JPS6332200B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a method for compressing audio patterns.

With a registration type speech recognition device, the speech to be recognized is registered in the recognition device in advance, and during recognition,
The input voice is recognized by comparing a plurality of registered voice patterns with the input voice pattern and finding the registered voice pattern with the most similarity. First, such a speech recognition device will be explained.

The audio signal is converted into an audio pattern as shown in FIG. 1A. That is, in Figure 1, 1
00 is a microphone, 101 is an amplifier, 102, 10
3,104 is a filter bank. The filter banks 102, 103, and 104 each have a first
As shown in Figure B, a bandpass filter 120;
It is composed of a rectifier circuit 121 that rectifies the output, and a low-pass filter 122. The output of the band pass filter 120 is an alternating current component, and the rectifier circuit 121 and the low pass filter 122 convert the output to a direct current level. filter bank 10
The center frequencies of filters 2, 103, and 104 are different, and although there are three filters in the figure, there are actually eight or more filters. The following explanation will be given by taking as an example a case where eight filters are provided.

105 is a control signal S from the acquisition control circuit 112
Outputs ₁ to ₈ from filters 102 to 104
This is a selection circuit that selects one of the following. 106
is an analog-to-digital converter. Using the signal from the oscillation circuit 111, the selection circuit 105 switches the levels of the signals ₁ to ₈ at regular intervals, converts them into digital signals via the A/D converter 106, and captures them into the capture area 107. Therefore, the audio signal is frequency-analyzed by eight filters and is captured into the capture area 107 at regular intervals. If no voice is being uttered, the sum of levels ₁ to ₈ will be close to zero, and if utterance is being made, the sum will be large. Therefore, if a certain threshold is set, and the total sum becomes larger than this level, the acquisition will start, and if the total sum continues to be small for a certain period of time, the frequency analysis of the uttered word will be obtained by stopping the acquisition at that point. , will be captured in the capture area 107. The acquisition control unit 112 performs this control.

108 is a normalization unit, which normalizes the sum of outputs from each filter taken in at the same time to be constant. In other words, a ₁ = ( ₁ / ₈ 〓 ⁱ⁼¹ i) × K, a ₂ = ( ₂ / ₈ 〓 ^{i = 1} i) × K ... a ₈ = ( ₈ / ₈ 〓 ^{i = 1} i) × K ... ...(1) where K is a constant value and ₁ , ₂ ... ₈ are the outputs from each filter. By doing this, the sum of the outputs from each filter becomes K, which is completely unrelated to the magnitude of the audio signal. Since the number of filters is 8, the audio signal at a certain point is 8.
It can be shown as a dimensional vector. If the nth captured audio signal is a _o , a _o = (a _o1 , a _o2 ,
A _o3 ,…, a _o8 ), and the audio signal ends up being a ₁ , a ₂ ,
It can be expressed as a vector sequence of a ₃ , ...a _o ...a _N. Therefore, N indicates the number of audio signals to be captured. This is called a voice pattern.

At the time of registration, the voice pattern is stored in the registration area 110 along the route e in FIG.

During recognition, the input speech pattern is output to the distance calculation unit 109 along the path γ, and pattern matching is performed between each speech pattern within the registered area to calculate the distance, and the standard pattern that is closest in distance is determined. Output as the recognition result of input speech.

At this time, since the same word has different temporal lengths due to differences in speaking speed, some kind of normalization on the time axis is necessary. As a normalization method, a method is known in which dynamic programming is used to nonlinearly expand or contract the time axis.

First, registered patterns a ₁ , a ₂ , a ₃ , ...a _o , ...a _N
and the input patterns b ₁ , b ₂ , b ₃ , ...b _n , ...b _M are arranged on the x-axis and the y-axis as shown in Fig. 2, and on the path _i- axis, Find the distance d (m, n) between a _o and b _n , and use the weighted average of d (m, n) found with respect to the load defined on that path to find the path between the registered pattern and the input pattern.
The distance regarding _i is defined as D( _i ).

Generally, the distance d(m, n) between a _o and b _n is
It is often defined by something called distance between city limits.

That is, in this case, d(m, n)=‖a _o −b _n ‖= ₈ 〓 ⁱ⁼¹ |a _oi −b _ni |.

Here, a _o = (a _o1 , a _o2 , ..., a _o8 ) b _n = (b _n1 , b _n2 , ..., b _n8 ).

At this time, as a route for comparing the two patterns, various routes can be considered as shown in FIG. In the case of FIG. 2, three routes are shown as an example, but various other routes are possible. Among these routes, one method for efficiently finding route ₀ with the smallest distance D( _i ) is pattern matching using dynamic programming. The distance related to route ₀ at this time becomes the distance between both patterns (description of dynamic programming will be omitted since it is separate from the present invention).

