JPS6131880B2 - - Google Patents

Description

[Detailed description of the invention]

The present invention relates to a speech recognition device with a high recognition rate. A conventional speech recognition device is shown in FIG. In Fig. 1, 1 is a microphone, 2 is a filter analysis section,
3 is a power detection unit, 4 is a sample data storage memory, 5 is a voice section extraction unit, 6 is a feature extraction unit, 7
8 is a voice feature registration memory, and 8 is an identification section. In general, recognition devices are broadly classified into specific speaker and non-specific speaker depending on the speaker. In a specific speaker recognition device, the speaker utters the target word once or several times to register the characteristics of his/her own voice in advance. (Hereinafter, this will be referred to as registration mode) In the case of unspecified speakers, there is no registration process. Most of the products currently available are for specific speakers, which will be explained below with reference to FIG. Input audio is converted into an electrical signal by a microphone 1 and divided into frequency components by a filter analyzer 2. The filter analysis section 2 generally consists of a group of bandpass filters, a group of full-wave rectifiers, a group of low-pass filters, a multiplexer, an AD converter, etc., and it converts the audio band of about 200Hz to 5KHz into about 10
It is divided into ~15 filter groups and the output of each filter is extracted at a cycle of 10 to 20 ms (hereinafter, this output is referred to as sample data). This process is not shown because it is a common method and is not a direct element of the invention. Note that the sample data has positive and negative polarity, or the absolute value of one polarity (positive and negative polarity).
Although there are other expressions such as peak-to-peak (peak-to-peak) data, from now on, for convenience of explanation, we will use absolute value expression. The sample data is sequentially sent to the power detection section 3, and when the sum or maximum value of the sample data output from each filter exceeds a predetermined threshold value, it is considered as the beginning of a voice section and is sequentially stored in the sample data storage memory 4. . Once the data for a certain period of time has been stored, this sequence ends and the voice section cutting section 5 operates next. The voice section extraction unit 5 newly detects the start and end of the voice section. In this method, thresholds 1 and 2 are set using audio power in the same way as the sample data storage method described above. A conceivable method is to set the end point at the beginning of sample data that lasts for a certain period of time or a point in time one sample before that point, and set the period in between as a voice section. Once the voice section is determined, the feature extractor 6 divides the voice section into equal parts, calculates the average value for each filter output within the divided time, uses this as a feature, and stores this feature in the registration memory 7 in the registration mode.
Store in. Once each word has been registered, it becomes possible to identify newly uttered words. The operation of the identification section 8 will be explained below. Let the utterance feature be B ^w (u, v) and the feature of the uttered phrase to be identified be A(u, v). However, w is n
The characteristic of the word/phrase registered No. 20, v is a number sequentially assigned to the divisions within the voice section, and u is a number sequentially assigned corresponding to each filter output. ^Bw
The distance M ^w between (u, v) and A(u, v) is defined as follows. M ^w indicates the degree of dissimilarity; M ^w is determined for all registered words and phrases, and the word corresponding to w for which M ^w is the minimum is the identification result. The process of performing this identification is hereinafter referred to as identification mode. However, with the above technology, even when the same word is uttered by the same person, the utterance time expands and contracts, and even partially expands and contracts each time the same person utters the same phrase.There are cases where linear matching, which divides the vocal interval equally, cannot cope with this problem. It has the disadvantage that it is very difficult to identify many words and phrases that are similar. The purpose of the present invention is to eliminate these drawbacks, by multiplying data for vowel sections with large expansion and contraction, silent sections, etc., dividing the vocal section into equal parts, finding the average feature, and performing the ordinary matching in the first stage. This method is characterized by a two-stage partial matching judgment that can handle words and phrases that are similar to each other, and will be described in detail below. FIG. 2 shows a first embodiment of the present invention, and the cutout portion is omitted because it is not a direct object of the present invention. 