JPH0449720B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Field of the Invention The present invention relates to a speech recognition device intended for unspecified speakers.

Configuration of conventional example and its problems Conventionally, in a speech recognition device, an n-dimensional feature vector sequence {a ₁ , a ₂ , ...a _I } obtained by analyzing an input speech signal is processed in advance as a dictionary. P standard pattern vector sequences registered in {b ¹ ₁ , b ¹ ₂ , ...b ¹ _J } ...{b ^p ₁ , b ^p ₂ , ...
b ^p _K }, the one with the closest distance or the one with the greatest similarity is used as the recognition result. In this case, the input vector sequence {a ₁ , a ₂ ,
..., a _I } and one of the standard pattern vector series, e.g. {b ^l ₁ , b ^l ₂ , ... b _M } (where l=1~
P), one-element vector ai of {a ₁ , a ₂ , ..., a _I } and one-element vector b ^l _n of {b ^l ₁ , b ^l ₂ , ..., b ^l _M } Most methods calculate the city distance or Euclidean distance, and then calculate the total distance between two vector sequences based on this using techniques such as dynamic programming or linear expansion/contraction.

However, the city distance and Euclidean distance are given by the following formula.

ai = {ai, ₁ , ai ₂ , ..., ai, N} b ^l _n = {b ^l _n,1 , b ^l _{n,2 ,} ..., b ^l _n,N }, then c _l,n = _N 〓 ^r=1 |ai, r−b ^l _n,r r| (urban distance) y _l,n = _N 〓 ^r=1 (ai, r−b ^l _n,r ) ² (Euclidean distance) However, In city distance and Euclidean distance, the speaker who extracted the registered standard pattern and the speaker who is actually trying to recognize are different. In the case of so-called speaker-independent recognition, a sufficient recognition rate was not obtained. This is due to minute variations in the spectral structure of each speaker.

When a section of an audio signal, for example about 10 msec, is cut out and frequency analyzed using means such as Fourier transform or a filter bank, peaks appear in some frequency bands. This is called a formant and is an important parameter that characterizes phoneme. A formant corresponds to the pole of a filter, or resonance point, when the human vocal tract is viewed as a filter with a certain transfer function. Among these, the one with the lowest resonance frequency is called the first formant, the second formant, ... the nth formant,
It is generally known that relatively low-order formants, particularly the first and second formants, play a very important role in characterizing phonemes.

Once the formant frequency and bandwidth are determined,
Although the phoneme can almost be determined, there is also variation depending on the individual, and this is the cause of a decline in the recognition rate in speaker-independent recognition.

For example, if you cut out a part of the voice waveform uttered as |a| (``a'') and perform frequency analysis using a relatively wide band hand-pass filter group that does not contain pitch components, the result will be as shown in Figure 1A. Two peaks are formed around kHz. This corresponds to the first and second formants (F1, F2). Also, the third formant (F3) appears near 3 and H2.

On the other hand, |i| (“I”) has F1300Hz,
F22.5kHz and F33kHz (Figure 1B) However, there are slight differences in the values of F1, F2, F3... depending on the individual. That is, even when the same sound is uttered as |a|, the positions of the formants are slightly different between speaker A and speaker B, as shown in FIG. 1, C and D. This variation in formant positions among speakers has been the cause of a decrease in recognition rates when conventional speech recognition devices are applied to unspecified speakers.

OBJECTS OF THE INVENTION In view of the above drawbacks, it is an object of the present invention to provide a speech recognition device that improves the reduction in recognition rate in speaker-independent recognition due to individual differences in formant frequencies.

Structure of the Invention The present invention includes a frequency analysis means for outputting a sequence of feature vectors, a storage means for storing a standard pattern vector sequence subjected to frequency analysis in advance, and a comparison between the output of the frequency analysis means and each of the standard pattern vector sequences. A speech recognition device is provided with a comparison means for determining a standard pattern vector that provides the minimum distance as a result of the comparison, and a determination means for determining, as a recognition result, a standard pattern vector that provides the minimum distance as a result of the comparison. Divide into pairs of adjacent frequencies, compare each pair while moving them in parallel, find the correspondence that minimizes the distance, and calculate the distance between the two vectors by taking the sum of the distances at that time. This method can reduce individual differences in formant positions and improve the recognition rate in speaker-independent recognition.

DESCRIPTION OF EMBODIMENTS An embodiment of the present invention will be described below with reference to the drawings.

