JPH0570840B2 - - Google Patents

Abstract

PURPOSE:To prevent lowering of recognition rate even when non-stationary noise is added by masking the part to which non-stationary noise is added, and not using the distance between patterns of the part to which noise is added. CONSTITUTION:Characteristic of inputted voice is extracted by a characteristic extracting section 1, and after normalization, stored in an input pattern memory 3. A noise detecting section 2 watches characteristic vector, and when the change is abrupt and in a short section, regards as addition of noise, and makes a mask series zero and stores in the input pattern memory section 3. Mask series that indicates pattern of each category and noise section are stored in a reference pattern memory section 4. Mask series read out from the input pattern memory section 3 and reference pattern memory section 4 are inputted in a distance mask section 6, output of a distance calculating section 5 is selected by the output and outputted.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a device for recognizing patterns of speech, characters, etc., and particularly to a pattern recognition device that eliminates the influence of noise and has high noise resistance.

(Prior Art and its Problems) Various methods have been tried in the past as means for recognizing patterns of sounds, characters, etc. Among them, the pattern matching method is a relatively simple and effective method. In the pattern matching method, the recognition target is expressed as a series of feature vectors and stored as a standard pattern.The pattern to be recognized (called the input pattern) is compared with the standard pattern that is closest, that is, the standard pattern with the smallest distance is selected. The category to which the standard pattern belongs is output as a recognition result.

On the other hand, when the recognition device is actually used, noise is often added to the pattern, causing a reduction in the recognition rate. A method for recognition by removing noise is described in Proceedings of the Acoustical Society of Japan, 3-4-6, ``An Experiment on Noise Processing in Word Recognition'' (hereinafter referred to as Document 1), published in March 1982. This method is called the noise subtraction method, and its outline is as shown below.

Now, when the recognition target is speech, the pattern can be expressed as a time series of spectra. The spectrum X(ω) observed in a noisy environment is the audio spectrum S(ω)
Since it can be considered that a noise spectrum N(ω) is added to the
The speech spectrum S^(ω) is estimated by subtracting the noise spectrum N^(ω) from (ω).

S^ (ω) = X (ω) − N^ (ω) ...(1) However, in the noise subtraction method, the noise is assumed to be stationary within a certain time of the voice and a constant value is subtracted, so Although this method is effective against regular noise, it is not effective against non-stationary noise.

As described above, in the conventional method, when non-stationary noise is added, the noise cannot be estimated and the noise cannot be removed. For this reason, even if the patterns belong to the same category, the distance between them becomes large, resulting in erroneous recognition and a low recognition rate.

(Objective of the Invention) An object of the present invention is to solve the above-mentioned drawbacks by masking the part to which unsteady noise is added and not using the distance between the patterns in the part to which the noise is added. It is possible to provide a pattern recognition device in which the recognition rate does not decrease even when noise is added.

(Structure of the Invention) According to the present invention, a feature extraction unit that obtains a feature pattern of an input pattern, an input pattern memory that temporarily stores the output of the feature extraction unit, a standard pattern memory that stores a standard pattern, and a noise added thereto. a distance calculation unit that calculates the distance between the input pattern and the standard pattern; a distance mask unit that sets the distance of the noise interval detected by the noise detection unit to zero; and the distance mask. There is obtained a pattern recognition device characterized by having a distance accumulator that accumulates outputs of the distance accumulator, and a determination unit that determines a recognition result from the output of the distance accumulator.

(Principle of the present invention) Next, the principle of the present invention will be explained. Input pattern A and standard pattern B ⁿ of the nth category are A = a ₁ , a ₂ , ..., a _I ... (2) B ⁿ = b ₁ ⁿ , b ₂ ⁿ , ..., b _I ⁿ ... (3) Then, the distance between the input pattern A and the standard pattern B ⁿ is D (A, B ⁿ )=1/I _I 〓 ⁱ⁼¹ d (a _i , b _i ⁿ ) ……(4) Given. Here, d (a _i , b _i ⁿ ) is the distance between the a _i vector and the b _i ⁿ vector, for example, d (a _i , b _i ⁿ )= _k 〓 ^R=1 | a _ik − b ⁿ _ik | ...(5) However, it is conceivable to use the Thievishev distance where a _i = (a _i1 , ..., a _ik , ..., a _ik ). Recognition by pattern matching searches for the standard pattern B ⁿ that is closest to the input pattern A.
That is, the recognition result is obtained as n^=argmnin D(A, B ⁿ ) (6). Here, argmxin y indicates the x when y is the minimum for all x.

