JPS6356700A - Continuous voice recognition equipment - Google Patents

Abstract

(57) [Abstract] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a continuous speech recognition device suitable for recognition processing of continuously uttered speech.

(Prior Art) Conventionally, various methods (hereinafter referred to as continuous word recognition techniques) for recognizing a string of words uttered continuously (hereinafter also referred to as continuous words) have been proposed. These can basically be divided into two types. One method uses units smaller than words (for example, syllables, phonemes, etc.) as the basic units of recognition, and uses these recognition results to identify gradually larger units such as words, phrases, and sentences. The other type uses words as basic units. Now, the basic unit! In the case of a P-word, the first step in the recognition process is to determine the interval in which a certain unit iΔ exists from among the sequences. ! 11. A common technique for detecting and identifying relatively long speech sections such as words is a method based on pattern matching, particularly a matching technique based on dynamic programming (hereinafter referred to as DP). Furthermore, there are techniques called continuous DP and lhp that are suitable for continuous word recognition. This method searches for the presence of predetermined words by continuously performing pattern matching on a given human voice, and is called word spotting. The idea of realizing continuous word recognition using this continuous DP is to detect and identify certain P-words from human speech, find word interval candidates (continuous DP), and determine the optimal word string from these candidates. There is a two-stage process.

What is the technology that has realized this idea in the past?
e28. +982). After describing the continuous DP 1000 method below, the basic idea of the conventional continuous word recognition technology using continuous DP will be explained.

The processing technology for continuous DP is as follows.

Characteristic parameter series of sound 7 at a time interval called C frame period (=human pattern) for human-powered voices
V(p, i) (where p=l, 2,3°..., P and i=1.2,3...., I) is calculated. In general, the characteristic pattern of speech is a two-dimensional pattern centered on the time axis and the spectral frequency.

Here, V(p, i) means the second value in the j-th frame, which is the first vector. Similarly to the input speech, the feature parameter series Sn (
p, j) (j=1.2゜..., , Jn, p+=l
, 2.3. ..., P) is calculated in advance. This feature parameter series So (ρ, j) is called a standard pattern.

The following symbols are defined here.

Dn (i, j) ・The first frame from frame m of the manual pattern “1” and the standard pattern S, (p, j)
The minimum value B, , (i, j): Dn (i
, j): the first frame of the human pattern and the standard pattern S. FIG. 3 shows the distance between (ρ, j) and the j-th frame in DP matching performed using the dynamic programming method used in this word spotting method.

When using the DP bus shown in FIG. 3, Dn (i.

J), B, (i, j) are determined by the following formula.

Dn (il, +) - Dn (+, j) + dn (il, +)
H+ HHH+ + (1) B1, (i, j)-B
n(i, j-1) ・= (2) However, i-arg
min D, , (i, j-1) i-2≦i'≦j. Equations (1) and (2) are recurrence equations, but by performing the following processing, any 1. In the j. (i,
J), Bn (t, J) can be calculated.

Performs initial setting processing for the start of a 0-person Kapatan.

D, (-1,j)=Dn (0,j)=Mj=1.2.
...+J11 However, M is a sufficiently large positive value, and J is the length of the standard pattern.

■Execute ■~■ for i-1, 2, ..., ■. However, ■ is the length of the large kapatan.

■When starting matching from the i-th frame,
That is, initialization processing for j=1 is performed.

D. (i, 1) = d, (i,
l) ・ ・ ・ −(3) B,
(i, 1) = i ・ ・−= ・ ・−−(4
)■j=2.3. ..., Equation (1) for Jo,
Execute (2).

Now, by the processing described above, Dn(i.

Jn), B, (i, Jn) can be calculated. on here
Considering the meaning of (i, Jn), the input pattern Bn (
i, Jn)~i-th frame and standard pattern 1~
This is the inter-pattern cumulative distance with the time axis normalized from the J-th frame.

