JP3104900B2 - Voice 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 an acoustic model corresponding to a number of partial hypotheses which can be connected under the control of a given grammar, such as phonemes, syllables, semi-syllables, words, etc. The present invention relates to a voice recognition method for searching for a candidate that is close to the input voice by comparing the input voice with the input voice.

[0002]

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

[0003] As signal processing in the analysis processing section 12,
Often used is linear predictive analysis (Linear P
and the characteristic parameters include LPC cepstrum, LPC delta cepstrum, mel cepstrum, and the like.
And logarithmic power. As the phoneme model 15, the probability
Hidden Markov Model (hereinafter HM) modeled based on statistical theory
M method) is the mainstream. Details of this HMM are disclosed in, for example, Seiichi Nakagawa, "Speech Recognition by Stochastic Model", edited by the Institute of Electronics, Information and Communication Engineers.

The search processing unit 13 evaluates the plausibility of the phoneme model of the partial hypothesis, which is a phoneme sequence that can be connected by grammar, with the input speech, and extends the phoneme to the partial hypothesis one by one. Proceed while searching. Here, the partial hypothesis indicates a phoneme sequence connected according to the restriction on the order of phonemes shown in the grammar.
Extension of a phoneme to a partial hypothesis means connecting one more phoneme to the phoneme sequence of the partial hypothesis according to the grammar.

For each of the partial hypotheses, 1. phoneme sequence; 2. A score function as 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 likelihood of the partial hypothesis for the input speech, are stored. If the identification number of the partial hypothesis is i and the time is t, the score function is represented as g _i (t). In the search processing unit 13,
First, the first phoneme permitted by the grammar is extended to a partial hypothesis, and the HMM corresponding to the phoneme is compared with the analyzed feature parameter vector data time series (input speech). Score function g _i (t) at time t
Ask for. There are a trellis method and a Viterbi method as a matching method with the HMM.
Is disclosed. An evaluation value of the partial hypothesis i is obtained from the score function g _i (t) by a method described later, and a phoneme sequence, a score function g _i (t), and an evaluation value are recorded for the partial hypothesis. And every time the subsequent phoneme expansion is performed,
The search process proceeds while obtaining an evaluation value for the partial hypothesis. If two or more types of phonemes can be extended due to grammatical restrictions with respect to the phoneme sequence of the partial hypothesis, the original partial hypotheses are duplicated by the number of types of phonemes that can be extended, and the partial hypotheses obtained by expanding each phoneme And calculate an evaluation value for them. The partial hypothesis for which it is no longer possible to extend the phoneme due to the grammar ends the phoneme extension as a hypothesis that the phoneme sequence has been accepted as a grammar. When phoneme expansion cannot be performed for all partial hypotheses, all phoneme strings allowed as grammars have been matched with the input speech,
The search processing 13 ends. The phoneme string of the hypothesis with the highest evaluation value at that time or the word or sentence corresponding thereto is recognized as the recognition result 1
Output as 4.

[0006] As described above, a search method in which the phonemes of partial hypotheses are extended so that the number of phonemes of all partial hypotheses (phoneme strings) becomes equal in the search processing is called a horizontal search method. When the horizontal search method is actually performed, calculations are performed for partial hypotheses corresponding to all phoneme sequences permitted by the grammar, and a large number of partial hypotheses must be calculated, requiring a lot of processing time. And For this reason, in a process of expanding a phoneme to a partial hypothesis, a method is often adopted in which only a partial hypothesis that is likely to be a final recognition result is left and other partial hypotheses are discarded. Specifically, it is determined whether or not the partial hypothesis is to be left based on the evaluation value of the partial hypothesis. As a determination method, a method of leaving a fixed number of partial hypotheses in order from 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 the two methods is used. In such a horizontal search method, a method of performing a search by leaving only promising partial hypotheses and discarding other partial hypotheses under certain conditions is called a beam search method.

More specifically, the case where a search process using an HMM is performed on a grammar expressed by a tree structure as shown in FIG. Assuming that the processing up to the fourth phoneme has been completed and the fifth phoneme is extended, in FIG. 3B, the partial hypothesis extended from the first phoneme # to the fourth phoneme is “# i k a ”,“ # i k i ","#
i m i ". here," "Is a symbol indicating a phoneme delimiter, and a phoneme # indicates silence.

