JPH0241760B2 - - Google Patents

Description

[Detailed Description of the Invention] This invention calculates the degree of similarity between a time series of features of input speech and a corresponding standard time series based on calculating the distance between elements of these time series. The present invention relates to a speech recognition device.

In a conventional speech recognition device, a time series A=a ₁ , a ₂ . . . a _l of features of input speech and a standard time series B=b ₁ , b ₂ ,
...b _n , the distance measure between the elements a _i and b _j was determined using a spectrum in which the total power in each element was normalized to a constant value. For this reason, there was a drawback that the distance could become quite small even between elements with extremely different vocal powers, such as a _i being a vowel and b _j being a consonant. In speech recognition, it is desirable to be able to recognize the voices of as many people as possible using one standard pattern, but in the case of conventional methods, distances based on spectra that are easily affected by individual differences are used, so vowels and consonants Errors such as erroneous recognition are likely to occur, and the recognition rate is particularly low when the input speech is not from the same speaker as the standard pattern.

Furthermore, as a distance measure between elements, one using a combination of spectral distance and power distance is known. For example, this is the Journal of the Institute of Electronics and Communication Engineers Vol.
J64-A.No.5.P.409-416 (May 25, 1981), especially as shown in the section 2 and 3 WeightedLiklihoodRatio on page 411, the spectral distance and power distance are given the same weight. , and its weight is fixed. Since this distance measure has a fixed power distance weighting, it is affected by variations in voice strength from utterance to utterance, deformation of power patterns due to differences in S/N ratio, voice cutting mistakes, devoicing, etc. If the influence of the weighting factor is too strong, mismatching may occur.

By using a matching distance measure between elements of a speech time series that combines spectral information and power information with little individual difference, this invention can more accurately recognize speech even for unspecified speakers than conventional methods. The drawings will be explained in detail below.

FIG. 1 shows an embodiment of the invention, in which input terminal 1
The audio signal added to the audio signal is digitized in the AD conversion and audio extraction section 2, and then only the audio portion is extracted, and the normalized power and autocorrelation coefficient are calculated at fixed time intervals in the autocorrelation analysis section 3. . As an example of normalized power, the maximum value is 1
The logarithmic power with the minimum value normalized to 0 is used.
The spectrum calculation unit 4 extracts parameters necessary for spectrum calculation, and the spectrum distance calculation unit 6 calculates the spectral distance between elements of the input voice and the standard pattern stored in the standard pattern storage unit 5. In the standard pattern, the logarithmic power is normalized so that the maximum power is 1 and the minimum power is 0 for each word. An example of spectral distance is the WLR scale (Sugiyama, Kano "LPC spectral matching scale with weighted peaks", Journal of the Institute of Electronics and Communication Engineers,
Vol. J64-A, No. 5, PP409-416, 1981 can be used.

Furthermore, the power distance calculating section 7 calculates the power distance between elements of the input voice and the standard pattern.
A distance calculation section 8 calculates an inter-element distance consisting of the spectral distance and power distance obtained in the calculation sections 6 and 7. Examples of this distance include D=WLR+6.0(P$-P ^* ) ² (1) D=WLR・(1+4.0(P$-P ^* ) ² ) ² (2). Here, P$ and P ^* are the normalized powers, P$ is the input voice, and P ^* is the power of each standard pattern. WLR represents WLR spectral matching distance. Since a better effect appears when the WLR scale is low-frequency weighted, the low-frequency weighting is performed in the spectrum calculation section 4 (Kano,
Sugiyama “Large word speech recognition experiment using WLR scale”
Proceedings of the Spring Conference of the Acoustical Society of Japan, 3-1-19, 1981.)

Time-normalized matching is performed in the time-normalized matching unit 9 using the inter-element distance obtained by equation (1) or (2) (for example, "Evaluation of LPC spectral matching scale in large word speech recognition" by Kano) , “Acoustical Society of Japan/Speech Research Group Material S80-60,
1980), the degree of similarity between the input voice obtained as a result and the standard pattern is calculated for several standard patterns for one input voice, and the similarity comparison unit 10 compares them and selects the one with the highest degree of similarity. The object is output to the recognition result output terminal 11.

However, the basic characteristic of the time-series pattern of voice power is that it is high in the vowel part and low in the consonant part.
This property remains unchanged even if the speaker is different. Figure 2 shows the power patterns of the word "Sapporo" uttered twice by each of the four speakers.
In FIG. 2, the logarithmic power pattern is normalized so that the maximum value and minimum value are constant. As can be understood from FIG. 2, the power pattern does not differ much even if the speaker changes. Therefore, it will be understood that the recognition rate can be improved by matching the input speech and the standard speech with respect to the time-series pattern of power as in the present invention, and also taking this into consideration.

Regarding the power time series pattern, when comparing input speech and standard speech, the logarithmic maximum value max
(logp(t)) is a constant value AL, e.g. 1, the logarithmic minimum value
When min(logp(t)) is normalized to a constant value IL, for example 0, the normalized pattern P(t) of the power pattern p(t) of one word is given by the following equation.

