JP2003005787A - Speech recognition device and speech recognition program - Google Patents

Abstract

(57) [Summary] In a large vocabulary speech recognition system, a speech recognition device that realizes real-time recognition processing within a frame length by reducing the processing amount of an output probability calculation unit is provided. SOLUTION: A relationship between a phoneme of a current frame and a phoneme considered to be generated in the next frame is stored in a continuous phoneme information storage unit 14.
Is stored in The phoneme selection unit 15 first selects a phoneme with a high output probability in the current frame, and then searches the phoneme that is considered to occur in the next frame of the selected phoneme by using the continuous phoneme information in the continuous phoneme information storage unit 14. At this time, it is extracted from the phoneme information storage unit 11 and stored in the selected phoneme information storage unit 16. The output probability calculation in the next frame is performed only on the phonemes stored in the selected phoneme information storage unit 16, so that the processing amount in the output probability calculation unit 10 is reduced.

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to a Hidden Makcov Model (hereinafter referred to as HMM).
The present invention relates to a voice recognition device using.

[0002]

2. Description of the Related Art In recent years, devices with a voice recognition function that can be operated by voice have become widespread in car navigation systems, portable terminals and the like.

As the voice recognition technology, at present, the voice recognition technology using HMM is predominant. The method using the HMM is to prepare a plurality of model state sequences representing a voice signal and determine from which model the input signal was generated. The model here includes a method for each word or each phoneme.

Specifically, when transitioning from one state to another, which state has the highest probability of transition next is determined, and each time the state transitions (including self transition). By accumulating the probabilities, the method with the highest cumulative probability finally becomes the recognition result.

FIG. 17 shows a conventional voice recognition device using an HMM. The conventional voice recognition device has an output probability calculation unit 2 for calculating which model the input voice is likely to be.
0, model that is the minimum recognition unit in advance (phoneme information)
Is stored in the phoneme information storage unit 21, a matching unit 22 that calculates the cumulative likelihood of each word, and a word dictionary 23 that holds a phoneme sequence of recognized words.

A series of operations of the conventional speech recognition apparatus will be described. First, the speech feature parameter for each frame length is input to the output probability calculation unit 20. Next, the output probability calculation unit 20 calculates the output probability of which phoneme the input voice is likely to have, based on the phoneme information in the phoneme information storage unit 21. The output probability calculation unit 20 calculates the total number of registered phonemes (the total number of distributions = the number of HMMs × HM).
The number of states of M × the number of mixtures) is repeated. The output probability calculated by the output probability calculation unit 20 is input to the matching unit 22. Next, the matching unit 22 accumulates the output probabilities for each word based on the word dictionary 23. Then, the characteristic parameter of the next frame is input to the output probability calculation unit 20, and the processing of the output probability calculation unit 20 and the processing of the matching unit 22 are executed until the voice input is completed,
Determine the final recognition result.

[0007]

When the voice recognition technology using the HMM is realized by a microcomputer having an upper limit of processing performance, in order to realize the recognition processing within the frame length in real time, the processing is It is essential to speed up. In the case of small vocabulary speech recognition, the total calculation amount of the output probability calculation unit,
The processing amount of the matching unit is almost the same. On the other hand, in the case of large vocabulary speech recognition, since the number of similar words increases, it is necessary to further subdivide and distinguish each phoneme in order to improve the recognition performance.

For example, in the case of small vocabulary voice recognition, "A"
In the case of large vocabulary speech recognition, the speech represented by the phoneme "a" is represented by "a-i" (i.e., the phoneme of the next frame is i), "au" (next). The phoneme of the frame is u (a), "a-sil" (the phoneme of the next frame is sil (silent)), and so on.
The same voice "a" is also classified into various "a" according to the conditions. An example in which a recognition word is described with 35 phonemes for small vocabulary recognition is shown in Table 1, and an example in which a recognition word is described with 288 phonemes for large vocabulary recognition is shown in Table 2.

[0009]

[Table 1]

[0010]

[Table 2]

As the number of states and the number of mixtures, values such as three states and four mixtures are generally used. Therefore, in the case of 35 phonemes, the total number of distributions is 35 phonemes x 3 states x 4
Mixing = 420 distribution, whereas for 288 phonemes,
288 phonemes × 3 states × 4 mixture = 3456 distribution, and in the case of large vocabulary speech recognition, the processing amount in the output probability calculation unit 20 increases significantly. As a result, the recognition processing within the frame length cannot be realized in real time.

In view of the above problems, the present invention can reduce the processing amount of the output probability calculation unit in the large vocabulary speech recognition system, and can provide the real-time execution processing within the frame length. The purpose is to provide a device.