The first phoneme starts from # and extends to the fourth phoneme.
One partial hypothesis, "# i k i ”, FIG.
As can be seen from B, as the fifth phoneme, three types of phonemes
k, o, and m are extensible. Also, the first phoneme starts with #
Another partial hypothesis that began and extended to the fourth phoneme,
"# i k "a" is the second phoneme as the fifth phoneme
m and n are extensible. Also, the partial hypothesis "# i m
"i" is completed with the fourth phoneme, and the phoneme expansion is performed
Not.

For each phoneme depth in tree structure grammar
In the beam search that discards the partial hypothesis,
For partial hypotheses with prime numbers, the evaluation values of these partial hypotheses are
Then, only a partial hypothesis with a good evaluation value is left under certain conditions.
Here, as a certain condition, the two highest ranking evaluation values
Only the partial hypothesis is left. As mentioned above,
The partial hypothesis extended to five phonemes is "# i k i
o "," # i k i k ”,“ # i k i
m ”,“ # i k a m ”,“ # i k a n "
And the evaluation value of each partial hypothesis is in this order
If it is high, the top two partial hypotheses "# i k
i o "and" # i k i k ”expands the next phoneme
Remain as partial hypotheses that can be extended, and eliminate other partial hypotheses.
Abandon

Thus, the phoneme is extended to the partial hypothesis,
The remaining partial hypotheses are limited according to certain conditions, phonemes are further extended to the remaining partial hypotheses, and the same processing is continued until the phonemes cannot be extended in all the partial hypotheses. Then, all the partial hypotheses for which the phoneme 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 the partial hypotheses estimated forward from the beginning of the voice is used. In this method, a difference between this and the score function g _i (t) of the partial hypothesis i is determined, and the value corresponding to the maximum value of the difference at time t is used as the evaluation value S _i of the partial hypothesis i. (Details of this method are described in, 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. 6-133939, Sound Recognition Method ").

A specific example of a method of obtaining the evaluation value of the partial hypothesis and
Hypothesis "#" extended to the fourth phoneme i k
Diagram showing how to calculate the evaluation value when phoneme o is extended to "i"
4 will be described. Figure 4 shows collation of phoneme sequence and input speech
The score function obtained by performing the trellis calculation
In a three-dimensional diagram with three axes of sequence, input voice, and score
Therefore, the curve 31 shows the partial hypothesis “# i k
i ”score function, g_i4(T) and time t₁In
Score value of 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 indicating the plausibility at the time of the assumption, and the time
t_TwoScore in_i4(T_Two) Indicates that the input voice is at time t_TwoMa
Assuming that this partial hypothesis was uttered the longest time
The score indicating the plausibility when the
t₁, T_TwoAnd the duration of phoneme o, the time t_ThreeDecide
In that section, the input voice
The column "# i k i o "is assumed to have been uttered
Curve 3 shows the plausibility (score) of the time
2, that is, the curve 32 is a partial hypothesis "# i
k i o ”score function g_i5(T).
That is, the partial hypothesis "# i k The score function 31 of “i”
Is calculated using the likelihood at each time as the initial value.
And accumulate the score at each time of phoneme o by trellis calculation
Then, "# i k i o ”score function 32
Confuse.

The trellis calculation is performed by an HMM indicating an acoustic model.
And vector time-series data of feature parameters obtained by analyzing the input voice. The probability of vector time-series data is calculated 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 a score (likelihood).

Next, in order to obtain the evaluation value of the partial hypothesis, the forward hypothesis obtained by the nongrammar common to each partial hypothesis estimated from the beginning of the speech (there is no restriction on the grammar and extension to any phoneme is allowed) is obtained. The heuristic function g ＾ (t) is obtained, and this is subtracted from the score function g _i (t) of the partial hypothesis as in the following equation (1) to obtain the maximum value S _i .
S _i indicates the plausibility of the partial hypothesis i,
By using this as the evaluation value of the partial hypothesis i, the evaluation value of the partial hypothesis normalized with respect to time can be obtained.

