JPH02293899A - Voice recognizing device - Google Patents

Abstract

PURPOSE:To improve recognition performance by providing a soundless section detecting section which detects soundless sections and stores the frame positions thereof and a similarity calculation control section which executes the computation of cumulative distances in accordance with a matching route. CONSTITUTION:The soundless section detecting section 11 which detects the soundless sections from an energy system outputted from a voice analyzing section 10 and stores the start frame and final frame thereof and the similarity calculation control section 12 which executes the the pattern matching computation with the start points obtd. in the respective soundless section detection of the input patterns and standard patterns and the first soundless section points as start point candidate points are provided. The pattern matching calculation is controlled with the start frame of the final soundless section on the time series of the voice patterns together with the terminals of the respective standard patterns and input voice patterns as end point candidates as well by taking the missing of the end of words into consideration, by which the similarity of both is determined. The accurate recognition is executed in this way even if the missing of the end of the words arises.

Description

DETAILED DESCRIPTION OF THE INVENTION Field of Industrial Application The present invention relates to a speech recognition device that derives recognition results through pattern matching between an input speech pattern and each standard pattern.

2. Description of the Related Art In this type of apparatus, a method is used in which the degree of similarity between an input speech pattern and each standard pattern stored in a dictionary is calculated, and the standard pattern with the maximum degree of similarity is determined as the recognition result. In order to calculate the similarity between two speech patterns, dynamic programming is used to perform pattern matching (hereinafter referred to as DP matching), which non-linearly expands and contracts the time axes of the two patterns. has been done. In particular, word speech recognition devices achieve high recognition rates using this DP matching method. For example, it is described in the following document.

Ant゜S Chiha゛ ``Evening Inami Fuku゜Following゜Think゜Off゜Timing゜tion゜Snaken Wort゛Recognition゛Eeeeeee Trance Acoustifuku, Speech, Sequential Music'' (
H. Sakoe and S. Ch1ba,
”Dynamic programming op.
timizat1on for sporken
IEEE
trans. Acoustlc, Speech, Si
gnal Processing), vol. A
SSP-27 pp, 3311i-349.1979
An example of a conventional speech recognition device will be described below with reference to the drawings.

In FIG. 7, reference numeral 1 denotes a speech analysis unit which converts the input speech signal into a time series of feature vectors and detects the uttered speech section. 2 is a word dictionary in which a time series of feature vectors of words to be recognized is stored in advance as a standard pattern. 3 is a similarity calculation unit that calculates the similarity between the input speech pattern and each standard pattern by pattern matching. 4
The determination unit determines the standard pattern with the maximum similarity as the recognition candidate speech.

Next, the operation of the above speech recognition device will be explained.

First, an audio signal input through a microphone or the like is converted into a time series of characteristic parameters such as linear prediction coefficients and cepsdrum coefficients, and an energy value series in the audio analysis unit 1.
A portion where the energy value exceeds a certain value is detected as a voice section. Here, it is assumed that the time series of feature vectors of the speech section portion of each recognition target word obtained by the speech analysis section 1 has already been stored in the word dictionary 2 as a standard pattern. Next, the similarity calculation unit 3 performs DP matching calculation between the time series of feature vectors of the input speech and the time series of feature vectors of each standard pattern stored in the word dictionary 3, and stores the respective similarities. Finally, the determination unit 4 recognizes the standard pattern that gives the maximum similarity among the similarities obtained by the similarity calculation unit 3. It is determined as a candidate voice and output as a recognition result to the outside.

