JPS60201395A - Voice recognition - Google Patents

Abstract

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

Description

DETAILED DESCRIPTION OF THE INVENTION (Technical Field) The present invention relates to a speech recognition method, and specifically, to a speech recognition method in which speech recognition is started by detecting the beginning of input speech before confirming the end of word input speech. Regarding recognition methods.

(Background technology) As a form of speech recognition, a feature pattern of speech is expressed as a frame time series of frequency components, the degree of dissimilarity between the input pattern and each standard pattern is calculated, and information containing at least that degree of dissimilarity is expressed. There are known methods for recognizing the category of input speech based on.

As methods for calculating dissimilarity, DP matching using dynamic programming and essentially linear linear matching are known for setting a matching path between an input pattern and a standard pattern.

The latter linear matching method is a simple method, and is applied when a relatively small number of recognition categories are targeted, as in, for example, Oki Research and Development No. 118, December 1980.

However, in the conventional linear matching method, the recognition operation starts after confirming the end of the word input speech, which has problems in terms of recognition response speed and requires a relatively large amount of operation memory. shall be.

Furthermore, in the linear matching method, various measures have been taken to absorb variations in speech rate, but these methods are difficult to apply to real-time dissimilarity calculations.

(Object of the Invention) An object of the present invention is to increase the recognition response speed and reduce the operating memory capacity by detecting the beginning of input speech and immediately starting recognition. The purpose is to absorb fluctuations in speaking speed by anticipating speed fluctuations and setting a plurality of essentially linear matching paths.

(Summary of the Invention) In the present invention, for example, although the pattern information of the last frame 1 of the standard pattern may be associated with multiple frames of the input pattern, it is essentially by a linear matching method.

The present invention detects the beginning of the input audio, calculates the distance between the input pattern and each standard pattern each time the input frame is updated, and corresponds to each standard pattern each time the input frame is updated, at least in the voiced state. Then, the degree of dissimilarity is calculated and updated and stored.

In this specification, the order of input patterns at least in the sound state is referred to as the audio frame number. The voice frame number has the same input turn order and tl, but it differs when a method of stopping distance calculation at power dip is adopted.

This dissimilarity calculation is performed by adding the distance at the audio frame number to the dissimilarity to the previous audio frame along the matching path, so after checking the end, the dissimilarity is calculated over the entire length of the matching path. Compared to the process of calculating similarity, the recognition response speed is much faster.

In another aspect of the present invention, dissimilarity calculation is performed in which variations in speaking speed are absorbed by setting a plurality of matching buses. In the present invention, one standard pattern code is selected based on the dissimilarity for at least each audio frame in the voiced state, and in another aspect, the dissimilarity is selected in order to reduce the number of selection processes. A memory with twice the capacity is prepared, and a standard pattern code is selected only for the degree of dissimilarity of the audio frame corresponding to the end of the input audio.

(Embodiment) FIG. 1 is a block diagram of a main part showing an embodiment of the present invention, and the explanation will be given below along with the figure.

52 is a frame counter that counts the number of input frames from the start of the input audio and outputs the audio frame number j, and its contents are set to 0 by a reset pulse when the start of the input audio is detected.
After that, the count is incremented by a count pulse every time the input frame is updated. The start of input audio is detected by exceeding a predetermined threshold, and as a noise countermeasure, the threshold is continuously exceeded for 3 frames (
However, if the length of one frame does not exceed 16 m5ec), recognition processing is performed by resetting the initial state. 60 is a standard pattern memory storing all standard patterns 5n (i, j), which is addressed by a standard pattern code (hereinafter referred to as standard pattern number) n, channel number i, and audio frame number j;
It is read out element by element.

61 is the input cover turn memory for one frame, and the input cover turn W for one frame is stored for each analysis frame of input audio.
(i, j) is input, addressed by channel number i, and read out one element at a time.

Reference numeral 63 denotes a distance calculator, which calculates and accumulates the distance value difference corresponding to each channel number i, thereby calculating the difference between the input cover pattern W(i, j) and each standard pattern 5n(i, j). Calculate the distance dn(j) for one frame. This calculation is performed every time the audio frame number is updated.

64 is an adder; 65 is a dissimilarity memory that stores the dissimilarity corresponding to each standard pattern; Addressing with the pattern number, read the dissimilarity Dn (j-=1) corresponding to the immediately preceding audio frame number,
An adder 64 calculates the dissimilarity Dn(j-1) and the distance d.
n(j) and again provides it to the dissimilarity memory 65, so that the dissimilarity Dn(j) of each standard pattern S'' (1tj) is updated and stored every time the audio frame number j is updated. do.