S _i = max ｛g _i (t) −g ＾ (t)｝ (1) max is the maximum value in {｝ for each t. When a search is performed by a no-grammar, an evaluation value close to the correct answer is obtained. Is obtained, but the number of partial hypotheses is remarkably large, and those having substantially the same evaluation value increase, making selection difficult. Therefore, as described above, the search is performed under the constraints of the grammar.

[0016]

In speech recognition, the amount of search processing is reduced, thereby shortening the recognition processing time and improving the ease of speech recognition in actual use. In addition, the reduction in the amount of search processing makes it possible to operate speech recognition practically on a computer having a low processing capability. In order to reduce the amount of search processing, it is necessary to discard unexpected partial hypotheses in the search process and reduce the number of partial hypotheses to be expanded. However, in the conventional beam search that holds a fixed number of partial hypotheses with high evaluation values, there is a partial hypothesis with a small evaluation value among partial hypotheses that hold a fixed number, that is, a partial hypothesis that can not be a plausible recognition result However, the partial hypothesis is not discarded, and wasteful processing is performed. In a beam search in which a threshold value is set and a partial hypothesis having an evaluation value higher than the threshold value is retained, a partial hypothesis with a small evaluation value is discarded, but the evaluation value is generally determined by the number of recognized words, the speaker, Since it is greatly affected by the input speech length, it is difficult to set a threshold value at which the partial hypothesis can be effectively discarded without dropping the correct partial hypothesis.

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

[0018]

According to the present invention, the utterance content of an input speech is determined in the search process, that is, for each depth of speech units (phonemes, syllables, semi-syllables, words, etc.) in a tree-structured grammar. Estimate the evaluation value when it is assumed that the answer is correct, and use it as the reference evaluation value, and evaluate the evaluation value obtained by connecting the speech units under the constraints of the conventional grammar and collating with the acoustic model. Normalization is performed using the evaluation value, and partial hypotheses whose normalized evaluation value is equal to or smaller than the threshold value are discarded.

By the above-described normalization, the influence of the number of recognized words, the speaker, the input voice length, and the like are removed from the evaluation value of the partial hypothesis.
During the search process, we can reliably discard unsuccessful partial hypotheses,
The search efficiency can be increased, and the amount of search processing can be reduced by using the normalized evaluation value.

[0020]

Embodiments of the present invention will be described below. The input voice is analyzed and processed in the same manner as in the related art to obtain time-series feature parameter vector data. In the search processing, a phoneme is used as a unit of speech extended to a partial hypothesis, and a phoneme-synchronized beam search in which the number of phonemes is constant in each partial hypothesis.
In this case, an embodiment to which the present invention is applied will be described with reference to FIG. Phoneme extension processing unit 42 using constraint conditions of grammar 41
To extend the phoneme to the partial hypothesis i, and the trellis calculation processing unit 43
Performs matching between the HMM corresponding to the phoneme sequence and the input speech. From the obtained score function g _i (t) of the partial hypothesis i, the evaluation value calculation processing unit 47 obtains the evaluation value S _i of the partial hypothesis i. In the conventional method, a beam search is performed in which a fixed number of partial hypotheses with a high evaluation value S _i of the partial hypothesis _i are held and the rest is discarded.
At 5, the score function g ₀ (t) for the reference evaluation value is obtained by a method described later, and the evaluation value calculation processing unit 48 obtains the reference evaluation value S _O 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.

More specifically, referring to the example of FIG.
A partial hypothesis that extends a phoneme from a prime partial hypothesis is "# i
k i o "," # i k i k ”,“ # i
k i m ”,“ # i k a m ”,“ # i k
a n ", and each partial hypothesis is a partial hypothesis
i, and the evaluation value of the partial hypothesis i is S_iAnd the reference evaluation value S
₀Then, the normalized evaluation of the partial hypothesis i is given 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 o "partial hypothesis is the most correct answer
The evaluation value increases soon. Also,"# i k a m ”
The partial hypothesis far from the correct answer
Is getting smaller. The reference evaluation value is
Estimated evaluation value assuming correct answer.
Grammar-free grammar,
Allowed to accept all acoustic model combinations without restrictions
Phoneme sequence that is the same as the content of the input speech
Or a match with something close to this, and the phoneme sequence is
It should be the combination with the highest evaluation value
The standard evaluation value is "# i k i o "
It becomes a value close to the evaluation value of the theory. Therefore, the normalized evaluation value
S_i'Is 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_i', This tendency is₀And S _iAre the same
S because it is made from one input voice₀And S_iSpeakers included
Are removed by subtraction in the normalized evaluation value, and the speaker
Hard to depend on. S for similar reasons_iSaid tendency of input sound
It does not depend on voice length. Also, keep the number of partial hypotheses constant.
In the beam search, the partial temporary
It is necessary to change the number of theories, but the evaluation value itself is recognized
Since the vocabulary number does not change, the normalized evaluation value
S_i'Has little effect on the number of recognized words.