Therefore, this stone. If the value of (i, Jn) is smaller than a certain predetermined value, it is a manual pattern B. (i.

, Jo) It is assumed that the same word as the standard pattern is detected between the i-frame and the i-frame.

It has been explained that by the processing described above, a specific word section existing at an arbitrary position in continuous speech can be detected.

Next, a conventional continuous word recognition technique using this result will be briefly described.

The processing described here is based on the document ``Large word rate continuous speech recognition system based on continuous DP'' (Kouchi Honen Ql Gakkai OnjI Study Group Material 582-11).

First, conditions for detecting a certain word n as a result of continuous DP are determined, for example, as shown below.

fE, (i) <If the word n exists consecutively for at least a frame from the time, and the word n has not been detected within b frames going back one time, then the word n or the time Suppose that it is detected at i. 1 However, En (i)
is based on the following formula.

Eo (i)=Dn (i, Jn)/J, ... (
5) However, when detecting a certain f-word, the detection conditions do not take into account the influence of the deformation of the spectrum pattern due to the concatenation from the single 5Δ after the city f, which is a characteristic of continuous speech. , the end position often shifts. Errors in detecting such blockages can be divided into three types.

<l> Erroneous detection: A word that does not originally exist in that section is detected.

<2> Not detected...The word existed but could not be detected.

<3> Segmentation error: The original word section and the detected word section are significantly different.

In particular, the example just described corresponds to the segmentation error (3), which causes a fatal error in subsequent processing. This is originally used to detect word sections in continuous speech.
It is necessary to incorporate processing that evaluates the connection state with subsequent word sections and sets the position where the evaluation value is the highest as the start or end. If such processing is incorporated, the segmentation error <3> will be greatly reduced.

Now, FIG. 4 shows an example of word section candidates detected by the conventional detection algorithm. As shown in this figure 4,
In many cases, each of the 4j word interval comparisons A to D overlaps by one on the time axis. That is, the simple i5 interval comparison A-D obtained as a result of continuous DP forms a lathe as shown by the arrow in the figure. From now on, this lattice will be called a word lattice. In order to find the optimal word string from this word lattice, the following algorithm is adopted in the conventional example.

(a) Creation of a word sequence group consisting of exclusive frame sections This process (a) means that if a correctly recognized sequence exists in a word lattice, the section that has a correctly recognized word sequence becomes exclusive. It is based on this idea.

(b) Determine the sequence with the highest likelihood among the above sequences. Next, determine the corresponding E for the word interval candidates included in the above word sequence.
The likelihood of the entire series is determined from the values of , , (i).

Here, if a certain word sequence vJ is composed of IJ word interval candidates, then En (i) of the second interval
When the value of is Fgo and t7, the likelihood LJ of the entire series is determined as follows.

Lj=−□□ Σ α(k)・F7−β・1°k~ (6) Here, α(k) and β are arbitrary values determined in advance by experience.

Therefore, the likelihood LJ of the entire series is the value of each B, that is, the continuous D
The sum of the distances between patterns at P is small, and the likelihood is large, and V
This shows that the larger the number of words included in j, the larger it becomes. Based on this likelihood, the word sequence with the largest J becomes the recognition result.

(Problems to be Solved by the Invention) However, in the conventional technology, a word section is detected independently for each word n from the result of continuous DP, so the connection from the preceding and succeeding words, which is also a characteristic of continuous speech, is not possible. The effect of deformation of the spectral pattern due to Furthermore, the precondition for finding optimal words from a word lattice, that is, ``the intervals of correctly recognized word sequences are mutually exclusive'' cannot be said to be perfect. For example, if "ichi" and -ichi are followed, "chi" at the end of the first "ichi"
The vowel part ``i'' is connected to the first ``i'' of the next ``ichi'', and the boundary between them is not clearly defined. Particularly when continuous DP is used to detect word sections, section candidates often overlap. Segmentation in this part usually requires consideration of the interview relationship or allows for a certain degree of 1[in the word boundaries, which poses the problem that the above-mentioned conditions do not apply.