10 is a resampling circuit, 20 is a resampling data storage section, 30 is a feature calculator, 40 is a feature storage section, 50 is a registered feature storage section, 60 is a first matching section, and 70 is a second matching section. In order to operate this, the resampling circuit 10 sequentially refers to the cut out voice section data (not shown) from the beginning to detect stationarity (corresponding to the vowel part), and detects the stationarity of each filter output value within one sample. Detects the maximum value of (detects silent parts) and normalizes the vocal power. FIG. 3 shows a detailed block diagram of the resampling circuit 10. 100 is a 1 sample data storage section, 101 is a maximum value detection section, 102 is a maximum value register, 103 is a comparator, 104 is an addition circuit, 105 is an addition register, 106 is a normalization section, 107 is a normalized data storage section, 108 is a difference polarity calculation unit, 109 is a current polarity register, 110 is a previous polarity register, 111 is a coincidence detection unit, 20 is a re-sampled data storage unit, 11
2 is a coincidence calculation section. The 1-sample data storage unit 100 sequentially stores 1-sample data within the cut-out voice section from the starting end. 1
When the sample data is stored, the maximum value detection unit 10
1 and addition circuit 104 sequentially examines these data and stores the maximum value and addition value in maximum value register 102 and addition register 105, respectively. The comparator 103 compares the output value of the maximum value register 102 with a predetermined constant value, and outputs the comparison result to the normalization section 106. Normalization unit 106
If the maximum value is larger than the constant value, the ratio (%) is calculated using the data in the 1-sample data storage section 100 and the value in the addition register 105.
Further, if the maximum value is smaller than the constant value, the ratio calculation is not performed and "0" is output. The output value is stored in the normalized data storage section 107. That is, the output value of each filter stored in the 1-sample data storage section 100 is assumed to be Fn(k). n is the number assigned to the filter, and k is the sample number of the voice section data. Maximum value MAX (k) = MAX {F ₁ (k), F ₂ (k), ......
Fn(k), ......, F _l (k)}. However, the number of filters is l. (1) In the case of MAX(k) constant value, the output NORMn(k) of the normalization unit 106 is NORMn(k)=Fn(k)/ADD(k)×100% (2) MAX(k)<constant value In the case of NORMn(k)=0 Next, the resampling operation will be explained. This detects stationarity in the time series of sample data (generally, it is well known that in speech data, vowel parts show stationarity, while consonant parts and transient parts show non-stationarity), and samples of stationary parts are detected. This is an action to make the surface coarser. The difference polarity calculation unit 108 refers to the data in the normalized data storage unit 107, calculates the difference between the output values of adjacent filters, and determines the difference polarity as three values. However, in the case of the above MAX(k) constant value, that is, for all NORMn(k), NORMn(k)
When =0 is not established, the following operation is performed. Difference value Dn(k)=NORMn(k)−NORMn ₊₁ (k) However, n= ₁ , ₂ , ......, l _-1 . (1) When |Dn(k)|△d (where △d is a predetermined constant) If Dn(k)0, the difference polarity
Sn(k)=S ⁺ If Dn(k)<0, Sn(k)=S ^- (2) If |Dn(k)|<△d, Sn(k)=S ⁰ . Here, S ⁺ , S ⁰ and S ^- are 2-bit representations, for example
Expressed as S ⁺ = (0, 1), S ⁰ = (0, 0), and S ^- = (1, 0). In this way, the code sequence of the difference of one sample data (slope identification code sequence) S ₁ (k), S ₂
Find (k),......S _l-1 (k). However, the above MAX(k)
<In the case of a constant value, that is, when all NORMn(k)=0, the difference polarity is not calculated, and a code sequence that does not appear in the difference polarity calculation is output. For example, S