70 is a comparator, and 71 is a minimum dissimilarity memory. At the beginning of updating the audio frame number j, the maximum value is set in the minimum dissimilarity memory 71, and the dissimilarity Dn(j) of the audio frame counter is set to the maximum value. Each time , the dissimilarity Dn(j) is read out, compared with the value in the minimum dissimilarity memory 21, and the smaller value is stored in the minimum dissimilarity memory 71.
This is to replace 1.

Therefore, when the processing for the audio frame number j ends, the dissimilarity that gives the minimum value along the matching bus from the start of the input audio to the frame number j is detected.

Reference numeral 72 denotes a frame code memory for storing standard pattern numbers; each time the comparator 70 detects a small degree of dissimilarity, it is rewritten with the corresponding standard pattern number, so that the recognition result at the time of the audio frame number j is stored. be done.

Reference numeral 75 is provided as a countermeasure against power dips, and is a code memory for storing a standard pattern number.The content of the frame code memory 72 is transferred and stored using a termination candidate clock indicating that a silent state has occurred. When the end of the code memory 75 is confirmed, the standard pattern number in the code memory 75 is recognized as the input voice.

That is, depending on the recognition target, there is a power dip with low power within the word (for example, between ``to'' and ``b'' of ``st'' and ``b''), so the end can be confirmed when the frame power continues for about 307 frames.

Therefore, it is necessary to continue the recognition operation until the end of the input audio is confirmed, and it is also necessary to save the recognition results at the end candidates whose frame power has fallen below the threshold. I am doing it.

With this configuration, a recognition response can be made immediately after the end is confirmed, and the input cover pattern memory only needs to be for one frame. In addition, if it is desired to avoid making a judgment for each audio frame in the recognition category, a separate memory similar to the dissimilarity memory 65 may be provided, and at the time an end candidate of the input speech is detected, the information at that time will be stored in the separate memory. The degree of dissimilarity is stored, and when the end is confirmed, the degree of dissimilarity is input to the comparator 70 and the minimum degree of dissimilarity memory 7.
1 may be used to detect the standard pattern number. In this case as well, the dissimilarity is calculated by adding the distance at that point to the previous dissimilarity every time the audio frame is updated, so after confirming the end, the recognition response takes about a few m5ec. can do. 2 and 3 are explanatory diagrams of another embodiment of the present invention. In this example, three frame numbers are generated for each standard pattern for each audio frame, and three matching passes are generated. is set, the degree of dissimilarity is calculated and updated for each path each time the audio frame number is updated, and the degree of dissimilarity is updated and stored, and a determination is made based on the degree of dissimilarity of the path corresponding to the path selected based on a certain criterion.

Figure 2 (a) shows the audio frame number j of the input pattern on the horizontal axis and the frame number of the standard pattern on the vertical axis, assuming a case of 20 bits faster, a standard case, and a case of 20 bits slower. Three matching paths 101 to 103 are shown, and FIG. 2(b) shows the frame power corresponding to the audio frame number j of the input pattern.

5L(n) in FIG. 2 is the frame length of the standard pattern with standard pattern number n.

The following formula calculates the frame number generated corresponding to an arbitrary standard pattern at audio frame number j by k, lc
', k' indicate the distance of one frame between the standard pattern and the input cover turn.

Distance to path 101 Distance to path 102 Distance to path 103...In the third formula path 101, the distance between the jth frame of the input pattern and the second frame of the standard pattern is calculated. In pass 102, the distance between the jth frame of the input cover turn and the ``th frame of the standard turn is calculated, and in the pass 103, the distance between the jth frame of the input cover turn and the k''th frame of the standard pattern is calculated. However, if k, k', k", which indicates the frame number of the standard pattern, is larger than the length of the standard pattern, 5L(n), then 5L(n) is calculated.
n). Note that [ ] in the first to third equations is a Gaussian symbol.

Next, the dissimilarity Dn(j) up to audio frame number j
, D'n(j) p denotes D''n(j).

Dissimilarity of path 101 Dn(j) = dn(j)+Dn(j-1) - 4th
Dissimilarity of expression path 102 D'n(j) = d'n(j)+D'n(j-1)
−=Dissimilarity degree of the fifth formula bus 103 D”n(j)=d”n(j)+D”n(j−1) ・・
・Equation 6, that is, calculation of dissimilarity at audio frame number j on each path, is calculated using the distance for each channel (
For example, IW(i#j)-8n(itk)l) for channels and the dissimilarity value (
For example, it can be obtained by adding it to Dn(j-1). These calculations are performed every time the voice frequency number j is updated. There are as many sets of these dissimilarities as there are standard patterns (assumed to be N). Category determination is performed using these dissimilarities. First, the dissimilarity for each path with respect to the nth standard pattern Dn(j), D'n(j)
One of tD"n(j) is selected. For this selection, L, L', and L" given by the following equation are used for the voice frame power j at the time of voice end detection.