[0013]

In order to solve the above-mentioned problems, a speech recognition apparatus of the present invention is a speech recognition apparatus using a hidden Markov model, wherein characteristic parameters of speech for each frame length are set for all phonemes. The output probability of each phoneme in the frame is calculated from the phoneme information storage unit held as a hidden Markov model, the hidden Markov model held in the phoneme information storage unit, and the characteristic parameter of the input speech for each frame length. Output probability calculation means, continuous phoneme information storage means for holding the relationship between the phoneme of the current frame and the phoneme considered to occur in the next frame, selected phoneme information storage means, and the output probability calculated by the output probability calculation means. And the phoneme information regarding the phoneme that is considered to occur in the next frame, based on the phoneme information stored in the phoneme information storage means A phoneme selection unit that is selected from the phoneme information storage unit and stored in the selected phoneme information storage unit, a word dictionary storage unit that holds a phoneme sequence of each recognized word, and an output probability calculated by the output probability calculation unit. And matching means for accumulating each word based on the word dictionary.

In the above configuration, the phoneme selection means is considered to occur in the next frame using the output probability calculated by the output probability calculation means and the continuous phoneme information stored in the continuous phoneme information storage means. Phonemes are extracted from the phoneme information storage means and stored in the selected phoneme information storage means. In the next frame, the output probability calculation means calculates the output probability not only for all phonemes but only for the phonemes selected and stored in the selected phoneme information storage means. As a result, the processing amount of the output probability calculation means can be reduced, and it is possible to provide a voice recognition device that realizes real-time execution processing within the frame length.

Further, in the speech recognition apparatus of the present invention, among the continuous phoneme information of the continuous phoneme information storage means, the relationship of phonemes that are considered not to occur continuously is determined by the phoneme of the word dictionary storage means. It is preferable to include continuous phoneme information optimizing means for deleting in advance based on the sequence.

According to the above configuration, the phoneme information stored in the continuous phoneme information storage means is not continuously generated based on the phoneme series of the recognized words stored in the word dictionary storage means. The phoneme relationship considered to be deleted is deleted, and continuous phoneme information is optimized. As a result, the output probability calculation is not performed for the phoneme that does not occur in the next frame, so that the processing amount of the output probability calculation means can be further reduced.

Further, in the voice recognition device of the present invention,
When the phoneme selection unit refers to the cumulative probability of each word obtained by the matching unit in addition to the output probability calculated by the output probability calculation unit and the continuous phoneme information, and the output probability is used as a reference It is preferable to select a phoneme that is not selected as a target phoneme, when it is considered that the phoneme of the current frame is likely to occur when the cumulative probability is used as a reference.

According to the above construction, even if the output probability in the current frame is not selected so small, if the cumulative probability from the recognition start frame to the current frame is high and there is a high possibility that it is a phoneme of the current frame, that phoneme is selected. select.
As a result, it is possible to prevent a reduction in recognition rate due to a phoneme selection error and reduce the processing amount of the output probability calculation means.

Further, the speech recognition apparatus of the present invention further comprises similar phoneme storage means for storing similar phonemes, wherein the output probability calculation means inputs the previous similar phoneme and selects the phoneme selected by the phoneme selection means. If the first phoneme belonging to the lower rank with respect to the sufficiency of the selection criterion is similar to the second phoneme belonging to the higher rank, the second phoneme is not calculated for the output probability. It is preferable to use the output probability calculated for the phoneme of.

According to the above configuration, when the phoneme belonging to the lower rank of the phonemes selected by the phoneme selecting means is similar to the phoneme belonging to the higher rank, the output probability calculating means. In, the output probability calculation is not performed for the phonemes belonging to the lower order,
The output probabilities calculated for the phonemes belonging to the higher order are substituted. As a result, it becomes possible to further reduce the amount of processing in the output probability calculation means.

In the speech recognition apparatus, it is preferable that the hidden Markov models stored in the phoneme information storage unit share common phoneme states. By sharing the states of similar phonemes, the number of distributions of the entire phonemes can be reduced, and the amount of processing in the output probability calculation means can be further reduced.

[0022]

BEST MODE FOR CARRYING OUT THE INVENTION A voice recognition device according to each embodiment of the present invention will be described below with reference to the drawings.

(First Embodiment) FIG. 1 is a block diagram showing the arrangement of a voice recognition apparatus according to the first embodiment of the present invention. The speech recognition apparatus of FIG. 1 has an output probability calculation unit 1
0, continuous phoneme information storage unit 14, phoneme selection unit 15, phoneme information storage unit 11, selected phoneme information storage unit 16, word dictionary 1
3 and a matching unit 12.

The output probability calculation unit 10 receives the feature parameters of the voice for each frame length as input, and calculates the output probability based on the phoneme information stored in the selected phoneme information storage unit. The continuous phoneme information storage unit 14 holds the relationship between the phoneme at time t and the phoneme that is considered to occur at time t + 1. The phoneme selection unit 15 selects a phoneme generated at time t + 1. The phoneme information storage unit 11 stores, for all phonemes, the characteristic parameters of the voice for each frame as an HMM. The selected phoneme information storage unit 16 extracts and stores only the phoneme information regarding the phoneme selected by the phoneme selection unit 15. The word dictionary 13 has registered phoneme sequences of words to be recognized in advance. The matching unit 12 accumulates output probabilities for each word using the output probabilities calculated by the output probability calculation unit 10 and the word dictionary 13.