^× , S ^× , S ^× , etc. However, S ^x = (1, 1). The difference polarity calculation result is set in the current polarity register 109, and at the same time, the contents of the current polarity register 109 up to that point are set in the previous polarity register 110. The initial state of the pre-polarity register (the state before it starts processing one voice) is the code sequence that does not appear in the difference polarity calculation, such as S ^× , S ^× , S ^× , etc. mentioned in the previous example. Assume that it has been set. The match detection unit 111 detects whether the contents of the current polarity register 109 and the contents of the previous polarity register 110 match. That is,
The coincidence detection unit 111 detects whether or not gradient identification code sequences (gradient identification code groups) that are adjacent on the time axis completely match each other. If they do not match, it is regarded as an unsteady point, and the 1-sample normalized data in the normalized data storage section 107 is stored in the re-sample data storage section 20. If they match, the match counting unit 112 counts the number of consecutive matches, and only when a certain count value (a predetermined number of times) is reached, the contents of the normalized data storage unit 107 are stored in the resampled data storage unit 20. At the same time, the count value is set to "0". All normalized resampling operations are performed on the voice section data in this manner. When the resampling and normalization operations are completed for the voice section sample data, the feature calculation unit 30 divides the voice section of the resampled data into equal parts and calculates the channel filter output value (normalized data) in each division. Find the average value and characterize it. Letting I be the voice section length of the resampled data and J be the number of equal divisions, the number (i) of data in one division can be found from I/J=i. In this case, if a remainder (r) is generated, one data is added to each division from the first division until there is no remainder left. For example, if r=3, the number of data in the first three divisions is i+1, and thereafter it is i. The formula for determining the average value is the average value Mj(n), and the normalized value is the NORMn(k). Let j be a number assigned to each division, n be a number assigned corresponding to a channel filter, and k be a number assigned to resampled data. However, j=1, 2,...J, △i=k ₂ −k ₁ +
1 and r=0 in division j, then △i=i, r
If ≠0, Δi=i+1. This process is shown in a flowchart in FIG. 4 and in a block diagram in FIG. In FIG. 5, 120 is a division unit calculation unit;
21 is a re-sample data reference address control unit; 1
22 is an adder, 123 is an addition result storage register,
124 is an average value calculation unit, 20 is a resampled data storage unit, and 40 is a feature storage unit. In the registration mode, the stored features are sent to the registration feature storage section 50 and stored therein. The saved features are hereinafter referred to as registered features. In the identification mode, these features (hereinafter referred to as input features) are sequentially compared with the registered features, and the words and phrases corresponding to the registered features with a small degree of dissimilarity become the identification results. The operation in the identification mode will be explained below. Input features are A(u, v), registration features are B ^w
Let it be (u, v). u is a number assigned corresponding to the filter output, v is a number assigned corresponding to division, and w is a number assigned corresponding to the registered word. Generally speaking, uttered sounds expand and contract in a non-linear manner, so simply matching corresponding divided parts may not match well. In the conventional method, a certain division part v
The dissimilarity n ^w (v _C ) of _C is I was calculating. In the present invention, in each combination (4 x 4 = 16 ways), newly calculated features (hereinafter referred to as overlapping features) for the divided part v _C are used, taking into account the overlap with adjacent divided parts. Let the minimum dissimilarity be the dissimilarity m ^w (v _C ) of the dividing point v _C . That is, m ^w (v _c )=min { _(C , _C) , _(C , _LC) , _(C ,
_RC) , _(C , _LCR) , _(LC , _C) , _(LC , _LC) ,
_(LC , _RC) , _(LC , _LCR) , _(RC , _C) , _(RC , _L
C) , _(RC , _RC) , _(RC , _LCR) , _(LCR , _C) ,
_(LCR , _LC) , _(LCR , _RC) , _(LCR , _LCR) } However, ₍ 〓, 〓 ₎ is a non-standardization of a certain partition v _C using the features F〓 and F〓 shown in Fig. 6 and Table 1. Indicates similarity. for example Therefore, the overall dissimilarity M ^w between the input features and the features of the registered word w is