Problems to be Solved by the Invention However, in the above-mentioned speech recognition device, if the section of the input speech is incorrectly detected (the section is detected due to a missing syllable at the beginning or end of the speech), it may cause erroneous recognition. often occurs. As an example of missing the beginning of a word, in the case of a voice that starts with a voiceless consonant (Fukuoka, Fukushima, Fukui, etc.), for example, rFUKUO.
When we consider the case of uttering ``KA'' (Fukuoka), the vowel rUJ of rFUJ at the beginning of the word is often devoiced, and as a result, the energy level at the beginning of the word becomes low, and when detecting the speech interval, the vowel rUJ of rFUJ is often devoiced. In many cases, only the However, when people vocalize, they tend to consciously try to pronounce it clearly in environments with high ambient noise levels.
The vowel rUJ of [U] may be voiced, and in this case it is detected as rFUKUOKAJ or rKUO.
The section KAJ may be detected. In this case, when there is a possibility that the beginning of a word may be missing depending on the situation during pronunciation, if the speech interval detection results of the standard pattern registered in advance in the dictionary and the input speech pattern are different (for example, the beginning of the word is rFUKUOKA).
J and the other is rKUOKA), the conventional DP matching method cannot cope with this problem and causes misrecognition. The same can be said about missing word endings.

Further, in conventional speech recognition devices using DP matching, pattern matching is performed over the entire speech pattern, so that erroneous recognition may occur due to extreme matching.

For example, considering Oita (OOITA) and Osaka (OOSAKA), the former has two energy peaks and the latter has three peaks of energy in their vocal patterns, and although they are clearly different energetically, the phonemes are similar. Therefore, erroneous recognition may occur. For this reason, there is a method of detecting the peaks and troughs of energy in a voice pattern to limit the pattern matching path, but if a detection error occurs, it may cause erroneous recognition.

In view of the above-mentioned problems with conventional speech recognition devices, the present invention is capable of accurately recognizing speech that may be missing the beginning or end of a word, and furthermore prevents misrecognition between similar words as much as possible. An object of the present invention is to provide a speech recognition device that can improve the project performance when inputting foreign words.

The present invention aims to solve the problems of the prior art.

Means for Solving the Problems A speech recognition device according to a first aspect of the present invention (claim 1) includes a speech analysis section that outputs a time series of feature vectors including an energy sequence from an input speech, and a speech analysis section that outputs a time series of feature vectors including an energy sequence from an input speech. A silent section detecting unit detects a silent section from an energy series and stores its start frame and final frame, and the start point is the start point obtained as a result of voice section detection and the first silent section point of the input pattern and the standard pattern, respectively. The present invention is characterized by comprising a similarity calculation control unit that performs a pattern matching operation on candidate points.

A speech recognition device according to a second aspect of the present invention (claim 2) includes a speech analysis section that outputs a time series of feature vectors including an energy series from an input speech, and a speech analysis section that extracts a silent interval from the energy series output from the speech analysis section. A silent section detector detects and stores the start frame and the last frame, and a pattern matching operation that limits the matching path at the intersection of the final frame of the silent section of each standard pattern and each input audio pattern is performed using a The first step to run on the route
and a second similarity that executes a pattern matching operation using the end point obtained as a result of voice section detection and the start frame of the last silent section as end candidate points for each of the input pattern and the standard pattern. It is characterized by comprising a calculation control section.

A speech recognition device according to a fourth aspect of the present invention (claim 4) includes a speech analysis unit that outputs a time series of feature vectors including an energy sequence from input speech, and a silent interval from the energy series output from the speech analysis unit. A silent section detector that detects and stores the start frame and the final frame, and a pattern matching operation that limits the matching path at the intersection of the final frame of the silent section of each standard pattern and each input audio pattern. The first to run for
and a second similarity that executes a pattern matching operation using the end point obtained as a result of voice section detection and the start frame of the last silent section as end candidate points for each of the input pattern and the standard pattern. The present invention is characterized by comprising a calculation control unit and a third similarity calculation control unit that executes a pattern matching calculation using the start ends of each of the input pattern and the standard pattern and the end frame of the last silent section as start end candidate points. .

Effects The first aspect of the present invention has the above-described configuration, and takes into account missing word endings and uses the start frame of the last silent section in the time series of the audio pattern as a candidate for the end point, as well as the end of each standard pattern and the input audio pattern. By controlling pattern matching calculations and determining the degree of similarity between the two, even if the ending of a word is missing, it can be recognized with high accuracy.

