However, when a beam search with a certain number of partial hypotheses is performed using this heuristic function, a recognition processing time of about 1.2 times that of the full search is required to obtain the same recognition rate. Therefore, even in the experiment under these conditions, the method of the present invention has a discarding function and the recognition processing time is shorter than that of the beam search with a constant number.

[0035]

As described above, the absolute value of the evaluation value of the partial hypothesis depends on the speaker, the number of recognized vocabularies, and the length of the input speech. Since it is normalized by the evaluation value, a normalized evaluation value that does not depend on the speaker, the number of recognized vocabularies, and the input speech length can be obtained, and it is possible to effectively discard unexpected partial hypotheses in the search process. . As a result, the threshold value for the normalized evaluation value is the same value, the voice recognition can be used for various purposes, and the setting load on the user can be reduced.

Further, when the input voice has a content that the grammar does not allow, the conventional search outputs the wrong result which is the closest candidate in the grammar, so that the utterance error of the user and the voice recognition error. Could not show distinction from cognition. However, in this case, in the present invention, all the partial hypotheses are discarded during the search process, and no recognition result is obtained, and the user can be notified of the utterance error. It is important for practical speech recognition to detect and show the utterance error of the user at an early stage.

The effects of the method of the present invention are listed below. -Effectively discarding partial hypotheses that are unlikely in the search process. -Since the threshold that must be set does not depend on the speaker, the number of recognized vocabularies, and the length of input speech, the burden of setting on the user can be reduced. -If the input voice has a content that the grammar does not allow, it can be detected that recognition cannot be performed early in the search process, and the user's utterance error can be notified.

[Brief description of drawings]

FIG. 1 is a diagram showing an example of a method of obtaining a normalized evaluation value of a partial hypothesis, which is a main part of the method of the present invention.

FIG. 2 is a diagram showing an example of an output probability density value obtained by a trellis calculation for explaining a score function calculation method 2 for a reference evaluation value.

FIG. 3A is a diagram showing a process of a speech recognition method in which a phoneme is a unit of recognition, and B is a diagram showing a grammar expressed by a tree structure.

FIG. 4 is a diagram showing a score function obtained as a result of trellis calculation.

FIG. 5 is a diagram showing the results of experiments conducted on the method of the present invention.

Claims (12)

[Claims] 1. Based on a grammar of a tree structure composed of voice units, one or a plurality of hypotheses regarding the generated content of an input voice is generated by gradually concatenating and branching voice units, and at that time, For each depth of the voice unit of the above tree structure, based on the acoustic model, the plausibility of the partial hypotheses up to that point for the input speech is evaluated to obtain the partial hypothesis evaluation value, and the In a voice recognition method for obtaining a recognition result from the likelihood, for each depth of the voice unit of the tree structure, the evaluation value when the utterance content of the input voice is assumed to be the correct answer is estimated as a reference evaluation value, A speech recognition method characterized by normalizing an evaluation value of a partial hypothesis of a corresponding depth with a reference evaluation value and discarding a partial hypothesis having a normalized evaluation value equal to or less than a threshold value. 2. A hypothesis related to the utterance content of the input voice is generated successively and continuously by adding voice units based on a grammar including the grammar, and the input voice is the sound corresponding to the partial hypothesis. 2. The voice recognition method according to claim 1, wherein the reference evaluation value is obtained by obtaining a score function by collating with a model. 3. A score function is obtained by collating the input speech with a partial hypothesis corresponding to at least one reference evaluation value acoustic model expressing an acoustic phenomenon including a recognition target, and obtaining the score function. The voice recognition method according to claim 1, wherein an evaluation value is obtained. 4. The partial hypothesis evaluation value is obtained by collating the input speech with an acoustic model corresponding to the partial hypothesis to obtain a score function. Speech recognition method. 5. The speech recognition method according to claim 4, wherein the acoustic model is a hidden Markov model. 6. A maximum value of all output probability density values of the hidden Markov model is calculated for each time, and the maximum value is accumulated to calculate a score function for obtaining the reference evaluation value. The voice recognition method according to claim 5. 7. A score function for selecting the maximum value among the output probability values of hidden Markov calculated for obtaining the partial evaluation value at each time and accumulating the maximum value to obtain the reference evaluation value. The voice recognition method according to claim 5, wherein 8. The voice recognition method according to claim 2, wherein the grammar including the grammar allows any combination of acoustic models corresponding to voice units. 9. The speech recognition method according to claim 8, wherein a constraint of a phoneme array structure peculiar to Japanese is used for a combination of acoustic models corresponding to the speech units. 10. The input speech is collated with the acoustic model corresponding to a partial hypothesis to obtain a score function to obtain the partial hypothesis evaluation value, and the reference evaluation value is the maximum value of the score function at each time. The speech recognition method according to claim 1, wherein the speech recognition method is obtained by determining 11. The evaluation value of the partial hypothesis is calculated as a forward heuristic function common to all partial hypotheses, the difference between the score function of each partial hypothesis and the forward heuristic function is calculated, and the maximum value of the difference is associated. The voice recognition method according to any one of claims 4 to 10, wherein the voice recognition method is obtained as a value. 12. The speech recognition method according to claim 11, wherein a score function obtained for obtaining the reference evaluation value is used as the forward-looking heuristic function.