By the way, as shown in FIG. 2, the pattern to which unsteady noise is added is more deformed than the original pattern in the section to which the noise is added. For this reason, the distance in the section where noise is added becomes abnormally large,
Distance D from the standard pattern of the correct answer, which is originally a small value
(A, B ⁿ ) became large, leading to misrecognition.

In the present invention, a noise-added section is detected based on a change in a feature vector, etc., and the distance is determined and recognized only in the section without noise, without using the distance related to the feature vector of the section. In other words, the mask sequence M=m ₁ , m ₂ , ..., m _i , ..., m _I ...(7) where m _i = 1or0 shall be.

is obtained for each pattern, and the distance between input pattern A and standard pattern B ⁿ is Dr (A, B ⁿ ) = 1/I _rI 〓 ⁱ⁼¹ m ^a _i・m ^bn _i・d (a _i , b ⁿ _i )
...(8) However, I _r = _I 〓 ^I=1 m ^a _i・m ^bn _i ...(9) Calculated. The recognition result is n^=argmin Dr(A,B ⁿ )...(10). Dr (A, B ⁿ ) obtained in this way
Since the influence of noise is removed, a recognition device with good noise resistance can be realized.

(Example 1) Next, an example of the present invention will be described using the drawings.
As shown in FIG. 1, the first embodiment includes a feature extraction section 1, a noise detection section 2, an input pattern memory 3, a standard pattern memory 4, a distance calculation section 5, a distance mask section 6,
It is composed of a distance accumulating section 7 and a determining section 8.

The first input voice is extracted by a feature extractor 1, such as a spectrum parameter, and converted into a feature vector, normalized on the time axis, and stored in the input pattern memory 3. Here, normalization of the time axis means expanding and contracting the length of each pattern in the time direction so that it is constant, and there are two ways to expand and contract, such as a linear method and a method that makes changes in the feature vector uniform. . Such a feature extraction unit 1
"Speech Recognition" published by Kyoritsu Shuppansha in October 1954.
It is possible to use the preprocessing unit, feature extraction unit, and time normalization unit described in FIGS. 4 and 7 on page 113 of (hereinafter referred to as Document 2).

The noise detection section 2 monitors the feature vector, and when the change is rapid and in a short period, it is assumed that noise has been added, and the mask sequence m _i is set to zero and stored in the input pattern memory section 3. The noise detection section 2 can be constructed from a well-known microprocessor, and is constructed as shown in FIG. 3, and operates according to the flowchart shown in FIG. 4. In other words, the change in the feature vector Δi = |a _i −a _i-1 | is monitored, and when Δi _s > TH and Δi _e > TH and i _e −i _s < TL ...〓 m _is , ... m _ie =0. Here, let TH be the threshold value for the amount of change, and let TL be the time interval.

The standard pattern memory 4 contains patterns B ⁿ of each category after similar feature extraction and noise detection.
and a mask sequence m ^bn indicating the noise section.

The distance calculation unit 5 is a part that calculates the distance between feature vectors using equation (5), and as shown in FIG. Read the a _i vector from 3 and the b _i ⁿ vector from the standard pattern memory, and write d(a _i , b _i ⁿ ) to the accumulator 54.
is required.

The distance mask section 6 is composed of a selection circuit 61 and an OR circuit 62, as shown in FIG. OR circuit 6
The mask sequences m _i ^a and m _i ^bn read out from the input pattern memory unit 3 and standard pattern memory unit 4 are input to 2, and the output d of the distance calculation unit 5 is
Select and output either (a _i , b _i ⁿ ) or 0. i.e. OR
When the output of the circuit 62 is 1, d(a _i , b _i ⁿ ) is output, and when the output of the OR circuit 62 is 0, 0 is output. In this way, in the distance mask section 6, m _i
This means that ^a・mi _bn ^・d(a _i , b _i ⁿ ) is being calculated.

The distance accumulator 7 includes an adder 71 as shown in FIG.
and the distance m _i ^a , m _i ^bn・d by the accumulator 72
(a _i , b ⁿ _i ) is calculated, the adder 73 and the accumulator 74 calculate the addition length Ir= _I 〓 ⁱ⁼¹ m _i ^a・m _i ^bn , and the divider 75 calculates the distance Dr(A , B ⁿ ) are calculated.