Next, the operation of the voice recognition device according to the above configuration will be described.

Firstly, the feature parameter y _t of the speech at time t, and inputs the output probability calculation unit 10.

Next, the output probability is calculated only for the phoneme H _p (p is a phoneme number) stored in the selected phoneme information storage unit 16. The selected phoneme information storage unit 16 holds the characteristic parameter of the voice for each predetermined time length (hereinafter, frame length) as an HMM. The HMM held here is phoneme information λ _p (μ, σ) such as the average value μ _p and the variance σ _p of the feature parameter for each phoneme. As an example, Table 3 shows μ _p and σ _p of the phoneme sil (silent).

[0028]

[Table 3]

The dimension in Table 3 is the characteristic parameter y _t.
Is a parameter used to represent a vector. Note that this value is a value corresponding to the number of states 1 and the number of mixtures 1, and in addition, values corresponding to the numbers of states 2 to 3 and the numbers of mixtures 2 to 4 actually exist. In addition, since it is a value calculated by our own calculation method, it is not a commonly used value.

The total number of phonemes is N, and the number of phonemes stored in the selected phoneme information storage unit 16 is n (0 <n ≦
N). The initial value of the selected phoneme information storage unit 16 is
It is sil (silence).

The output probability calculation unit 10 and the feature parameter y _t of the input speech at time t, and inputs the phoneme information lambda _p stored in the selected phoneme information storing section 16 (mu, sigma), which input speech is Output probability b whether it is likely a phoneme
Calculate _i ^p (y _t ). Here, i represents the number of states. The output probability calculation is performed only on the phonemes stored in the selected phoneme information storage unit 16. Selected phoneme information storage unit 1
Assuming that the number of phonemes stored in 6 is n and the number of states and the number of mixtures are 3 and 4 respectively, the total distribution number is n × 3.
State × 4 mixture = 12n, and output probability calculation is performed 12n times.

Next, a series of processing flow of the phoneme selection unit 15 will be described.

The phoneme selection unit 15 first receives the output probability b _i ^p (y _t ) at time t calculated by the output probability calculation unit 10. Next, the phoneme selection unit 15 uses the continuous phoneme information storage unit 1.
Based on the continuous phoneme information stored in 4, the time t + 1
Select a phoneme that is thought to occur in.

The continuous phoneme information storage unit 14 holds the relationship between the phoneme at time t and the phoneme that is considered to occur at time t + 1. For example, it is assumed that a (a) at which i (a) comes at time t + 1 is described as “a-i”. Then, if the phoneme at time t is "a-i", the phoneme considered to occur at time t + 1 is "i
-○ "(where ○ is an arbitrary phoneme). FIG. 4 shows a phoneme worthy of “i− ◯” as a phoneme that may occur at the time when “ai” occurs. For example, the phoneme generated at time t is “a
-I ", the phonemes considered to occur at time t + 1 are" i-a "and" i- ", as shown in FIG.
There are four possible phonemes: "b", "id", and "ii". Such a relationship between the phoneme at time t and the phoneme that is considered to occur at time t + 1 is described for all phonemes and stored in the continuous phoneme information storage unit 14.

It is considered that the phoneme selection unit 15 will occur at time t + 1 based on the output probability b _i ^p (y _t ) calculated by the output probability calculation unit 10 and the continuous phoneme information in the continuous phoneme information storage unit 14. Only the phoneme information regarding the selected phoneme is selected from the phoneme information storage unit 11 and stored in the selected phoneme information storage unit 16. The phoneme information storage unit 11 stores phoneme information λ for each frame length for all N registered phonemes.
_p (μ, σ) is retained. The selection method from the phoneme information storage unit 11 has the following two patterns.

(Pattern 1) FIG. 2 is a flowchart of the processing of the phoneme selection unit 15 according to pattern 1.

The phoneme selection unit 15 inputs the output probability calculation result b _i ^p (y _t ) (step S1), and then b _i ^p
(Y _t ) is sorted (step S2). Then, the top m phonemes having a large value are selected from the sorting result (step S3). Then, for the m phonemes, the phonemes considered to occur at time t + 1 are selected using the continuous phoneme information in the continuous phoneme information storage unit 14 (step S4). Then, the phoneme information regarding the selected phoneme is extracted from the phoneme information storage unit 11 (step S
5) and stores it in the selected phoneme information storage unit 16 (step S
6).

(Pattern 2) FIG. 3 is a flowchart of the processing of the phoneme selection unit 15 according to pattern 2.

The phoneme selection unit 15 first inputs the output probability calculation result b _i ^p (y _t ) (step S11). Next, a threshold th that is set in advance for use in phoneme selection is referred to, and a phoneme for which b _i ^p (y _t ) is larger than th is selected (step S12). For the selected phoneme,
Time t is calculated using the continuous phoneme information in the continuous phoneme information storage unit 14.
A phoneme considered to occur at +1 is selected (step S13). Then, the phoneme information regarding the selected phoneme is extracted from the phoneme information storage unit 11 (step S14) and stored in the selected phoneme information storage unit 16 (step S15).