These values L, L', L" are similar to the formula giving the number of frames of the standard pattern corresponding to a voice frame, but are not limited by the length of the standard pattern 5L(n). Therefore, L, L', and L'' are unrelated to the type of standard pattern. These.

Select only the dissimilarity corresponding to the path that is closest to the length 5L(n) of the standard pattern among L' and L''.For example, if L' is the closest to 5L(n), path 102 is selected. corresponds to it, and the dissimilarity D'nl) corresponding thereto is selected.The selected dissimilarity is assumed to be DDn.

These selections are made for each standard pattern.

A minimum value is determined for the degree of dissimilarity DDn for each standard pattern obtained in the determination step.

The category added to the standard pattern that gives this minimum value becomes the recognition result for the audio frame number j.

As shown in FIG. 2(b), the audio frame number j.

If the frame power is below the threshold at +1 and a silent state is detected, the standard pattern number determined by the audio frame number j1 is stored as a recognition category candidate, and the recognition operation is interrupted until the silent state continues for 30 frames. do not.

In FIG. 2(b), as a result of continuous recognition, one standard pattern number at audio frame number j2 is updated and stored as a recognition category candidate, and the end is confirmed at audio frame number j3, and that standard pattern number is Determined as input audio.

FIG. 3 is a block diagram of this embodiment, and the following description will be made with reference to this diagram.

52 is an audio frame counter that outputs audio frame number j.

It is a ROM that is addressed by a path signal indicating the type of path and an audio frame number j, and outputs t that can correspond to frame numbers k, k', k'' of the standard pattern.

Reference numeral 56 denotes a ROM which stores the entire standard pattern frame length 5L(n) and outputs it when addressed by the standard pattern number n.

57 is a selector that compares the outputs t and 5L(n) of both ROMs.

Both the ROMs 54 and 56, the comparator 57, and the selector 58 perform the RO
The frame number of the standard pattern stored in M54 is output from the selector 58, and if A>5L(n), the final frame number equal to the frame length 5L(n) is output.

In this way, for each audio frame update, three frame numbers k, k', k are assigned corresponding to each standard pattern.
" is generated, and the matching path explained in FIG. 2 is set.

60 is a standard pattern memory, 61 is an input pattern memory, 62 is an absolute value calculator, 64 is an adder, and 65 is a dissimilarity memory. D, each path 101 to 1 of each standard pattern every time the audio frame number j is updated
Corresponding to 03, the pattern for one frame is read element by element from the standard pattern memory 6o in synchronization with channel number i, and the calculation IW(itj)-8n(itk)I is performed element by element at 62. Then, calculation of the dissimilarity by the adder 64 and the dissimilarity memory 65 is performed element by element in synchronization with the channel number i as shown in the following equation.

Contents of dissimilarity memory ← Contents of dissimilarity memory at the previous channel number i + IW (' * j) - 8n (1
t k) l ・”Equation 8.

In this example, although the process of summing the distance corresponding to channel number i over one frame is not taken, the distance between the input cover turn W (i, j) and the standard pattern 5n (i, k) is calculated for one frame. and the previous audio frame number j
This is substantially the same as calculating the dissimilarity Dn(j) at the frame number j by adding it to the dissimilarity Dn(j-1) of −1, and the dissimilarity memory 65 contains , the dissimilarity of each standard pattern corresponding to each path Dn(j) t D'n(
j) t D'n(j) is updated and stored.

67 is a ROM that outputs a path selection signal P that is addressed by an audio frame number j and a standard pattern number n and specifies the path to be selected in the audio frame number j, and uses the above-mentioned formula 7. As shown, the path type to be selected is stored as path selection information in association with the audio frame number for each standard pattern.

68 is a comparator that compares the path selection signal P and the path signal, and 69 is a converter that outputs either the maximum dissimilarity stored in advance or the dissimilarity read from the dissimilarity memory 65. .

After the calculation of the dissimilarity for each audio frame number j is completed, the address is sequentially read from the dissimilarity memory 65 using the standard pattern number n and the path signal, and the path created in the ROM 67 and the comparator 68 is read out sequentially from the dissimilarity memory 65. The dissimilarity corresponding to the selection signal P is output from the converter 69, and the maximum dissimilarity is output in other path signal periods, thereby selecting a path, that is, one dissimilarity for each standard pattern. .

20 is a comparator, 71 is a minimum dissimilarity memory, 72
is the frame code memory, 25 is the code memory,
By these, one standard pattern number is recognized as a category of input audio, and in this configuration, the maximum dissimilarity is output from the converter 69 corresponding to the non-selected path, so the This is the same as the embodiment shown in FIG.