The determination unit 8 sequentially outputs the category n of the standard pattern to the standard pattern memory unit, and determines the distance from the input pattern.
Receive Dr(A, B ⁿ ), calculate its minimum value,
Output the category n^ for which the minimum value was obtained.

(Example 2) Next, a second example of the present invention will be described using the drawings. In the first embodiment, the feature extractor performs time normalization so that the pattern length (number of feature vectors) is constant. However, since voices and the like individually have subtle expansion and contraction, a DP matching method is known that calculates the distance between patterns while finding the expansion and contraction that best matches each standard pattern. This details
Nikkei Electronics issue published November 7, 1983
“Two-stage DP matching for efficient recognition of continuously uttered word sounds” from pages 171 to 208 (Reference 3)
). The outline is as follows.

A function (time normalization function) that maps the time axis j of the standard pattern B ⁿ to the time axis i of the input pattern is defined as i=j
When (i), time-normalized distance between patterns
D _d (A, B ⁿ ) is the solution of the minimization problem where D _d (A, B ⁿ )=1/Imin _I 〓 ⁱ⁼¹ d(i, j) ……(13) j=j(i) obtained as. This minimization problem is obtained by sequentially calculating the following recurrence formula.

Initial condition: g(0,0)=0, g(i,j)=d(i,j)+ming(i-1,j) g(i-2,j) g(i-3,j)... (14) Calculate the recurrence formula in the range j-rij+r, and g (I, J _o ) at the terminal I of the input pattern and the terminal J ⁿ of the standard pattern become the sought D _d (A, B ⁿ ). .

The second embodiment is an apparatus incorporating time normalization by DP matching, and is configured as shown in FIG. The second embodiment has the same parts as the first embodiment except for the feature extracting section 1 and the cumulative distance section 7. The feature extractor 1 is composed of the feature extractor 1 of the first embodiment except for the time normalizer, and obtains a feature vector of a pattern. The cumulative distance unit 7 is composed of a recurrence formula calculation unit 721, a DP matching control unit 722, and a G memory unit 723, and the distance obtained from the distance mask unit 6 from which noise has been removed is
The time-normalized distance D _d ( _A , B ⁿ ₎ is determined by the DP matching method using m ^a _i ·m ^bn _j ·d (a i , b ^{n j} ). This recurrence formula calculation section 721, DP matching control section 7
22, G memory unit 723 can use the functions described in FIG. 11 of Reference 3, respectively.

Although the present invention has been described above based on examples, these descriptions do not limit the scope of the rights of the present invention. In particular, in this example, the distance between feature vectors was used as the Tievishiev distance shown in equation (5).
Euclidean distance d(a _i , b _i ⁿ )= _K 〓 ^K=1 (a _ik −b ⁿ _ik ) ² ...(15) or inner product d(a _i , b _i ⁿ )=− _K 〓 ^K=1 a _ik・b ⁿ _ik ...(16) can be used. Furthermore, various methods described in Document 3 can be used as the DP matching method. These obvious modifications are within the scope of the present invention.

(Effects of the Invention) As explained above, in the pattern recognition apparatus of the present invention, the distance of the section to which unsteady noise is added is assumed to be zero, and the distance is recognized by the cumulative value of only the distance of the section without noise. This removes the influence of noise and prevents a reduction in recognition rate due to addition of noise.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing a first embodiment of the present invention, FIG. 2 is a diagram for explaining the principle of the present invention,
Figures 3, 5, 6 and 7 are respectively
FIG. 4 is a block diagram showing a part of the detailed configuration of the embodiment shown in the figure, FIG. 4 is a flowchart of a noise detection section, and FIG. 8 is a block diagram showing a second embodiment. In the figure, 1 is a feature extraction section, 2 is a noise detection section, 3 is an input pattern memory section, 4 is a standard pattern memory section, 5 is a distance calculation section, 6 is a distance mask section, and 7
is a cumulative distance section, and 8 is a determination section.

Claims (1)

[Claims] 1. A feature extractor that obtains a feature pattern of an input pattern, an input pattern memory that temporarily stores the output of the feature extractor, a standard pattern memory that stores a standard pattern, and a noise that detects a section where noise is added to the input pattern. a detection section, a distance calculation section that calculates the distance between the input pattern and the standard pattern, a distance mask section that sets the distance of the noise section detected by the noise detection section to zero, and a distance that accumulates the output of the distance mask section. A pattern recognition device comprising: an accumulating unit; and a determining unit that determines a recognition result from the output of the distance accumulating unit.