The phoneme selection unit 15 performs steps S12 to S1.
The process of 5 is repeatedly executed until the process is completed for all phonemes.

In step S5 of pattern 1 or step S14 of pattern 2 described above, for example, when the output probability of the phoneme "a-i" is high at time t, what is considered as the phoneme at time t + 1 is shown in FIG. As described above with reference to "i-a", "i-b", "i-".
There are four types, "d" and "ii". The phoneme selection unit 15
The four types of phoneme information λ _p (μ, σ) are extracted from the phoneme information storage unit and stored in the selected phoneme information storage unit 16.

As described above, after the output probability calculation unit 10 has performed the output probability calculation, the matching unit 12 refers to the word dictionary 13 and sets the output probability calculated by the output probability calculation unit 10 to the recognition word. It accumulates every time. The word dictionary 1
3 holds the phoneme sequence of each recognized word as shown in Table 2 shown previously.

After accumulating the output probabilities for each word in the matching unit 12, the characteristic parameter y _{t + 1} at time t + 1 is input to the output probability calculation unit 10, and the output probability b _i ^{p at} time t + 1.
Calculate (y _{t +1} ). When calculating the output probability, only the phonemes held in the selected phoneme information storage unit 16 are calculated.

In this way, the calculation of the output probability and the cumulative probability is repeated for the voices of all the frames, that is, the voices until time T (0≤t <T). Finally, the word with the highest cumulative probability is used as the recognition result.

Next, how much the calculation amount of the output probability calculation unit 10 is reduced in the present embodiment will be described by showing a concrete example.

First, at time t, it is assumed that the phoneme stored in the selected phoneme information storage unit 16 is only sil (silent). Therefore, the output probability calculator 10 calculates once. The phoneme selection unit 15 selects a phoneme with a high output probability, but in this case, only the output probability of sil is input, so only sil is selected. Next, the phoneme selection unit 15 uses the continuous phoneme information in the continuous phoneme information storage unit 14 to select a phoneme that is considered to occur at time t + 1. FIG. 5 shows continuous phoneme information regarding sil. As shown in FIG. 5, the phoneme selection unit 15 selects 23 phonemes that are considered to occur at time t + 1, phoneme information regarding these phonemes is extracted from the phoneme information storage unit 11, and the selected phoneme information storage unit 16 is selected. Stored in.

At the next time t + 1, the selected phoneme information storage unit 1
Since 6 phonemes have 23 phonemes stored therein, the number of calculations performed by the output probability calculation unit 10 is 23. After calculating the output probability, it is assumed that pattern 2 above is used as the phoneme selection method, and si
It is assumed that the output probabilities of lk and sil-To are larger than the threshold th. Next, the phonemes considered to occur at the next time of sil-k and sil-To are selected from the continuous phoneme information storage unit 14.
FIG. 6 shows continuous phoneme information regarding sil-k, and FIG. 7 shows continuous phoneme information regarding sil-To. As shown in FIG. 6, 14 phonemes that are considered to occur at the next time of sil-k and 1 phoneme that is considered to occur at the next time of sil-To, as shown in FIG. The phoneme information of 15 phonemes is extracted from the phoneme information storage unit 11 and stored in the selected phoneme information storage unit 16.

At the next time t + 2, the selected phoneme information storage unit 1
Since 15 phonemes are stored in 6, the number of calculations in the output probability calculation unit is 15.

A comparison of the number of times of the output probability calculation from time t to t + 2 between the conventional method and the method according to the present embodiment (indicated as “proposed method” in Table 4) is as follows.
It is Table 4.

[0050]

[Table 4]

As described above, it can be seen that the processing amount of the output probability calculation unit 10 is significantly reduced in this embodiment as compared with the conventional method.

Among the conventional examples, at least two types of phoneme information (HMM: simple HMM and detailed HMM) for the same phoneme.
There is a method of reducing the processing amount of the output probability calculation unit by providing MM). In this method, one simple HMM is used to calculate the output probabilities of all phonemes, and then the other detailed HMM is used to calculate the output probabilities for the higher ones of the output probabilities. However, in this method, since two kinds of HMMs are held, the storage memory amount is increased and the output probability calculation is performed twice by using both the simple and detailed HMMs.
It is thought that processing overhead will also occur.

The phoneme information stored in the phoneme information storage unit 11 may be shared if different phonemes have similar states. By sharing the states, the number of distributions of the entire phonemes can be reduced, and the processing amount in the output probability calculation unit 11 can be reduced.

As described above, according to the first embodiment, a phoneme having a high output probability is first selected at time t, and the phoneme generated at time t + 1 is selected from the selected phonemes based on continuous phoneme information. Then, select the phoneme
At time t + 1, the output probability calculation is performed only on the selected phoneme, so that the processing amount of the output probability calculation unit can be reduced. As a result, the processing rate of the output probability calculation unit is reduced, which enables faster recognition processing, and real-time recognition processing within the frame length when speech recognition is performed with limited processing performance such as a microcomputer. It becomes possible to do.