Table 1 shows the results of a recognition experiment to find out what kind of matching path is effective to set in order to overcome the difference in speaking speed using the relative configuration shown in Figure 3. Based on these results, it was concluded that three matching passes assuming a difference in speaking speed of ±20 degrees are effective.

Table 1 As evaluation voices, the voices of 20 men and 20 women were recorded using a microphone into a tape recorder.

The words used in this recognition experiment were the eight words shown in Table 2.

Table 2 #-v Hisashi 1 of 12! ! Intestinal last name F1 E Zpekal L positive #
The analysis was performed by logarithmically transforming the frequency analysis results and using the frame average power, low frequency average power, and high frequency average power of each frame.

(Effects of the Invention) As explained above, in the present invention, each time the audio frame number is updated, the distance corresponding to one frame is calculated, and the distance is added to the dissimilarity at the previous audio frame number. Since the degree of dissimilarity is calculated every time the frame number is updated, there is an advantage that the recognition result can be obtained immediately after the termination is confirmed, and the capacity of the input cover pattern memory and other operation memories is also relatively small.

Furthermore, since the speaking speed is predicted and a plurality of linear paths are set for each standard pattern, there is an advantage that fluctuations in speaking speed can be absorbed with a relatively simple configuration.

[Brief explanation of the drawing]

Fig. 1 is a block diagram showing one embodiment of the present invention, Figs. 2 and 3 are explanatory diagrams of other embodiments of the present invention, and Fig. 2 is an explanatory diagram of dissimilarity calculation; Figure 3 is a professional drawing. 52...Frame counter; 54 is a ROM that stores the frame number of the standard turn in correspondence with the audio frame number; 56 is a ROM that stores the frame length of the standard turn; s7 is a ROM that stores the frame number of the standard turn; Konno Kureta, 58
is a selector, 60 is a standard number turn memory, 61 is an input cover turn memory, 62 is an absolute value difference calculator, 63 is a distance calculator, 64 is an adder, 65 is a dissimilarity memory, and 67 is each standard number ROM which stores turn path selection information corresponding to audio frame numbers, 68 is a comparator, 69 is a controller, 70 is a comparator, 71 is a minimum dissimilarity memory, 72 is a frame code memory, and 75 is a code memory . Patent applicant Oki Electric Industry Co., Ltd.■, Case indication 1982 Patent Application No. 057280 2, Name of the invention Speech recognition method 3, Relationship with the person making the amendment case Patent application person office (〒 105) 1-7-12 Toranomon, Minato-ku, Tokyo
No. 6, Contents of the amendment (1) On page 19, lines 9 to 10 of the specification, the statement "The process of summing the distances corresponding to channel number i over one frame" has been changed to [channel-related distances]. The process of summing over one frame is corrected as ``.

Claims (2)

[Claims] (1) Store a standard pattern expressed as a frame time series of frequency components corresponding to each standard voice, a) extract the input pattern from the input voice as a frame time series of frequency components, and b) extract the input voice as a frame time series of frequency components. C) an essentially linear relationship to the audio frame number; d) Every time the audio frame number is updated, calculate the distance between the input cover turn and each standard pattern between the frames associated with the matching path, e. ) The cumulative value of the distance along the matching bus from the start of the input audio to an arbitrary audio frame is defined as the dissimilarity, and the immediately preceding dissimilarity and the distance at the audio frame number are added and temporarily stored. f) Separately store all the dissimilarities, or at least update and store the dissimilarities corresponding to each standard pattern; and f) Separately store all the dissimilarities, or at least g) After confirming the end of the input audio, one standard pattern code selected based on the standard pattern code or the dissimilarity that is separately updated and stored is stored. A speech recognition method characterized by recognizing standard pattern codes as categories of input speech. (2) Store a standard pattern expressed as a frame time series of frequency components corresponding to each standard voice, a) extract the input pattern by expressing it as a frame time series of frequency components from the input voice, b) ) Detect the start of the input audio and start counting that frame, and update the audio frame number each time a frame is updated at least while detecting the active state; C) Essentially change the audio frame number A plurality of matching passes are set by generating a plurality of frame numbers of the standard pattern in a linear relationship, and d) each time the audio frame number is updated, the input cover pattern and each standard pattern are e) The cumulative value of the distance along the matching path from the start of the input audio to an arbitrary audio frame is defined as the dissimilarity, and the dissimilarity of the immediately preceding dissimilarity and the audio frame are calculated. By adding the distance in terms of number and storing it once, the degree of dissimilarity is updated and stored corresponding to each matching pass of each standard pattern. g) After confirming the end of the input audio, store the dissimilarity separately or at least update and memorize one standard pattern code selected based on the dissimilarity; A speech recognition method characterized in that one standard turn code selected based on the degree of dissimilarity updated and stored in the input speech is recognized as a category of input speech.