When discarding a partial hypothesis having a low normalized evaluation value S _i ′ in beam search, a threshold value L is determined and S _i ′ <L
The partial hypothesis is discarded, but L is set to a fixed 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, the value is set to a large negative value in the above example in accordance with this. Is also good. Equation (1) is used as a 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 equation (3), the amount of calculation processing for the normalized evaluation value S _i ′ can be significantly reduced.

S _i ′ = max {g _i (t) −g ₀ (t)} (3) max is the maximum value in {} for each t. The score function g ₀ (for the reference evaluation value S _O The method for obtaining t) will be described below. <Method 1 for calculating score function for reference evaluation value> Each phoneme HMM usually has about three states, and each state has an output probability density distribution of a weighted sum of a plurality of probability density functions. Here, the characteristic parameters of the input speech at each time are given to all output probability density distributions, the highest output probability density value is selected, and the log-like maximum likelihood at each time is obtained. The cumulative value of the maximum likelihood in the time progression is obtained, and this is used as a score function for the reference evaluation value. Oτ
Is the feature parameter at time τ, and p _j (Oτ) is the output probability density value obtained by giving the feature parameter to the output probability density distribution j, and g ₀ (t) is given by equation (4).

G ₀ (t) = {maxp _j (Oτ) (4)} is from τ = 0 to t, and max is the maximum value for all _j in p _j (Oτ), usually from one HMM to another HMM
Transition from the end state of one HMM to the other HM
Although the transition is made under the condition that the transition to the first state of M is made, this score function eliminates the transition condition and grammatical restrictions, and from any state of all HMMs to any state of any HMM. A score function when Viterbi calculation is performed with a transition to a state permitted and the transition probability set to 1 is shown. As the search progresses, most of p _j (Oτ)
Since the trellis calculation is performed in the search process, the result can be used and the calculation amount can be reduced.

<Method 2 of calculating score function for reference evaluation value> In the above-described calculation method 1, the score function was obtained from the maximum value of the output probability density values obtained from the output probability density distributions of all states of all HMMs. In this calculation method 2, the output probability density value is obtained from the maximum value of the output probability density values for all the output probability density distributions that have been calculated by the trellis calculation up to the present time in the search process. For example, as shown in FIG. 2, when the output density distributions p ₁ , p ₂ , p ₃ ... In each state of each HMM are plotted on the vertical axis and the time t is plotted on the horizontal axis, the silence # in the example of FIG. In this example, the silence length for which the output probability density value is predicted with respect to the output density distribution of each state of the HMM is calculated from time 0 to 3 (the region where the calculated value is embedded is indicated by 51), and the shortest silence length is calculated. The output probability density values for the output density distributions of the respective states of the HMM of the next phoneme i are calculated from the time 2 following the end time 1 after the end time 1 to the time 4 following the end time 3 of the longest silence. The area in which the calculated value is embedded is indicated by 52. Similarly, the output probability density value of each state of the HMM of the phoneme k is calculated as an area 53 in FIG. Although such calculation is advanced by the search, each time 0, 1, in FIG.
The maximum value of each calculated output probability density value in 2,. The maximum values are 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 restricted by the grammar in the search process, a score function closer to the actual grammar is 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 required. Even in such a calculation method, the output probability density value of a portion not restricted by the grammar is considered to be mostly smaller than that obtained by trellis calculation, and the number g ₀ (t) is correctly estimated.

<Method 3 of calculating score function for reference evaluation value> As described in the description of the horizontal search method, a phoneme is extended to a partial hypothesis, and a score function is obtained by performing collation such as trellis calculation. . In this case, the phoneme is extended by a grammar that can extend an arbitrary phoneme to each partial hypothesis, that is, a non-grammar, and at each time of the score function obtained by comparing the corresponding acoustic model with the input speech. 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 method, a highly accurate normalized evaluation value S _i ′ can be obtained. The amount of calculation also increases.