(Second Embodiment) FIG. 8 shows the configuration of a speech recognition apparatus according to the second embodiment of the present invention.

As shown in FIG. 8, the second embodiment of the present invention has a configuration in which a continuous phoneme information optimizing unit 47 is provided between the word dictionary and the continuous phoneme information storage unit of the first embodiment. is there. Output probability calculation unit 40, selected phoneme information storage unit 46, phoneme information storage unit 41, phoneme selection unit 45, matching unit 42
Have the same operation as the components having the same names in the first embodiment, and detailed description thereof will be omitted.

As shown in the flowchart of FIG. 9, the continuous phoneme information optimizing unit 47 selects the phoneme sequence of each recognized word described in the word dictionary 43 from the phoneme information contained in the continuous phoneme information storage unit 44. Performs a process of deleting the relationship (phoneme connection) between the unused phoneme at time t and the phoneme at time t + 1.

In the original continuous phoneme information storage section 44,
The relationship between the phoneme at time t and all the phonemes that may occur at time t + 1 is stored. However, it cannot be considered that all phoneme connections are generated depending on the word to be recognized. Therefore, the continuous phoneme information optimization unit 47 refers to the word dictionary 43 in which the phoneme sequence is described, and determines the phoneme connections that are considered not to occur, that is, the phoneme connections that are not included in the word dictionary 43. It is deleted from the storage unit 44 in advance. as a result,
The output probability calculator 40 uses the word dictionary 4 currently used.
Within 3, the unnecessary output probability calculation is not performed for phonemes that do not continuously occur. Thereby, the processing amount of the output probability calculation unit 40 can be further reduced.

A specific description will be given below with reference to an example.

Here, it is assumed that the word dictionary 43 has the words shown in Table 2 above. Further, it is assumed that there are 14 pieces of continuous phoneme information stored in the continuous phoneme information storage unit 44 for sil-k as shown in FIG. However, according to the word dictionary 43, the phonemes connected to sil-k are kj
And ko, only phoneme connections that do not occur are deleted from the continuous phoneme information storage unit 44 by sil-
There are only two pieces of continuous phoneme information about k shown in FIG. Therefore, as compared with the case where the output probability calculation is performed for all 14 phonemes connected to sil-k, the output probability calculation unit 40 according to the present embodiment only needs to calculate twice. It is possible to further reduce the processing amount of.

As described above, according to the second embodiment, based on the phoneme sequence of the word dictionary, the phoneme information which is considered not to be generated by the word recognition currently used is generated from the continuous phoneme information. By deleting, the number of phoneme information can be reduced, and thus the processing amount of the output probability calculation unit can be reduced.

(Third Embodiment) A third embodiment of the present invention is a phoneme selection unit in the first embodiment, which uses information of cumulative probabilities for each word in the matching unit.

FIG. 11 shows the configuration of a voice recognition device according to the third embodiment.

Output probability calculation unit 50, selected phoneme storage unit 5
6, the phoneme information storage unit 51, the continuous phoneme information storage unit 54, and the word dictionary 53 have the same operations as the components having the same names in the first embodiment, and thus detailed description thereof will be omitted.

The phoneme selection unit 55 uses the output probabilities calculated by the output probability calculation unit 50, the phoneme information stored in the continuous phoneme information storage unit 54, and the cumulative probabilities calculated by the matching unit 52. Select a phoneme that is thought to occur in the next frame.

In the first embodiment, the phoneme considered to occur at time t + 1 is selected based on only the output probability and the continuous phoneme information. According to this, the cumulative probability calculated from the recognition start time to the time t is high,
There is a possibility that some phonemes have not been selected because the output probability at time t is low. Even if the output probability at time t is low, if the cumulative probability up to time t−1 is high, there is a sufficient possibility that it will be a recognition result. Therefore, the speech recognition apparatus of the present embodiment also inputs the cumulative probability of each word to the phoneme selection unit 55 so as to select the phonemes that can be the recognition result when the cumulative probability is also taken into consideration. Based on both, select a phoneme that is likely to be the phoneme at time t,
Next, based on the continuous phoneme information from the continuous phoneme information storage unit 54, a phoneme considered to occur at time t + 1 is selected.

The processing flow of the phoneme selection unit 55 of this embodiment is shown in the flowchart of FIG.

The phoneme selection unit 55 first inputs the output probability b _i ^p (y _t ) which is the calculation result from the output probability calculation unit 50 (step S31), and further, the matching unit 52.
The cumulative probability is input from (step S32). continue,
It is determined whether the output probability b _i ^p (y _t ) is larger than the set threshold th (step S33). The determination result is YE
If S, the threshold value th for which the cumulative probability is set
It is determined whether it is larger than 2 (step S34). When the determination result of step S33 is NO, or when the determination result of step S33 is YES and the determination result of step S34 is YES, the continuous phoneme information in the continuous phoneme information storage unit 54 is used to generate at time t + 1. A possible phoneme is selected (step S35). Then, the phoneme information regarding the selected phoneme is extracted from the phoneme information storage unit 51 (step S36) and stored in the selected phoneme information storage unit 56 (step S37).