The purpose of this invention is to obtain an optimal word sequence that also takes into account the context by performing another DP process on the distance calculated from continuous DP without using the above-mentioned preconditions. This paper proposes a continuous speech recognition device that can perform continuous speech recognition.

(Means for Solving the Problems) In order to achieve this objective, the continuous speech recognition device of the present invention has the following features: a) Extracts a sequence of speech feature parameters from human speech at fixed time intervals called frames. b) a standard dictionary unit that stores predetermined standard patterns of recognition target words; and C) performs continuous DP matching processing between the standard patterns and the feature parameter series to determine the features for each frame. Continuation D that calculates the minimum cumulative distance between the parameter series and the standard pattern and the matching start frame position that gives the minimum cumulative distance.
and d) an optimal Ilt that performs processing for determining, for each recognition target word, a word that minimizes the sum of the minimum cumulative distance and the cumulative inter-pattern distance value of the optimal word series up to the matching start frame position. ;? It is characterized by comprising an A-series determination section.

In implementing the present invention, preferably, a) the above-mentioned minimum cumulative distance is calculated using a matching bus normalized by the number of frames on the standard pattern side in continuous DP matching processing between the feature parameter series and the standard pattern, and b) ) It is preferable to configure the above-mentioned optimal word sequence so that the inter-pattern distance cumulative value is expressed by the sum of minimum cumulative distances.

(Operation) As described above, according to the continuous speech recognition device of the present invention, word spotting is performed for each word in continuous single JΔ recognition, and as a result, the cumulative distance between patterns calculated for each frame is further dynamically detected. It is configured to calculate the optimal word string by applying programming processing, so real-time operation is possible, reducing the amount of memory required.
Furthermore, it is possible to improve the recognition rate.

(Embodiments) Hereinafter, embodiments of the continuous speech recognition device of the present invention will be described with reference to the drawings.

FIG. 1 is a functional block diagram showing an embodiment of the present invention, and 100 is a human voice terminal. After the voice manually inputted from the input terminal 100 is A/D converted in a preprocessing unit lot, a characteristic parameter series V(p, i) is calculated for each frame period. This V (ρ, i) is expressed by the voice section detection unit 102
At the same time, it is also sent to the continuous DP section 103. The continuous DP unit 103 calculates the inter-pattern distance from the feature parameter series based on the pattern in the standard pattern dictionary unit 104 in which the standard pattern group of t'thousand is stored.

Furthermore, the optimal word sequence determination unit 105 determines the optimal word sequence using the inter-pattern distance. The optimum word sequence determining unit 105 also receives a detection signal of the end of a voice interval issued from the voice interval detecting unit +02, and outputs the final recognized word sequence to an output terminal +06 as a recognition result.

Each operation of each of these constituent components 101 to 105 can be processed by software using a microcomputer or the like.

The audio signal input manually from the input terminal 100 is sent to the preprocessing unit 10.
1, it is converted into a vector series representing the feature, a feature parameter series V(p, i). This characteristic parameter series V(p, i) is generally obtained by sampling in-band frequency components extracted by P bandpass filter groups having different center frequencies for each frame period. In this case, a prediction result may be calculated for each frame period using another method called linear prediction analysis, sampled, and returned to the feature parameter series V(p, i). This feature parameter series is generated by the continuous DP section 103 and the voice section detection section +02.
It will be sent to Incidentally, since the configuration and operation of this preprocessing section 101 have already been proposed, further detailed explanation will be omitted.