By performing pattern matching as described above, the registered pattern with the least similarity is selected, and the input pattern is identified.

In the case of such a method, the capacity of the registration area 110 and the capacity of the input pattern are as follows.
In other words, if the capture interval is Tt seconds, the maximum length of input audio is Ti seconds, and the number of filters is F, then (Ti/Tt) x F samples is the memory required to store the input pattern. The capacity is
The capacity of the registration area is this value multiplied by the number K of speech words to be recognized. In addition, the processing time required for pattern matching is approximately the time required to calculate the distance d (m, n) between two vectors, where Td is the time required to calculate the dynamic programming recurrence formula for one grid point. If the required time is TD, it will be (Td + TD) x (/Tt) x K x W. W
is a value called a matching window, and W is a value of 2 or more.

is the average of Ti.

The present invention aims to reduce the capacity of the registration area and the processing time required for pattern matching through pattern compression.

The present invention calculates the distance between two adjacent vectors in a normalized pattern, and when the distance is large, the two vectors are left as is, and when the distance is small, the distance between the two vectors is determined. On the other hand, it attempts to compress the pattern by using the average as a new vector.
An explanation of this is shown in FIG. When the voice patterns are a ₁ , a ₂ , a ₃ , ...a _o , ...a _N , (P ₀
Indicated by ) Find the vector distance between a ₁ and a ₂ , and when the distance d ₁ , ₂ is greater than a certain threshold S, a ₁ ,
When both a ₂ are smaller than the residual S, leave either a ₁ or a ₂ . In the case of Figure 3,
d ₁ . ₂ is greater than S, leaving a ₁ and a ₂ . then a ₃
Find the vector distance d ₃ . ₄ between and a ₄ , and in the case of figure
d ₃ . ₄ is less than S, leaving a ₃ . The same process is performed below. The compressed pattern sequence is denoted by _P1 . If all d _o,o+1 are less than S, the pattern will be 1/2 the length. all
When d _o,o+1 is smaller than S, it is extremely rare, so in reality, the length is 1/2 or more and 1 or less.
In general, in the vowel part, d _o,o+1 is smaller than S and the speech pattern almost always includes the vowel part, so it is unlikely that it will not be compressed. In the consonant part, the changes are drastic, and d _o,o+1 takes a large value, resulting in
The consonant part remains intact and the vowel part is compressed.

If the above operation is repeated several times, it will be further compressed. In the pattern row of P ₁ in Figure 3,
Further similar processing is performed to obtain a compressed pattern of _P2 . Further compress the pattern sequence of P ₂ ,
Get _P3 . Do the same below. By doing this, the compression ratio will further increase.
As a result, the distances between adjacent vectors are all larger than S, and unnecessary vectors are deleted.

However, in reality, the length of the vowel part, the length of the consonant part, etc. are also necessary amounts of information, and in consideration of this, care should be taken not to overdo the compression.
Depending on the word to be recognized, the number of compressions is determined through experiments based on the recognition results. For example, if you try to distinguish between "KEY" and "KE", if you compress too much, the "I" of "KEY" will
The ``key'' part is missing, making it difficult to distinguish between ``key'' and ``ki''. If the words to be recognized include words whose vowel parts are compressed in this way, making it difficult to identify them, the number of compressions should be reduced.

The threshold value S, like the number of compressions, is also determined through experiments.

FIG. 4 shows an example of a specific circuit configuration of the compression device of the present invention. FIG. 5 is a flowchart explaining the operation. 400 is a shift register,
This includes the audio patterns a ₁ , a ₂ ... to be compressed.
...a Suppose that _N is entered as shown in the figure. E is a code indicating the end. Similarly, 401 is a shift register in which a compressed pattern is stored.
402 and 403 are shift clocks CLK ₁ and CLK ₂
These clocks are generated by signals from the control circuit 407 and the compression determination circuit 406, respectively, and the shift registers 400, 401
are shifted independently. A distance calculation unit 404 calculates a distance d from the contents of the _first stage R ₂ and R of the shift register 400. 405 is a circuit that detects the end code E. 406 is a compression determination circuit that detects whether the distance d is larger or smaller than the threshold value S.

The operation will be explained below with reference to FIG.