<Method of calculating score function for reference evaluation value
Method 4> Calculation method 3 of score function for reference evaluation value
Is not a grammar that can extend any phoneme,
By using a grammar that allows only Japanese phoneme sequence structures, the likelihood
Calculate, and use the resulting score function
Tick function. Allow Japanese phoneme array structure
The phoneme sequence is, for example, “o m o sh i r o
i "or" s u t o r a i k u "
As a general rule, constrained consonants do not follow consonants.
are doing. "S t r ai chain of phonemes "k"
Satisfies the phoneme sequence structure in English, but sounds in Japanese
It does not have a prime array structure.

G ₀ in calculation methods 3 and 4
The grammar that extends phonemes when calculating (t) can be said to be a grammar that includes a grammar for creating a partial hypothesis for search. <Method 5 for calculating score function for reference evaluation value> The final correct partial hypothesis often has a score function larger than other partial hypotheses. Therefore, the score functions g ₁ (t), of all partial hypotheses calculated in the search process
g ₂ (t), g ₃ (t),...
₀ (t). The expression is as follows.

G ₀ (t) = maxg _i (t) (5) max is the largest of all _i of g _i (t). This calculation method requires almost no calculation amount for g ₀ (t). do not do. <Score Function Calculation Method 6 for Reference Evaluation Value> In the calculation of the score function g ₀ (t) for obtaining the reference evaluation value S _O , there is no need to identify phonemes, and it is sufficient to obtain a score. Therefore, it is not necessary to use an HMM for each phoneme, and the acoustic model 1 for recognition is used as shown by the dotted line in FIG.
5 may be used. As the acoustic model 46, for example, even if one or several acoustic models are provided, a large number of states are provided so that an acoustic phenomenon including a recognition target is included. May be configured so that all can be expressed, and this one acoustic model is repeatedly used,
Alternatively, in the case of several acoustic models, these may be arbitrarily selected and connected, collated with the input speech, and a plausible one may be obtained to obtain g ₀ (t).

In the above description, a forward heuristic function was obtained in order to obtain an evaluation value. For example, “Large vocabulary continuous speech recognition algorithm for number guidance,” Minami et al. 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 used as a score function g _i (t).
May be added as the evaluation value S _i . Further, the present invention is applied not only to speech recognition in units of phonemes, but also to recognition in units of syllables, semi-syllables, words, and the like.

An experimental example will be described below. Phoneme balance 216
In the word recognition for the 108 odd-numbered words, the speech data of 108 odd-numbered words as words in the vocabulary and 108 words of even-numbered words as non-vocabulary words were given and evaluated. During the search, the recognition rate for word recognition in the vocabulary is the overall recognition rate, and the percentage of words recognized as "no recognition result" in the vocabulary is the false rejection rate, In the word recognition, the ratio of “the recognition result is not rejected” 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 determination rate. In other words, it is considered that discarding performance is good when the erroneous determination rate can be kept low while the recognition rate is maintained.

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

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

[0035]

According to the present invention, while the absolute value of the evaluation value of the conventional partial hypothesis depends on the speaker, the number of recognized vocabularies, and the input speech length, in the present invention, the evaluation value of the partial hypothesis is obtained from the same input speech. Because the evaluation value is normalized, a normalized evaluation value that does not depend on the speaker, the number of recognized vocabulary words, and the input speech length can be obtained, and it is possible to effectively discard an unlikely partial hypothesis in the search process. . As a result, the threshold value for the normalized evaluation value is the same, and speech recognition can be used for various purposes, and the burden on the user for setting can be reduced.

If the input speech does not allow the grammar, the conventional search will output an incorrect result that is the closest candidate in the grammar, resulting in a user utterance error and a speech recognition error. No distinction from cognition could be shown. However, in this case, according to the present invention, all partial hypotheses are discarded in the search process, no recognition result is obtained, and the user can be notified of the utterance error. It is important in practical speech recognition to find and show user utterance mistakes early.