In addition to the effects of the first aspect of the invention described above, the second aspect of the present invention has the configuration described above. Even if the detection is incorrect,
The rejection performance when non-target words are input is improved, and misrecognition between similar words can be minimized.

According to the fourth aspect of the present invention, in addition to the effects of the second aspect of the present invention described above, the fourth aspect of the present invention has the configuration described above, and also takes into account the omission of the beginning of a word and performs the same on the chronological order of the voice pattern as well as the beginning of each standard pattern and input voice pattern. By controlling the similarity calculation using the end frame of the first silent section as a starting point candidate, even if the beginning of a word is missing, it can be recognized with high accuracy.

Examples Examples of the present invention will be described below with reference to the drawings.

FIG. 1 is a block diagram of a speech recognition device according to a first embodiment of the present invention. In FIG. 1, reference numeral 10 denotes a voice analysis section, which has the same function as the voice analysis section 1 in FIG. Reference numeral 11 denotes a silent section detecting section, which detects a silent section from a time series of audio energy values and stores its starting frame. Reference numeral 12 denotes a similarity calculation control unit which controls the similarity calculation to be performed only at a plurality of terminal candidate points of the speech pattern. A pattern matching control unit 13 controls the overall pattern matching calculation.

14 is a word dictionary that stores the time series of feature vectors of words to be recognized and frame positions of silent sections. 15 is a determination unit that determines the standard pattern with the maximum similarity as the recognition candidate speech. FIG. 2 is a pattern diagram for explaining the device shown in FIG. 1.

Next, the operation of the speech recognition device in the first embodiment will be explained.

First, a voice signal inputted through a microphone or the like is converted from analog to digital by the voice analysis section 10 into a time series of voice feature vectors and an energy value series, and the uttered voice portion is detected as a section. Next, the silent section detecting section 11 detects, as a silent section, a section in which the energy value is lower than a predetermined threshold for a certain period of time from the energy value series obtained by the speech analyzing section 10, and Store the starting frame of the interval. It is assumed that the word dictionary 14 has already stored the time series of feature vectors of speech to be recognized and information on silent intervals.

Here, we consider the following equation as a recurrence equation for pattern matching.

g(1.1):d(1.1)
- - - (+)g(1 +J)=mln(g(1
-1+J) +g(1-1+j-1)+g(+-1+j
−2))+d(1,J) 1− −
- (2) Here, g (I, j) and d (Xl j) are the i frame of the input audio pattern and the j frame of the standard pattern, respectively.
It represents the cumulative distance in a frame and the distance between feature vectors. Note that the recurrence formula may be other than the above formula. Furthermore, for simplicity, we will not consider limiting the matching path by the matching window. As an example, consider a case where pattern matching is executed between an input voice pattern and a standard pattern as shown in FIG. In FIG. 2, the envelope for the speech pattern shows the change in energy over time. Here, the start frame of the last silent section of the input audio pattern is 1 0E1, and the start frame of the last silent section of the standard pattern is j OE.

By the way, especially when the speech to be recognized ends with a voiceless consonant or the like, there is a possibility that the final section of the word will be detected incorrectly in both the standard pattern and the input speech pattern.

In this way, if there is a possibility that the ending of the word is missing in both patterns, the end points of the pattern matching path are (I.J), (I QEwJ), as shown in Figure 2.
There are three possible points (I. joε), and by selecting the optimal path from the three matching paths that end at these three points, pattern matching that can deal with missing word endings can be realized.

The pattern matching control unit 13 starts calculating the cumulative distance according to recurrence formula (2). First, with i=1, the cumulative distance is calculated in the vertical direction from the first frame of the standard pattern, that is, j=1. When this is completed, the same process is continued while sequentially incrementing i. When the processing up to i=I is completed, the similarity calculation control unit 12 calculates (I. J) (I,
Let the smaller cumulative distance in Joε) be the similarity between the two patterns.