The speech section detection unit +02 uses the feature parameter series V(p,
Based on i), detect the voice section, that is, the start and end of the voice. This detection algorithm is a complex algorithm that performs detection using the voice power found from the feature parameter series V(p, i) (Patent Application No. 108668/1986)
There are simple algorithms and other algorithms that define the point at which the voice power exceeds a predetermined threshold as the start of the voice, and the point at which it falls below the threshold as the end of the voice, but I somehow followed an appropriate algorithm. All you have to do is make it work. Since the configuration and operation of this voice section detection section 102 have already been proposed, further detailed explanation will be omitted.

Next, regarding the operation of the continuous DP section, the details of this block are the same as the configuration and operation according to the algorithm described in the conventional example, except for a certain part. Therefore, in the embodiment of the present invention, only the operation of the conventional example and the older part will be described.

First, in calculating the recurrence formulas of equations (1) and (2) described above, as already explained in the conventional example, initialization processing for the starting end of the input pattern is first performed, or In the embodiment, the frame ■ determined to be the start of the voice by the voice section detection unit 102 is the start of the human pattern.
8 is adopted. That is, i representing the manual frame number used in the conventional example is a relative frame number from Is in this embodiment.
The section up to frame ■6, which is determined to be the end of the audio in step 02, is defined as the length ■ of the human pattern.

Therefore, I=1. −1q+1・・・・・・(7) After the end of the voice is detected, if the start of the next voice is detected, the frame number is set to 5 again and the same process is repeated. Thus, this continuous DP section is provided with means (not shown) for detecting a frame Is instead of i representing the conventional input frame number.

From now on, unless otherwise specified, all manual frame numbers i will be 1. Let it be a relative frame number from . Now, the final result calculated here, that is, the i-frame ends 01;1
Inter-pattern distance fiD with each standard pattern n when (
i, Jo) and matching start frame number Bn (
+, Jn) are directly transferred to the optimal word sequence determination unit 105, which is a characteristic part of the present invention.

Next, the operation of the optimal word sequence determining section 105, which is a main part of the embodiment of the present invention, will be explained according to the flowchart of the operation shown in FIG. In the following description, each processing step will be represented by S.

Here we define the following symbols.

DD(i): Cumulative value of inter-pattern distances of the optimal word sequence up to i frame X (i): Number of words in the optimal word sequence that gives DD (i) The starting frame number of the word in the X(i)th digit that gives the word name BB (i)+DD (i) in the i)th digit.Now, the optimal word sequence that gives DD (IE) is the recognition result. When DD (IE) is used with dynamic programming, it is sufficient to solve the following recurrence formula.

DD(i)=min[DD(m-])+Dn(i, Jo
)] ...(S8) However, m=Bo (i, J
n ) X(i)=X(m)+1−・・・−=・(9)N(i
)-n・・・・・・・・・・・・・・・ (10) BB(i
)=m・・・・・・・・・・・・ (11) Here, n is the word name n that gives the minimum value of equation (8), and m is m=n
This is the value of m when .

By sequentially calculating equations (8) to (11) for 1≦i≦I, DD (1,) can be found.

Therefore, the following processes (1) to (2) are performed.

■First of all, continuous D P ats 103 Karari. (i, J
n) and B. Take in the information of (i, Jn) (Sl).

0 Next, perform initial setting processing for the start of the voice section (
S2).

DD (o)=X (o)=BB (o)=0 ■Next,
For n=1.2, ..., ■, the above equations (8) to (1
1) is executed (34-S7).

In this case, first, (DD (
m-1) +D, (i, Jo)), and then the formula (
8) DD (i) is determined (S4).

Next, let the llj word name 0 that gives this minimum value be n, and the
(i) Set the word name of the digit as m=N(i) (S5
).

Next, let m where m=n be m, and at this time, the Xth (i
) is set as the starting frame number BB(i)-m of the word (S6).

Next, the number of words X(i) in the optimal word sequence that gives DD(i) in equation (9) is determined (S7).

[The above-described processes (S2 to S7) are executed until the final frame I (S8), and then the process proceeds to the next process.