Next, the above operation will be further described using a concrete example.

Here, it is assumed that the word dictionary 53 has the words in Table 2 and the output probability calculation result at time t when the voice "Kyoto" is input is shown in Table 5 and accumulated up to time t-1. The probabilities are shown in Table 6.

[0071]

[Table 5]

[0072]

[Table 6]

Assuming that the threshold value th is 90 when selecting a phoneme at time t based on this output probability, Table 5
More phonemes oo are selected, but jo and ar are not. However, the input voice is "Kyoto (sil-k kj jo oo
ot to oT sil) ”, a recognition error occurs if jo is not selected. Therefore, the voice selection unit 55 performs phoneme selection using not only the output probability at time t but also the cumulative probability up to time t−1 calculated by the matching unit 52.
Also in this case, the threshold value th2 is set and it is determined whether or not the cumulative probability of each recognized word is larger than the threshold value th2. It is assumed that the threshold value th2 set here is 1000. As can be seen from Table 6, "Kyoto (sil-k k-
j jo oo ot to oT sil) ”is a partial word
Since the cumulative probability of “k kj” is equal to or greater than the threshold value th2, jo can be selected.

As described above, according to the third embodiment, the phoneme selection unit selects the phoneme for which the output probability calculation is required by using the calculation result of the cumulative probability in the matching unit. It is possible to prevent the recognition rate from decreasing and to reduce the processing amount in the output probability calculating unit.

(Fourth Embodiment) In the fourth embodiment of the present invention, similar phonemes are stored in the speech recognition apparatus according to the first embodiment as an input of the output probability calculation unit. A similar phoneme storage unit is added.

FIG. 13 shows the configuration of a voice recognition device according to the fourth embodiment. A selected phoneme storage unit 66, a phoneme information storage unit 61, a continuous phoneme information storage unit 64, a matching unit 62,
Each of the word dictionaries 63 has the same operation as that of the component having the same name in the first embodiment, and thus detailed description thereof will be omitted.

The output probability calculation unit 60 of this embodiment uses the characteristic parameter y _{t at} time t, the selected phoneme information stored in the selected phoneme information storage unit 66, and the similarity stored in the similar phoneme storage unit 67. The phoneme information is input and the output probability is calculated.

The similar phoneme storage unit 67 is a phoneme information storage unit 6.
If there are similar phonemes among all the phonemes stored in 1, it is said which is the similar phoneme (it does not matter even if there are two or more similar phonemes). It stores information. For example, phoneme H _p
And H _q are similar. The information is used as an input to the output probability calculator 60.

In the first embodiment, the output probability calculator 10
However, the output probability calculation is performed using only the phoneme information stored in the selected phoneme information storage unit 16. On the other hand, in the present embodiment, in order to further reduce the amount of processing in the output probability calculation unit 60, in the selected phoneme information selected by the phoneme selection unit 65, the phoneme selected lower by the selection criterion. If there is a phoneme that is similar to the upper selected phoneme, the output probability calculation is not performed for the lower selected phoneme, and the output probability calculation result of the similar upper phoneme is used instead. And

Here, the flow of processing of the output probability calculation unit 60 will be described with reference to the flowchart of FIG. The output probability calculation unit 60 inputs the characteristic parameter y _{t at} time t and the phoneme information in the selected phoneme information storage unit 66 (steps S41 and S42). Then, the output probability calculation unit 60
It is determined whether or not there is a similar phoneme H _{q in the} selected phonemes H _p (step S43). If the decision result in the step S43 is YES, the output probability calculation unit 60 decides whether or not the phoneme at the previous time of the phoneme H _p is higher than the output probability of the phoneme at the previous time of the similar phoneme H _q. Yes (step S4
4).

If the determination result of step S43 is NO or the determination result of step S44 is YES, b
_i ^p (y _t ) is calculated (step S45). On the other hand, the determination result of step S43 is YES, and step S43
If 44 of the determined result is NO, without calculating the b _i ^q (y _t), using the value of b _i ^p (y _t) as b _i ^q (y _t) (step S46). The output probability calculation unit 60 performs the above-described processing of steps S43 to S46 for all phonemes.

For example, at time t-1, the phoneme selection unit
H by 65_pAnd H_qIs selected, λ _p(Μ, σ) and λ
_q(Μ, σ) is stored in the selected phoneme information storage unit 66.
Suppose. However, H_pThe phoneme at the previous time is H_qBefore
It is assumed that the output probability is higher than that of tick phonemes. Smell at time t
If the output probability is calculated by using the
H stored in the selected phoneme information storage unit 66 according to the report.
_pAnd H_qB are similar to each other, so b_i ^p(Y_t)
Only calculated, b_i ^q(Y_t) ≒ b_i ^p(Y_t)
And by b_i ^q(Y_t) Is omitted. As a result,
It is possible to further reduce the amount of processing in the output probability calculator.
Become.