Here, since we use a slope restriction like the nonlinear equation (2), the cumulative distance obtained at point (I, J) is already the second
This is the result when the optimal route is selected from routes a and b in the figure. The degree of similarity obtained in this way becomes the distance between patterns when the optimal route is selected from among the three types of matching routes considering the three end point candidate points in FIG. 2. Note that the similarity calculation is performed using the point (I. J) (I, Jo
ε) (ICIEIJ) may be normalized by the length of the input speech pattern, and the minimum value among the values may be taken as the similarity between patterns.

After all the similarities between each standard pattern stored in the word dictionary 14 and the input speech pattern are calculated, the determination unit 1
In step 5, the standard pattern with the highest degree of similarity is determined as the recognition candidate speech and output to the outside.

As described above, according to the present embodiment, the silent section detection unit 11 detects a silent section of speech, and the similarity measure calculates the similarity between patterns from the cumulative distance at a plurality of terminal candidate points in consideration of missing word endings. By being equipped with the calculation control unit 12, even if the ending of a word is missing in either the standard pattern or the input speech pattern, calculations can be performed that consider all possibilities with a single pattern matching operation, reducing the amount of calculation. It becomes possible to accurately recognize speech with unstable endings without increasing the number of words.

Next, a speech recognition device according to a second embodiment of the present invention will be described with reference to the drawings.

FIG. 3 is a block diagram of a speech recognition device according to a second embodiment of the present invention. In FIG. 3, 20 is a speech analysis unit that extracts the feature vector and energy time series of hoarseness;
Reference numeral 21 denotes a silent interval detection unit that detects the start frame and end frame of a silent interval in a voice pattern, 22 a first similarity calculation control unit that controls pattern matching calculation that limits the matching path in the silent interval, and 23 each A back pointer storage section 24 stores the cumulative distance at a point as well as which intersection of silent section points the matching path has passed before; A similarity calculation control unit, 25 is a pattern matching control unit that controls the overall cumulative distance calculation, 26 is a word dictionary that stores feature vectors of standard patterns and frame positions of silent sections, and 27 recognizes the standard pattern with the maximum similarity. This is a determination unit that determines that the candidate is a candidate. FIG. 4 is a pattern diagram for explaining the device shown in FIG. 3.

Next, the operation of the speech recognition device in the second embodiment will be explained.

First, the input audio signal is converted by the audio analysis section 20 into a time series of feature vectors and a time series of energy values. Next, the silent section detecting section 21 detects, for example, a section where the energy value is less than a certain value from the energy value series obtained by the speech analyzing section 20 as a silent section, and stores the start frame and end frame thereof. It is assumed that the time series of feature vectors of the speech to be recognized and the start and end frames of each silent period are stored in the word dictionary 26 in advance. Next, the first similarity calculation control unit 22 executes a pattern matching calculation between the input audio pattern and each standard pattern in which the matching path is constrained at the end frame of the silent interval detected by the silent interval detection unit 21. To go. Also second
The similarity calculation control unit 24 controls the cumulative distance calculation using the end of the input speech pattern, the standard pattern, and the start frame of the last silent section of each as end candidates. For simplicity, the pattern matching operation will be described below for the case where there are two silent sections in the input pattern and one silent section in the standard pattern.

Here, the start and end of the input pattern are respectively Is(
1), Is(1), and the starting and ending ends of the standard pattern are Ji(1) and J2(J), respectively.

Further, the end frame of the silent section of the input pattern is I++ Is from the one closest to the start, the start frame of the last silent section is I2°, and the standard pattern is similarly J+, J2, j. These situations are shown in FIG. Here, for the sake of simplicity, we will not consider restrictions due to matching windows.

Pattern matching is performed by constraining the matching path on the intersection of the input pattern and the end frames of the silent sections detected on the standard baton, but if the silent section is detected correctly (if it is not carried out) Taking this into account, it is necessary to consider the four types of abcd routes shown in FIG. The recurrence formulas for pattern matching are the same as formulas (1) and (2) in the first embodiment. Here, let the back pointer at point (Lj) be b(i, j) and the initial value is b(1. 1
) (1.1).