P(t)=log(p(t))-IL/AL-IL By normalizing in this way, even if the intensity ratio of speech and background noise changes, the unvoiced part can be made to the same value ( Also, due to the nature of the logarithmic function, the compression ratio is high in the parts with high power intensity, so all vowel parts take values close to 1. The power pattern is less deformed due to fluctuations.

For the 641 city name speech data, the power pattern was normalized as described above, and the average power distance between the same words and the average power distance between different words were calculated using the conventional time normalized matching method. Within participants, the distance between the same words is 0.0374 (maximum is 1, minimum is 0), but the distance between different words is
The distance between the same words between different speakers is 0.0530, which is smaller than the distance between different words within the same speaker, and the distance between different words is 0.147, which is the same as the distance between different words within the same speaker. It has become quite large.
From this, it can be seen that referring also to the power time series pattern according to the present invention can improve recognition in speech recognition from a plurality of speakers.

Therefore, the spectral matching distance Ds (i, j) between input audio frame i and standard pattern frame j
and the power distance Dp(i,j), the distance D(i,j) between these audio frames is given by the following formula, and D(i,j)=F(Ds(i,j),Dp( i, j) ∂D/∂Ds≧0 ∂D/∂Dp≧0, Ds≧0, Dp≧0 Using this distance D(i, j), time normalized matching of audio frame sequences is performed. If this is done, parts with similar spectra and power levels will be matched, and matching between vowels and consonants can be avoided.The simplest form of the function F is D(i,j)=Ds(i , j) + Dp (i, j) D (i, j) = Ds (i, j) · Dp (i, j) In this case, in this invention, the variation in voice intensity for each utterance, In order to avoid effects such as deformation of the power pattern due to the difference in S/N ratio and mistakes in audio extraction, the conventional spectrum matching distance excluding the power component is used as Ds (i, j), and the power pattern is always Taking into consideration that variations are involved, Dp(i, j) has pure sensitivity to the power difference when the power difference between the input voice and the standard pattern is small, and has sharp sensitivity when the difference is larger than a certain level. Let the function be such that
As the simplest of these functions, the square function can be used as shown in equations (1) and (2).

When using the cepstral distance WLR measure as Ds(i,j), on the other hand, the power distance Dp
The optimal weighting value of (i, j) and the matching distance between frames within the same speaker and between different speakers for 100 city name data are: D ² = WLR + a (P$ - P ^* ) ² D ² = WLR・(1+a(P$-P ^* ) ² ) ² , and examine changes in recognition rate and error rate by changing coefficient a for these, and as shown in equations (1) and (2), optimal coefficient a= 0.0 and a=4.0 were obtained, respectively.

Using equations (1) and (2) above, we conducted a large word speech recognition experiment for 641 city names and compared it with the conventional method using only spectral distance. As a result, (1),
There was no substantial difference in the first-place recognition rate between equations (2).
When using equations (1) and (2), compared to the conventional method, the recognition results within the same speaker were approximately the same, but between speakers of different speakers, the recognition was significantly improved from 82.4% to 88.0%. The recognition rate was improved, and it was confirmed that this invention is excellent.

As described above, according to the present invention, since the input voice and the standard pattern are matched using the distance consisting of the power distance and the spectral distance,
For standard patterns with content different from the input speech, the time-series patterns of power are different from each other, resulting in poor similarity. Therefore, there is an advantage that the recognition ability can be improved in multi-speaker speech recognition where recognition errors are likely to occur when using spectral distance alone. Furthermore, in this invention, since the power distance is used as a matching distance between elements that are taken into account by changing the sensitivity according to the power difference, it is possible to prevent variations in voice strength from utterance to utterance, deformation of the power pattern due to differences in S/N ratio, It is not affected by audio cut-out mistakes.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the concept of the speech recognition device according to the present invention, and FIG. 2 is a diagram showing the speech power time series pattern of "Sapporo". 1: Audio input terminal, 2: A/D conversion/sound extraction section, 3: Autocorrelation analysis section, 4: Spectrum calculation section, 5: Standard pattern storage section, 6: Spectral distance calculation section, 7: Power distance calculation section , 8: distance calculation section, 9: time normalization matching section, 10: similarity comparison section, 11: recognition result output section.

Claims (1)

[Claims] 1. Means for determining the spectral distance Ds between the elements of the feature time series of input speech and standard pattern speech, and normalization so that the maximum and minimum values of input speech and standard pattern speech are respectively constant values. A method for determining the power difference between an input voice and a standard pattern using the voice power obtained by performing
The difference between Ds and power is changed by using a function that changes the weighting for the difference between Ds and power, and makes the sensitivity to the power difference dull when the power difference is smaller than a predetermined value, and sharper when it is larger than the predetermined value. 1. A speech recognition device comprising means for determining an inter-element matching distance from the above, and means for calculating the similarity between an input speech and a standard pattern speech by time-normalized matching using the inter-element matching distance.