FIG. 2 is a block diagram of a speech recognition device in one embodiment of the present invention. In the figure, reference numeral 1 denotes a parameter analysis unit that analyzes the parameters of the input voice and sequentially converts it into an N-dimensional parameter vector sequence {a ₁ , a ₂ , ..., a _I }. It consists of an analyzer. 2 is a switch which is switched to the B side when creating a standard pattern and to the A side when comparing patterns. 3 is a pattern storage unit, which stores N created by the parameter analysis unit 1.
The array of dimensional parameter vectors is defined as a standard pattern {b ¹ ₁ , b ¹ ₂ , ..., b ¹ _J }, ..., {b ^p ₁ , b ^p ₂ , ...b ^p _K
}
be memorized as

4 is a phase shifter consisting of K phase shifters, which converts one vector b ^l _n belonging to the standard pattern vector sequence into b ^l _n = {b ^l _n,1 , b ^l _n,2 ..., b ^l _{n, N} }, this is K
set of pieces The signal is divided into 2 and each of them is transferred to a corresponding phase shifter and shifted. 5 is a partial distance calculation unit, which consists of K partial distance calculators, and calculates the output of each phase shifter, which is sequentially output while shifting, and one of the input parameter vectors ai=, which is also divided into K sets. The distance is calculated for each set of {ai, 1, a _i,2 , ... a _i,r } ... {a _i,t , a _i,t+1 ... a _i,N }.

Reference numeral 6 denotes a partial determining unit, which is composed of K partial determining units, and selects and outputs the minimum among the calculated phase shifter outputs while sequentially shifting. Reference numeral 7 denotes a total distance calculation unit which calculates the sum of the K values obtained from the partial determination unit 15 and sequentially cumulatively adds the results of performing the above operations for i=1 to I of the input parameter vector sequence. 8 is the judgment section, which performs the above operations,
This is performed for standard pattern vectors l=1 to P, and the one with the minimum distance obtained as a result is output to the signal line 9 as the recognition result.

Next, the operation of the apparatus configured as described above will be explained separately for the time of standard pattern creation and the time of pattern comparison.

First, when creating a standard pattern, the switch 2 is connected to the B side, and the input audio signal is processed by the parameter analysis section 1 into an N-dimensional parameter vector sequence {a ₁ ,
a ₂ , . . . , a _I } are sequentially converted and then stored in the pattern storage unit 3. By repeating this operation nine times, standard pattern vector sequences {b ¹ ₁ , b ¹ ₂ , ..., b ¹ _J }, ... {b ^p ₁ , b ^p ₂ , ...
,
b ^p _K } is stored.

Next, the case of pattern comparison will be explained. For pattern comparison, the switch 2 is connected to the A side, and the parameter analysis section 1 converts the input voice into an input parameter vector sequence {a ₁ , a ₂ , ..., a _I }, and inputs it to the partial distance calculation section 5. . On the other hand, the pattern storage unit 3 stores one of the standard pattern vector sequences {b ^l ₁ , b ^l ₂ , . . .
..., b ^l _M } are divided into K groups and input to K phase shifters of the phase shifter 4. i.e. vector
Let b ^l _n be one vector layered in {b ^l ₁ , b ^l ₂ , ..., b ^l _M }b ^l _n = {b ^l _n,1 , b ^l _n,2 ...b ^l _n,N } , then this This is used as input to the phase shifter 4.

Each phase shifter in the phase shifter 4 shifts its output each time a partial distance is calculated in the partial distance calculator 5 in the next stage, and compares it with each vector in the input parameter vector sequence grouped in the same way. Distance calculations are performed while shifting the pattern between the two.

In other words, the input parameter vector sequence {a ₁ , a ₂ ,
..., a _I }, each element of one vector ai (i=1 to I) is similarly divided into K sets. i.e. Each of these sets is used as one input of each partial distance calculator of the partial distance calculating section 5, and that of the standard pattern vector is used as the other input. At this time, when the distance of the first set is expressed as a city distance, it is expressed as c _1d = _S 〓 ^v=1 | a _i,v −b ^l _n,(v+d) | ...(1). d at this time becomes the shift amount in the phase shifter 4. The distance c _k,v can be defined in the same manner up to the Kth group (k=1 to K).

In the partial determination unit 6, the distances c _k,v sequentially given by the partial distance calculation unit 5 (however, -D1dD2:D1,
D2 is a constant indicating the amount of shift).
c _k,vnio is determined and input to the comprehensive distance calculation section 7.
The total distance calculation section 7 calculates the sum of the K partial distances obtained from the partial determination section 6, and further converts this sum into a standard pattern vector sequence {b ^l ₁ , b ^l ₂ , . . . b ^l _n , . . . b ^l _M } over m=1 to M, and output this to the comprehensive determination unit 8 as the distance cl from the input pattern vector sequence {a ₁ , a ₂ , . . . a _I }. That is, cl= _M 〓 ^m=1 _K 〓 ^k=1 c _k,vnio ...(2) The comprehensive judgment unit 8 performs the above operation for l=1 to P of the standard pattern vector sequence, and then gives the minimum distance. A standard pattern vector sequence is output to the signal line 9 as a recognition result.