First, the pattern matching control unit 25 sets i=1, j=1, increments j, and calculates the cumulative distance i according to the recurrence formula.
! When tftg is calculated and the processing is completed until j=J, i is incremented and the same processing is continued.

The back pointer at each point (i, j) is determined as follows.

i ” When I k and j = J+, b (1+ J) (Ik+ J+)i≠Ib
Or when j f−L, the point (i, j) is Ik<i:SIb−+ and J+<j
sJ+. Let b(i,j) be the content of the back pointer corresponding to gO that gives the minimum value in the first term on the right side of recurrence formula (2) when it is within the range of 1. However, here, the point (I., J,) that can be taken as b (i, j) is m = k and H: aL
Or n=1 and m≦k. The back pointer b(i, j) thus obtained is stored in the back pointer storage section 23.

When the processing up to i=I is completed, the subsequent processing is performed by the second similarity calculation control section 24 and the judgment section 27, respectively, which perform the same operations as the similarity calculation control section and the judgment section of the first embodiment. The degree of similarity between the input pattern and the standard pattern obtained here is the degree of similarity when the optimal route is selected from the four types of routes shown in FIG. 4.

As described above, according to the present embodiment, the calculation of the similarity between the input pattern and the standard pattern is controlled by the silent section detection unit 21 that detects silent sections of audio and pattern matching in which the path is restricted by the intersections of the silent sections. A first similarity calculation control unit 22, a back pointer storage unit 23 that stores intersection points of silent sections passed in the past, and a second similarity calculation control unit that performs a pattern matching operation in consideration of a plurality of termination candidates. By providing 24, it is possible to consider all possibilities of path restriction in a single pattern matching operation, and at the same time, by calculating the similarity with multiple terminal candidates that take into account missing endings, it is possible to eliminate silence. Accurate recognition is possible even when a section is incorrectly detected, and a voice pattern with a missing ending can be recognized correctly regardless of whether it occurs in the input pattern or the standard pattern.

Next, a speech recognition device according to a third embodiment of the present invention will be described with reference to the drawings.

FIG. 5 is a block diagram of a speech recognition device according to a third embodiment of the present invention. In FIG. 5, the voice analysis section 30,
Silent section detection section 31, back pointer storage section 33, first similarity calculation control section 32, second similarity calculation control section 3
4. The pattern matching unit 38, the word dictionary 37, and the determining unit 38 have the same functions as the units with the same names corresponding to those in FIG. 3 of the block diagram of the second embodiment. 35 is the third
The similarity calculation control unit controls the pattern matching calculation using the starting edges of the input pattern and the standard pattern and the end frame of the first silent section as starting edge candidates. FIG. 6 is a pattern diagram for explaining the device shown in FIG. 5.

Next, the operation of the speech recognition device in the third embodiment will be explained.

First, the input audio signal is converted by the audio analysis section 30 into a time series of feature vectors and a time series of energy values. Next, the silent section detecting section 31 detects, for example, a portion where the energy/L/Gy value is less than a certain value from the energy value series obtained by the speech analyzing section 30 as a silent section, and stores the start frame and end frame of the section. . It is assumed that the time series of feature vectors of the speech to be recognized and the start and end frames of each silent section are stored in the word dictionary 37 in advance.

Next, control of pattern matching calculation will be explained.

Here, we will consider the case where there are two silent sections each in the introductory pattern, tone pattern, and standard pattern, as shown in FIG. Here, let the start and end of the input/f turn be Is(1) and I, respectively, and Je(1) and J for the standard pattern as well. Also input the end frame of the silent section in chronological order and the start frame of the last silent section in order for z<turn III Ill I2
'JI*Js+J2' for the target Ml Hataan.

By the way, when the speech to be recognized starts with an unvoiced consonant or the like, there is a possibility that the initial section of the word will be detected incorrectly in both the standard pattern and the input speech pattern.