The above operation will be further described using a concrete example. Here, it is assumed that the information stored in the similar phoneme storage unit 67 is as shown in FIG. Also,
FIG. 16 shows the phonemes stored in the selected phoneme information storage unit 66 in the descending order of the phoneme output probability at the previous time.

As can be seen from FIGS. 15 and 16, the upper phoneme kj stored in the selected phoneme information storage unit 66 of FIG. 16 is similar to the lower phoneme ko. Therefore, when the output probability regarding the phoneme ko is obtained, the output probability regarding the phoneme kj is substituted instead of performing the calculation, so that the calculation amount in the output probability calculating unit 60 can be further reduced.

As described above, according to the fourth embodiment, if there is a similar phoneme among the phonemes selected by the phoneme selection unit, the output probability calculation result of the upper phoneme is set as the lower phoneme. By substituting the output probability, it becomes possible to further reduce the processing amount in the output probability calculation unit.

In the above-mentioned first to fourth embodiments,
An example of implementing the present invention as a voice recognition device has been described.
The present invention can be implemented as a program that causes a computer to execute the operations of the above-described units. In this case, by reading this program into a computer and executing the program, the computer can be operated as the voice recognition device according to the present invention.

[0087]

As described above, according to the present invention, by providing the continuous phoneme information storage unit that stores the phoneme information of the phonemes considered to occur at time t + 1, the phoneme having a high output probability is selected at time t. Then, the phoneme that is considered to occur at the next time of the phoneme is extracted, and at time t + 1, the output probability is calculated only for the extracted phoneme. As a result, even when the number of phonemes increases in large vocabulary speech recognition, it is possible to reduce the increase in the processing amount of the output probability calculation unit, and it is possible to realize real-time recognition processing within the frame length. .

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration of a voice recognition device according to a first embodiment of the present invention.

FIG. 2 is a flowchart showing an example (pattern 1) of the operation of the phoneme selection unit according to the first embodiment.

FIG. 3 is a flowchart showing another example (Pattern 2) of the operation of the phoneme selection unit according to the first embodiment.

FIG. 4 is an explanatory diagram showing an example of continuous phoneme information stored in a continuous phoneme information storage unit.

FIG. 5 is an explanatory diagram showing an example of continuous phoneme information regarding a phoneme sil.

FIG. 6 is an explanatory diagram showing an example of continuous phoneme information about phoneme sil-k.

FIG. 7 is an explanatory diagram showing an example of continuous phoneme information about phoneme sil-To.

FIG. 8 is a block diagram showing a configuration of a voice recognition device according to a second embodiment of the present invention.

FIG. 9 is a flowchart showing the operation of the continuous phoneme information optimizing unit according to the second embodiment.

FIG. 10 is an explanatory diagram showing an example of continuous phoneme information regarding a phoneme sil-k optimized by the word dictionary and the continuous phoneme information optimizing unit shown in Table 2.

FIG. 11 is a block diagram showing a configuration of a voice recognition device according to a third embodiment of the present invention.

FIG. 12 is a flowchart showing the operation of the phoneme selection unit according to the third embodiment.

FIG. 13 is a block diagram showing the configuration of a voice recognition device according to a fourth embodiment of the present invention.

FIG. 14 is a flowchart showing the operation of the output probability section according to the fourth embodiment.

FIG. 15 is an explanatory diagram showing an example of information stored in a similar phoneme storage unit according to the fourth embodiment.

FIG. 16 is an explanatory diagram showing phonemes stored in the selected phoneme information storage unit according to the fourth embodiment in the order of higher output probability of phonemes at the previous time.

FIG. 17 is a block diagram showing a conventional configuration example of a voice recognition device.

[Explanation of symbols]

10, 20, 40, 50, 60 Output probability calculator 11, 21, 41, 51, 61 Phoneme information storage unit 12, 22, 42, 52, 62 Matching unit 13, 23, 43, 53, 63 word dictionary 14, 44, 54, 64 Continuous phoneme information storage brain 15, 45, 55, 65 Phoneme selection section 16, 46, 56, 66 Selected phoneme information storage unit 47 Continuous phoneme information optimization unit 67 Similar Phoneme Storage

Claims (14)