[Formula]. Note that in FIG. 6, v _C , v _L , and v _R are numbers given to the divisions of the voice section divided into n equal parts.

【table】

[Table] Calculate the degree of dissimilarity between all registered features and the input features, and if there is one registered feature that is less than a predetermined threshold, the corresponding phrase of that registered feature is output as the recognition result. . If the number is 0, it is assumed that there is no target word and a result indicating that it is unrecognizable (reject) is output. If there are two or more matching units, the next second matching unit 70 operates. FIG. 7 shows a detailed block diagram of the first matching section 60. 40 is a feature storage unit, 50 is a registered feature storage unit, 130 is an input feature overlap feature calculation unit, 13
1 is a registered feature overlap feature calculation unit, 132 is an absolute value calculation unit, 133, 137 is an addition unit, 134, 13
6, 138 are registers, 135, 139 are comparison units, 140 is a first matching result storage unit,
The first comparing unit 135 calculates the minimum dissimilarity of the dividing points, and the adding unit 137 calculates the total dissimilarity of all dividing points, and compares it with a predetermined threshold value.
9, and if it is less than the threshold value, the division number w is stored in the storage unit 140. A control unit (not shown) outputs the results of all registered words using the above-mentioned judgment conditions. The second matching unit 70 calculates the degree of dissimilarity between the registered features corresponding to the candidate number w stored in the first matching result storage unit 140 in pairs for each division unit (m ^w (v
(equivalent to _C )), sequentially select p division numbers and formats of overlapping features at that time (constant determined in advance) from those with the highest degree of dissimilarity between each registered feature and input feature. Only for these candidates, the degree of dissimilarity is calculated in the form of the selected overlapping feature, and candidates with a large degree of dissimilarity are deleted from the first matching result storage unit 140. All the candidates are examined in pairs in this way, and if there is a candidate that finally remains in the storage section 140, that number becomes the recognition result, otherwise the result is that it is unrecognizable. For example, if there are three candidate numbers α, β, and γ, there are three pairs each of (α, β), (α, γ), and (β, γ), so first, let's compare the registered features corresponding to α and β. The degree of dissimilarity corresponding to m ^w (v _C ) is checked for each division part, and for the division part with a large degree of dissimilarity, input features and registered features α and input features and registered features β are calculated using the same overlapping feature format. The degree of dissimilarity is calculated between the candidates, and the candidate with the larger value is dropped. (α,
Similar processing will be performed for γ) and (β, γ). FIG. 8 shows a detailed block diagram of the second matching section. 40 is a feature storage unit, 50 is a registered feature storage unit, 150 is a dissimilarity calculation unit for each division unit in FIG. 7, 151 is a result storage unit, 152 is a second matching unit selection unit, 153 is an addition unit, 154,155
is a register, and 156 is a comparison section. The degree of dissimilarity between the registered features is calculated for each division part, and a predetermined number of division parts with a large degree of dissimilarity are calculated by the second matching selection part 152, and the results are transferred to the control part. . A control unit (not shown) extracts input features and registered features according to the information, calculates them in a dissimilarity calculation 150 (lower part of FIG. 8), and stores the sum of one of the pairs in a register 155. It is stored and compared with the other sum in the comparison section 156, and the result is transferred to the control section. As explained above, in the first embodiment, by resampling the constant part, the data of the vowel part and the data of the consonant part become involved in recognition to the same extent (generally, the duration of the vowel part is more is sufficiently long compared to the duration of the consonant part), it is possible to extract well-balanced features. In addition, the vowel part has a large temporal expansion and contraction, but this effect can be suppressed to some extent by resampling, and by using overlapping features and normalizing the resampled data, the magnitude of vocal power can be suppressed. I can deal with differences. There is an advantage that similar phrases can be recognized by focusing only on parts where the difference in registered features is large (parts that are not similar). In the first embodiment, the overlapping features are calculated for the input features and the registered features, but even if it is performed only for one side and the other side does not overlap, it can sufficiently cope with the temporal expansion and contraction of speech. Further, as the overlapping feature, the overlapping feature F _LCR between the divided portion of interest and the divided portions on both sides thereof is sufficiently effective even if not used. For the convenience of explanation, sample data is taken in once and then re-sampled and normalized, but it is also possible to process each sample data at the same time when inputting it.
It is clear that the redundant features may be calculated first and stored as registered features without being calculated at the same time.
Further, although the second matching section performs matching using a predetermined number of division parts in the first embodiment, the total dissimilarity may be calculated only for division parts having a large degree of dissimilarity. Furthermore, a method in which the second matching section is also subjected to multi-stage processing and the number of division sections is gradually reduced is also effective, although it takes longer processing time. Since the present invention includes a resampling circuit, a normalization circuit, and a partial matching circuit, it is possible to perform sufficiently high recognition and can be used in a speech recognition device.