In this way, if there is a possibility that the beginning of a word is missing in both patterns, (1, 1) is used as the starting point of the pattern matching path.
Three points, (L, 1) and (1, J+), are considered, and if the optimal route is selected from all routes starting from these three points, pattern matching that can deal with missing beginnings of words can be realized. Equations (1) and (2) of the first embodiment are used as recurrence equations for pattern matching.

First, the third similarity calculation control unit 35 sets the following equation as an initial setting.

d(1,j)=φ However, j≠1, J, d(i.1)=■ However, i > I + As a route starting from point (I+,1), point (1,1
) to point (II, 1) is considered as a pseudo route by allowing it on the matching route.

This setting imposes a restriction so that the pattern matching route does not start from any point other than the three starting end candidate points. Thereafter, calculation of cumulative distance is started from i=1 according to recurrence formula (2). Thereafter, the first similarity calculation control unit 32 executes a pattern matching calculation in which the matching path is limited at the intersection of the silent sections, taking into account the detection error of the silent sections.

Further, the second similarity calculation control unit 34 controls bang matching using the end of the input pattern, the standard pattern, and the start frame of the last silent section of each as end candidates.

Considering a plurality of start and end candidates and silent section detection errors as described above, many possible matching paths can be considered as shown in FIG. 6.

The processing contents performed by the first similarity calculation control unit 32 and the second similarity calculation control unit 34 from this point onwards are the same as those of the first similarity calculation control unit 22 and the second similarity calculation control unit 34 of the second embodiment, respectively. Since it is the same as the calculation control section 24, the explanation will be omitted here.

After the degrees of similarity with all the standard patterns stored in the word dictionary 37 are calculated in this way, the determination unit 38 determines the standard pattern with the maximum degree of similarity as the recognition candidate speech and outputs it to the outside.

As described above, according to the third embodiment, the silent section detecting section 31 detects a silent section of audio, and the first silent section detecting section 31 executes pattern matching to limit a matching path at the intersection of the silent sections.
a second similarity calculation control unit 34 that performs a pattern matching calculation taking into account a plurality of end point candidates; 3 similarity calculation control unit 3
By providing 5, it is possible to recognize accurately even when input audio is missing the beginning or end of a word.
In addition, even if a silent section is detected incorrectly, all possibilities of route restriction can be considered in one pattern matching, improving the recognition performance between similar words and rejecting as much as possible even when untargeted speech is input. .

Effects of the Invention As described above, the first aspect of the present invention includes a silent section detection unit that detects a silent section in a speech pattern and stores its frame position, and a terminal candidate point for multiple pattern matching that takes into account missing word endings. Equipped with a similarity calculation control unit that performs cumulative distance calculations according to the matching path, it is possible to eliminate missing endings from either the input speech pattern or the standard pattern when there is an unstable ending in the speech to be recognized. Even so, it is possible to provide a speech recognition device that can perform all pattern matching in consideration of missing beginnings of words in one calculation, and can improve recognition performance without substantially increasing the amount of calculation.

In addition, the second invention performs a pattern matching operation that limits the matching path at the intersection of the final frames of the silent sections of the standard pattern and the input audio pattern on many possible paths, taking into account silent section detection errors. a first similarity calculation control unit for calculating the cumulative distance at each point; and a back pointer for storing the intersection point of the end frame of the silent section that the matching route has passed through before reaching that point when calculating the cumulative distance at each point. As a result, even if a silent section detection error occurs, the pattern matching operation along the matching path that may match the energy envelope can be executed in one calculation. It is possible to provide a speech recognition device that can improve performance and minimize misrecognition between similar words.