[Claims] 1. A speech recognition apparatus using a hidden Markov model, comprising: phoneme information storage means for retaining, as a hidden Markov model, feature parameters of speech for each frame length for all phonemes; and retained in the phoneme information storage means. An output probability calculation unit that calculates the output probability of each phoneme in the frame from the hidden Markov model that is being used and the feature parameters of the input speech for each frame length, and the phoneme of the current frame and the phoneme that is considered to occur in the next frame. Based on the continuous phoneme information storage unit that holds the relationship with the selected phoneme information storage unit, the output probability calculated by the output probability calculation unit, and the continuous phoneme information of the continuous phoneme information storage unit, in the next frame. Only phoneme information relating to a phoneme that is considered to occur is selected from the phoneme information storage means, and stored in the selected phoneme information storage means. A phoneme selection unit to be stored, a word dictionary storage unit that holds a phoneme sequence of each recognized word, and a matching unit that accumulates the output probability calculated by the output probability calculation unit for each word based on the word dictionary. A voice recognition device comprising: 2. The phoneme selection unit sorts the output probabilities calculated by the output probability calculation unit in descending order of value,
Selecting a predetermined number of phonemes in the upper order among them, and selecting a phoneme that is considered to occur in the next frame when the current frame is that phoneme based on the continuous phoneme information of the continuous phoneme information storage means; Item 1. A voice recognition device according to item 1. 3. The phoneme selecting means provides a predetermined threshold value to the output probability calculated by the output probability calculating means, and selects a phoneme whose output probability exceeds the threshold value.
2. The speech recognition device according to claim 1, wherein a phoneme that is considered to be generated in the next frame when the current frame is the phoneme is selected based on the continuous phoneme information in the continuous phoneme information storage means. 4. A continuous deletion of phoneme relationships that are considered not to occur continuously among the continuous phoneme information stored in the continuous phoneme information storage means based on the phoneme series stored in the word dictionary storage means. The voice recognition device according to claim 1, further comprising a phoneme information optimization unit. 5. The phoneme selection unit refers to the cumulative probability of each word obtained by the matching unit, in addition to the output probability calculated by the output probability calculation unit and the continuous phoneme information, and the output probability. Regarding a phoneme that is not a selection target when the reference is based on, the phoneme is selected if it is considered that the phoneme of the current frame is highly likely when based on the cumulative probability. The voice recognition device according to any one of claims. 6. A similar phoneme storage means for storing similar phonemes is further provided, wherein the output probability calculation means inputs a similar phoneme in the previous period, and regarding the satisfaction degree of the selection criterion among the phonemes selected by the phoneme selection means. If the lower first phoneme is similar to the higher second phoneme, then the first
4. The speech recognition apparatus according to claim 1, wherein the output probability calculated for the second phoneme is used for the phoneme of No. 1 without using the output probability calculation. 7. The speech recognition according to claim 1, wherein the hidden Markov models stored in the phoneme information storage unit share common phoneme states. apparatus. 8. A phoneme information storage unit that holds a feature parameter of speech for each frame length as a hidden Markov model for all phonemes, and a continuation that holds the relationship between the phoneme of the current frame and the phoneme considered to occur in the next frame. A phoneme information storage unit, a selected phoneme information storage unit, and a hidden Markov model stored in the phoneme information storage unit, which is read by a computer that can use a word dictionary storage unit that holds a phoneme sequence of each recognized word, From the characteristic parameters of the input speech for each frame length, the output probability calculation process of calculating the output probability of each phoneme in the frame, the output probability calculated by the output probability calculation process, the continuation of the continuous phoneme information storage means Based on the phoneme information, only the phoneme information regarding the phoneme considered to be generated in the next frame is stored in the phoneme information storage means. A phoneme selection process of selecting and storing in the selected phoneme information storage unit, and a matching process of accumulating the output probability calculated by the output probability calculation process for each word based on the word dictionary are executed on the computer. A voice recognition program characterized by: 9. The phoneme selection process sorts the output probabilities calculated by the output probability calculation process in descending order of value,
In the process of selecting a predetermined number of upper phonemes among them, and selecting a phoneme that is considered to occur in the next frame when the current frame is that phoneme based on the continuous phoneme information of the continuous phoneme information storage means. The voice recognition program according to claim 8, which is provided. 10. The phoneme selection process provides a predetermined threshold value to the output probability calculated by the output probability calculation process,
A phoneme whose output probability exceeds the threshold value is selected, and a phoneme considered to be generated in the next frame when the current frame is the phoneme is selected based on the continuous phoneme information of the continuous phoneme information storage means. The speech recognition program according to claim 8, which is a process for performing. 11. A continuous deletion of phoneme relations, which are considered not to occur continuously, among the continuous phoneme information stored in the continuous phoneme information storage means based on the phoneme sequence stored in the word dictionary storage means. The phoneme information optimizing process is further executed by the computer.
10. The voice recognition program according to any one of 10. 12. The phoneme selection process refers to the cumulative probability of each word obtained by the matching process in addition to the output probability calculated by the output probability calculation process and the continuous phoneme information, and the output probability The process of selecting a phoneme that is not a selection target based on the above, when it is considered that there is a high possibility that the phoneme is the current frame based on the cumulative probability. -10
The speech recognition program according to any one of 1. 13. The computer can further utilize a similar phoneme storage means for storing a similar phoneme, wherein the output probability calculation process inputs a similar phoneme in the previous period, and outputs the phoneme selected in the phoneme selection process. Among them, if the first phoneme belonging to the lower rank with respect to the satisfaction degree of the selection criterion is similar to the second phoneme belonging to the higher rank, the first phoneme
The speech recognition program according to any one of claims 8 to 10, which is a process of using the output probability calculated for the second phoneme without performing the output probability calculation for the phoneme of. 14. The computer according to claim 8, wherein the computer refers to a hidden Markov model in which similar phoneme states are shared as a hidden Markov model stored in the phoneme information storage means. The speech recognition program according to claim 1.