The effects of the method of the present invention are listed below.・ Effective discard of partial hypotheses that are unlikely in the search process can be performed. The threshold value to be set does not depend on the speaker, the number of recognized vocabulary words, and the input voice length, so that the burden of setting by the user can be reduced. -If the input speech does not allow the grammar, it is possible to detect that recognition cannot be performed early in the search process, and to notify the user of a speech error.

[Brief description of the drawings]

FIG. 1 is a diagram showing an example of a technique for 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 trellis calculation for explaining a score function calculation method 2 for a reference evaluation value.

FIG. 3A is a diagram illustrating a process of a speech recognition method using phonemes as a unit of recognition, and FIG. 3B is a diagram illustrating a grammar represented 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 an experiment performed on the method of the present invention.

────────────────────────────────────────────────── ─── Continuation of the front page (56) References JP-A-8-6588 (JP, A) JP-A-2-300798 (JP, A) Proceedings of the Acoustical Society of Japan Spring Meeting 2007 , 1-Q-28, Yoshiaki Noda, “Beam search by forward heuristic function with dynamic reject function” p. 151-152 (issued on March 14, 1995) IEICE Technical Report [Voice], Vol. 94, no. 91, SP94-23, Yoshiaki Noda, "HMM-LR speech recognition by A * beam search using forward likelihood," p. 1-8 (Issued June 17, 1994) (58) Fields investigated (Int. Cl. ⁷ , DB name) G10L 15/00-17/00 JICST file (JOIS)

Claims (12)

(57) [Claims] 1. A method for generating one or more hypotheses relating to the content of an input speech by gradually adding and connecting speech units based on a tree-structured grammar composed of speech units. For each depth of the tree-structured speech unit, based on the acoustic model, the plausibility of the previous partial hypothesis in each hypothesis with respect to the input speech is evaluated to obtain a partial hypothesis evaluation value. In the speech recognition method for obtaining a recognition result from likelihood, an evaluation value when assuming that the utterance content of the input speech is correct is estimated as a reference evaluation value for each depth of the speech unit of the tree structure. A speech recognition method comprising: normalizing an evaluation value of a partial hypothesis having a depth corresponding to a reference evaluation value, and discarding a partial hypothesis whose normalized evaluation value is equal to or smaller than a threshold value. 2. Based on a grammar including the grammar, a hypothesis relating to the utterance content of the input speech is generated by adding speech units gradually and continuously, and the input speech is generated by the sound corresponding to a partial hypothesis. 2. The speech 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 comparing the input speech with a partial hypothesis of at least one reference evaluation value acoustic model representing an acoustic phenomenon including a recognition target, and obtaining a score function. The speech recognition method according to claim 1, wherein an evaluation value is obtained. 4. The partial hypothesis evaluation value according to claim 2, wherein the input speech is collated with an acoustic model corresponding to the partial hypothesis, a score function is obtained, and the partial hypothesis evaluation value is obtained. Voice recognition method. 5. The speech recognition method according to claim 4, wherein said acoustic model is a hidden Markov model. 6. A score function for obtaining a maximum value of all output probability density values of the Hidden Markov Model at each time and accumulating the maximum values to calculate the reference evaluation value. The speech recognition method according to claim 5, wherein 7. A score function for selecting a maximum value among output probability values of hidden Markov calculated to obtain the partial evaluation value at each time, and accumulating the maximum value to obtain the reference evaluation value. The speech recognition method according to claim 5, wherein is calculated. 8. The speech recognition method according to claim 2, wherein the grammar including the grammar allows any combination of acoustic models corresponding to speech units. 9. The speech recognition method according to claim 8, wherein a constraint of a phoneme array structure unique to Japanese is used for the combination of the acoustic models corresponding to the speech units. 10. The input speech is collated with the acoustic model corresponding to a partial hypothesis, a score function is obtained to obtain the partial hypothesis evaluation value, and the reference evaluation value is set to a maximum value of the score function at each time. 2. The speech recognition method according to claim 1, wherein the speech recognition method obtains the following. 11. An evaluation value of the partial hypothesis is obtained by finding a forward heuristic function common to all partial hypotheses, taking the difference between the score function of each partial hypothesis and the forward heuristic function, and determining the maximum value of the difference. 11. The speech recognition method according to claim 4, wherein the value is obtained as a value to be obtained. 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 heuristic function.