Further, the fourth aspect of the present invention is provided with a third similarity calculation control unit that executes a pattern matching operation using the start point of the speech pattern and the end frame of the first silent section as start point candidate points, taking into account the beginning of a word drop. As a result, all pattern matching operations that take into account missing word endings can be performed in a single calculation, regardless of whether the missing beginning of a word occurs in the standard pattern or the input speech pattern. Recognition performance is improved with almost no increase in the amount of calculation. It is possible to provide a speech recognition device that allows

[Brief explanation of drawings]

FIG. 1 is a block diagram of a speech recognition device according to a first embodiment of the present invention, FIG. 2 is an explanatory diagram of the operation of the same device, FIG. 3 is a block diagram of a speech recognition device according to a second embodiment, and FIG. 5 is a block diagram of the speech recognition device of the third embodiment, FIG. 8 is a diagram illustrating its operation, and FIG. 7 is a block diagram of the conventional speech recognition device. be.゛30●●●Speech analysis section, 31●●●Silent section detection section, 3
2●●Fist first similarity calculation control unit, 33●●●back pointer storage unit, 34●●●second similarity calculation control unit,
35●●φ third similarity calculation control unit, 36●●●pattern matching control unit, 37●●●word dictionary, 38●●●
Judgment department. Name of agent Patent attorney Shigetaka Awano and 1 other person Figure 1 Voice input #fruit output Fuzufuzu S to fruit output Figure 7 Figure 7 Turn at Otoichi no Hachijo

Claims (4)

[Claims] (1) A speech analysis section that outputs a time series of feature vectors including energy sequences from the input speech signal and detects uttered speech sections; a silent section detecting section that detects a speech portion and stores its start frame, and detecting the end of the speech section detected by the speech analyzing section of each of the standard pattern and input speech pattern stored in advance as speech to be recognized, and the silence section. Similarity is calculated by pattern matching between the standard pattern and the input speech pattern, using the end frame of the last silent section in the speech pattern time series detected by the section detection unit as the end frame of the speech pattern. A speech recognition device comprising: a calculation control section. (2) A speech analysis unit that outputs a time series of feature vectors including energy sequences from the input speech signal and detects uttered speech sections; A pattern is created at the intersection of the silent section detecting section that detects the audio part and storing its start frame and ending frame, and the end frame of the silent section obtained by the silent section detecting section of the input pattern and each standard pattern stored in the dictionary in advance. a first similarity calculation control unit that performs pattern matching distance calculation for limiting the matching path on a matching path limited by intersections of end frames of all possible silent periods in consideration of a silent period detection error; , for each standard pattern and input audio pattern, the end of the audio section detected by the audio analysis section and the start frame of the last silent section in the audio pattern time series detected by the silent section detection section are used as the audio pattern. A speech recognition device comprising: a second similarity calculation control unit that controls similarity calculation between a standard pattern and an input speech pattern as a terminal candidate point. (3) When calculating the cumulative distance at each point, the first similarity calculation control unit uses a back pointer that stores the intersection point of the end frame of the silent section that the pattern matching path has passed before reaching that point. 3. The speech recognition device according to claim 2, further comprising a speech recognition device. (4) A speech analysis unit that outputs a time series of feature vectors including energy sequences from the input speech signal and detects uttered speech sections; A silent section detecting section detects a speech portion and stores its start frame and end frame; and a silent section detecting section detects a speech portion and stores its start frame and end frame; A third control unit that controls similarity calculation between the standard pattern and the input speech pattern by using the starting point and the end frame of the first silent section in the speech pattern time series detected by the silent section detection unit as starting point candidate points for each speech pattern.
The pattern matching distance calculation, which limits the pattern matching path at the intersection of the end frame of the silent section obtained by the similarity calculation control section of the input pattern and the silent section detection section of each standard pattern, takes into account silent section detection errors. The first similarity calculation control section performs the matching on a matching path limited by the intersections of the end frames of all possible silent sections, and the speech analysis section detects each standard pattern and input speech pattern. the end of the voice section,
A second degree of similarity that controls similarity calculation between the standard pattern and the input voice pattern, using the start frame of the last silent period in the voice pattern time series detected by the silent zone detection unit as the end candidate point of the voice pattern. A speech recognition device comprising: a calculation control section.