(2) Find the optimal word sequence by backtracing N (i) and BB (i). In other words, let n=1 and S
9), Next, using the starting point BB (1) of the word N(I) that gives the X(1)th digit, write the word name N of the X(1)-1st digit.
(BB (1)-1) and its starting point BB
(BB (1)-1) is determined (S10.5Il).

The result is obtained by performing the same process up to the first digit (S
l2).

Note that the flow of the processing operations described above is an example, and is not limited to this flow.

As is clear from the following, Dn (i, Jn) and Bn (i
, Jn), the optimal word sequence determination unit 105 has determined all the values necessary for the calculation of equation (8)=(+1), so the calculation process can be started at that point. In other words, the processing based on the algorithm of this embodiment is capable of processing a complete frame quiz.

At the time when the end of the voice section is detected, the processing except for the back trace in process 4 is completed, and this is suitable for real-time processing. In addition, the memory required by the optimal word sequence determination unit 105 for holding intermediate results is also
(i), X(i), N(i), and BB(i), it can be realized with an extremely small amount of memory. Furthermore, in practice, if the number of digits
Since the 01 arithmetic operation can be omitted, the structure can be configured with even less memory.

(Effects of the Invention) As is clear from the above description, according to the continuous speech recognition device of the present invention, dynamic programming is further applied using only the results for each frame calculated by continuous DP matching. Techniques for calculating the optimal word sequence by
This is a device that uses a 3D device, which has the advantage of being able to operate in real time and requiring very little memory.

In addition, from the results of continuous DP matching obtained for each frame as in the conventional example, the relationship between the preceding and succeeding eruption sections was considered 1.h.
This method does not employ an algorithm that includes information compression processing at a lower level, in which a word section is directly determined without any additional processing, and then an optimal Qj-word string is determined using only this word section. Therefore, there is an advantage that a high recognition rate can be obtained because errors in lower-level processing seen in such an example do not accumulate.

[Brief explanation of the drawing]

Fig. 1 is a functional block diagram for explaining an embodiment of the continuous speech recognition device of the present invention, Fig. 2 is a flowchart of the operation of the maximum 18Q1 word sequence determination section for explaining the invention, and Fig. 3 is a functional block diagram for explaining the embodiment of the continuous speech recognition device of the present invention. FIG. 4 is an explanatory diagram of word section candidates used to explain the present invention and the conventional continuous speech recognition apparatus. 100... Human voice terminal I, lot... Preprocessing unit 102... Voice section detection unit, 103... Continuous DP unit If) 4... Standard pattern dictionary unit 105... Optimum IP- Various series determination unit 106...output 6;g:f-0 Patent applicant: Oki Electric Industry Co., Ltd.Kono Bf4
Shoes #3? Recognition ■ V death ■ construction map Figure 1 + party association DPI 7) Parameter Figure 3 □ - □ engineering Tsuboma axis Figure 4

Claims (2)

[Claims] (1) a) A preprocessing unit that extracts a sequence of audio feature parameters from input audio at fixed time intervals called frames, and b) A standard dictionary unit that stores predetermined standard patterns of recognition target words. and c) performing continuous DP matching processing between the standard pattern and the feature parameter series, and determining a minimum cumulative distance between the feature parameter series and the standard pattern for each frame, and a matching start frame position that provides the minimum cumulative distance. and d) for each word to be recognized, a word that minimizes the sum of the minimum cumulative distance and the cumulative inter-pattern distance value of the optimal word series up to the matching start frame position. What is claimed is: 1. A continuous speech recognition device, comprising: an optimal word sequence determination unit that performs processing for determining an optimal word sequence. (2) a) the minimum cumulative distance is calculated by a matching pass normalized by the number of frames on the standard pattern side in continuous DP matching processing between the feature parameter series and the standard pattern, and b) the minimum cumulative distance 2. The continuous speech recognition apparatus according to claim 1, wherein the cumulative distance value between patterns of a series is expressed by the sum of the minimum cumulative distances.