The end code detection unit 405 performs operations 450 and 451. If not an exit code, R ₁ ,
The distance d is calculated by the distance calculation unit 404 based on the contents of _R2 , and the compression judgment circuit 404 determines whether it is larger or smaller than the threshold value S.
06 is determined. If d<S, the clock generation circuits 402 and 403 each generate a clock once using a signal from 406 (as shown in 452).The contents of _R1 are transferred to 401, and 400 is incremented by one.
Shift to the right. Next, the control circuit 407 shifts 400 to the right once more by the operation 453 so that the distance between the next adjacent vectors can be calculated. For dS, 454,4
As shown in 55, clocks CLK ₁ and CLK ₂ are generated twice, and the contents of R ₁ and R ₂ are transferred to shift register register 4.
01 and shift register 400 to 2.
It operates so that the next distance calculation between adjacent vectors can be performed by shifting the vector twice. If the content of _R1 is the end code E, only the end code is transferred to the shift register 401 (as shown in 456). If R ₂ is an end code, the contents of R ₁ and R ₂ are transferred to shift register 4.
01 (shown at 457, 458).

Compression can be achieved through the operations described above.

If the contents of the shift register 401 are transferred to the shift register 400 and the operations described above are performed, further compression can be achieved. Alternatively, the circuit configuration may be such that the shift resists 401 and 400 are replaced.

The control circuit 407 receives a signal from the compression determination circuit 406 and issues a command for the operation 453 to be carried out by the circuit 402, and also sends a command to the end code detection circuit 4.
05, the compression operation is terminated, and the end of the compression is notified to other circuits by the END signal. It also receives the STAR signal and issues commands to each circuit to start operating.

When the method of the present invention is used in a speech recognition device, the compression function described above is performed after the output of the normalization unit 108 in FIG. At the time of registration, the compressed pattern is registered as is, the input pattern is similarly compressed, and pattern matching is performed with the pattern group in the registration area.

Note that when the distance between two adjacent vectors is small, one of them is left in Figures 3 and 4, but it is also possible to take the average of the two vectors and combine them into one. . Furthermore, when S is relatively large, taking the average of the two vectors is more faithful to the original pattern. The averaging method is defined as follows.

a′ _oi = a _oi + a _o+1,i /2 i = filter channel number a′ _oi = [a _o1 , a _o2 …a _oi , …a _o,I ] a′ _oi is the averaged vector .

As described above, according to the method of the present invention, the capacity of the registration area can be reduced.

Furthermore, the processing time for pattern matching can be shortened. In other words, the aforementioned (Ti/Tt)
As a result, the number of times the distance d(m, n) and recurrence formula are calculated becomes smaller. at the time of compression
Since d _o,o+1 is calculated, it increases by that amount, but
On the other hand, due to compression, d during pattern matching
The number of calculations for (m, n) is reduced, and the overall processing time is shortened. The greater the number of recognized words, the greater the compression effect.

Furthermore, patterns in which only the consonant parts differ can be more accurately discriminated. For example, SAGA
In the case of KAGA, only the first S and K differ, and the distance between the two patterns is extremely large. When compressed, the vowel part is compressed,
The proportion of S in SAGA increases. KAGA
However, as the proportion occupied by K increases, the distance between the two patterns becomes smaller, and it becomes possible to more accurately identify these similar sounds.

In this example, the degree of difference between two feature vectors was explained using the concept of distance, so when the value is large, the difference between the two is large, and when the value is small, the difference between the two is small. Of course, it is also possible to use the concept of similarity, in which case it would represent the degree to which the two are similar; when the value is large, the difference between the two is small, and when the value is small, the difference between the two is expressed. is large. The essence of the principle of the present invention is the same no matter which concept is used.

[Brief explanation of drawings]

FIG. 1A is a block diagram showing the configuration of a speech recognition device, FIG. 1B is a block diagram of a part thereof, and FIG. 2 is a diagram explaining pattern matching in speech recognition.
FIG. 3 is a diagram illustrating an example of the compression method according to the present invention, FIG. 4 is a block diagram illustrating an example of the configuration of the apparatus according to the present invention, and FIG. 5 is a flowchart illustrating its operation. 400, 401...shift register, 402,
403...Shift clock generation circuit, 404...
Distance calculation unit, 405... End code detection unit, 40
6... Compression determination unit, 407... Control unit.

Claims (1)

[Claims] 1. A first storage means for holding a pattern represented by a series of feature vectors; a distance calculation means for calculating a distance between adjacent vectors without allowing overlap;
compression determining means for determining whether the distance calculated by the distance calculating means exceeds a predetermined threshold; and if the distance does not exceed the threshold, either one or a second storage means for replacing the average value of both with the two adjacent vectors, and storing the two adjacent vectors as they are when the distance exceeds the threshold; A pattern compression device characterized in that the number of feature vectors constituting the original pattern is reduced by transferring the contents of the means to the first storage means and repeating similar processing using the various means.