Next, the shift operation of the phase shifter 4 and pattern comparison will be further explained using FIG. 3.

FIG. 3A shows one input pattern vector sequence ai uttered by speaker A, where ai={a _i,1 , a _i,2 , . . . a _i,N }.

C in the figure shows the vector b ^l _n corresponding to the input pattern vector in the standard pattern vector sequence, b ^l _n = {b ^l _n,1 , ..., b ^l _n,N }, and these are divided into K pieces. Each divided into blocks
B1 to B4.

At this time, the operation of calculating the partial distance c _1d = _S 〓 ^v=1 ｜a _i,v −b ^l _n,(v+d) ｜(−D1dD2) is standard as shown in Figure 3B. pattern of
This is nothing more than shifting part B1 one sample at a time from left to right and calculating the distance one by one.

As described above, according to this embodiment, the phase shifter 4 divides the standard pattern vector into K equal parts and shifts them sequentially, and the partial distance sequentially calculates the distance between the input pattern divided into K equal parts and the output of the phase shifter 4. A calculation unit 5 and a partial determination unit 6 that determines the minimum distance among its outputs.
By providing this, it is possible to realize a pattern comparison method that corrects shifts in formant positions due to different speakers by comparing each part of the standard pattern while moving it parallel to the input pattern.

The urban distances c _{and d} given by Equation (1) in the main text can be similarly realized using other distance measures such as Euclidean distance or LPC distance.

Further, the calculation of the cumulative distance cl given by equation (2) in the comprehensive distance calculation section can also be performed using linear expansion/contraction and DP matching techniques.

Effects of the Invention As described above, the speech recognition device of the present invention divides the standard pattern and the input pattern into K sets, moves each set individually in parallel, and calculates the sum of the distances when the distance is the minimum. By setting the distance between the two patterns as the distance between the two patterns, it is possible to reduce the distance error caused by individual differences in formants and improve the recognition rate in speaker-independent speech recognition, which has great industrial value. be.

[Brief explanation of the drawing]

Figures 1A to D are waveform diagrams for explaining differences in spectral shapes, Figure 2 is a block diagram of a speech recognition device in an embodiment of the present invention, and Figure 3 explains a pattern comparison method in the same embodiment. FIG. 1...Parameter analysis section, 2...Switch, 3
...Pattern storage unit, 4...Phase shift unit, 5...Partial distance calculation unit, 6...Partial judgment unit, 7...Comprehensive distance calculation unit, 8...Comprehensive judgment unit.

Claims (1)

[Claims] 1. A frequency analysis means that frequency-analyzes an input audio signal and outputs an N-dimensional feature vector sequence {a ₁ , a ₂ ..., a _I }, and P sets of standard pattern vector sequences {b ¹ that have been frequency-analyzed in advance ₁ , b ¹ ₂ , ..., b ¹ _J }...
..., {b ^p ₁ , b ^p ₂ , ..., b ^p _K };
One element vector ai (i=1 to I) of the input feature vector sequence {a ₁ , a ₂ , ... a _I } and the standard pattern vector sequence {b ¹ ₁ , b ¹ ₂ , ..., b ¹ _J ｝……,
One _- element ^vector b _l ⁿ ₍ ^{l =} ₁
~P), the elements of b ^l _n {b ^l _n,1 , b ^l _n,2 ,
..., b ^l _n,N } is a set of adjacent frequencies {b ^l _n,1 ,
b ^l _n,2 ,..., b ^l _n,S }...{b _n , ^l _t , b ^l _n,t+1 ,...b ^l _n
,N }
, move each pair in parallel around the corresponding frequency band of ai divided in the same way, find the correspondence that minimizes the distance, and then create a vector ai with the sum of the distances of each pair. and b ^l _n , and by this measure, the output of the frequency analysis means {a ₁ ,
a ₂ , ..., a _I } and the standard pattern vector sequence {b ¹ ₁ , b ¹ ₂ , ..., b ¹ _J } ..., {b ^p ₁ , b ^p ₂ , ..., b ^p _K
}
a standard pattern vector {b ^l ₁ , b ^l ₂ ,
...b ^l _M } as a recognition result.