[Brief explanation of the drawing]

Fig. 1 is a block diagram of a conventional speech recognition device, Fig. 2 is a block diagram of an embodiment of the present invention, Fig. 3 is a detailed block diagram of the resampling circuit, and Fig. 4 is a detailed flowchart of the feature calculation section. , FIG. 5 is a block diagram thereof, FIG. 6 is an explanatory diagram of overlapping features, FIG. 7 is a block diagram of the first matching section, and FIG. 8 is a block diagram of the second matching section. 1...Microphone, 2...Filter analysis section,
3... Power detection unit, 4... Sample data storage memory, 5... Voice section extraction unit, 6... Feature extraction unit, 7... Voice feature registration memory, 8... Identification unit,
DESCRIPTION OF SYMBOLS 10...Resampling circuit, 20...Resampling data storage part, 30... Feature calculation part, 40... Feature storage part, 50... Registered feature storage part, 60... First
Matching section, 70...Second matching section, 10
0...1 sample data storage section, 101...
MAX value detection section, 102...MAX value register, 1
03...Comparator, 104...Addition circuit, 105...
...Addition register, 106...Normalization unit, 107...
...Normalized data storage unit, 108...Difference polarity calculation unit, 109...Current polarity register, 110...Previous polarity register, 111...Coincidence detection unit, 112...
Match counting section, 120... Division unit calculation section, 121
... Re-sample data reference address control unit, 12
2... Addition unit, 123... Addition result storage register, 124... Average value calculation unit, 130, 131...
...Overlapping feature calculation unit, 132...Absolute value calculation unit, 1
33,137...addition section, 135,139...comparison section, 134,136,138...register, 1
40...first matching result storage section, 150...
Division dissimilarity calculation unit, 151...result storage unit,
152...Second Matching Division Selection Division, 153...
Addition unit, 154, 155...Register, 156...
...Comparison section.

Claims (1)

[Claims] 1 Sampling that divides the input audio signal into multiple frequency components and samples them at regular time intervals and outputs them as primary audio data.
filter means; silence detection means for comparing the maximum value of the primary audio data at each sampling time point with a constant value to identify a sound time point and a silent time point;
normalizing means for normalizing the primary audio data at a time when there is a sound by audio power; and identifying a difference value between adjacent primary audio data in the frequency axis based on the output of the normalizing means, and the identification value thereof. a slope code creation means for creating a slope identification code group according to the above, and a steady state detection means for determining whether or not adjacent slope identification code groups on the time axis completely match each other to distinguish between a steady time point and an unsteady time point. Then, the primary audio data at a certain number of continuous silent points are represented by the data at one sampling point, and the primary audio data at non-stationary points are made to correspond to those at each sampling point, and the primary audio data at a certain number of continuous points at a steady point are represented by those at one sampling point. The primary audio data of
resampling means for outputting secondary audio data, the audio section is equally divided into a specific number, and for each divided section, the average value of the secondary audio data belonging to the divided section is calculated as a type 1 feature. Then, the average value of the secondary audio data belonging to the divided interval and the immediately preceding divided interval is calculated as the second type feature, and the average value of the secondary audio data belonging to the divided interval and the immediately preceding divided interval is A speech recognition device characterized in that the average value of next speech data is calculated as a third type feature, and the degree of dissimilarity between an input speech phrase and a registered speech phrase is calculated by selectively using these three kinds of features.