JPS617896A - Word voice recognition method - 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 Field of the Invention The present invention relates to a word speech recognition method for recognizing words by comparing input speech with a word dictionary in which phonemes are expressed.

Structure of the conventional example and its problems FIG. 1 is a functional block diagram of a device for executing the conventional word speech recognition method. In FIG. 1, 1 is a parameter extraction unit that creates a time series of parameters from input speech, 2 is a probability density calculation unit that collates phoneme standard patterns and calculates the probability density of phonemes, 3 is a segmentation unit for each phoneme, Likelihood calculation.

This is a word recognition unit that calculates word similarity. In addition, 4 shows the distribution of various parameters for each phoneme, the average value for each phoneme (μ2), and the covariance matrix (Σ
, ), and a word dictionary section 5 stores a word dictionary in which all words to be recognized are expressed as symbol strings in phoneme units.

For example, the words "Asahi" and "Tochi" are [
ASAHIJ.

It is written as "TOCI" etc.

Next, the operation of the above conventional example will be explained. In the parameter extraction unit 1, input audio is analyzed for each 10 msec frame, parameters are extracted, and a parameter time series is created. Next, the probability density calculation unit 2 calculates the probability density of the parameters and phonemes obtained for each frame. Next, the word recognition unit 3 performs phoneme segmentation for each dictionary item according to the dictionary phoneme series that constitutes that dictionary item, and according to the following formula 0, segments are segmented according to the type of phoneme and the phoneme. The likelihood l is calculated, and the degree of similarity is obtained as the average of the likelihoods of each phoneme in the dictionary entry. Here, if the phoneme is X, the frame numbers at the start and end of the segmented section corresponding to X are Ns and Ne, and the value of each parameter in the second frame is Cn, then the likelihood of phoneme X is lx is defined by the following formula.

φL(Cn) represents the probability density of a certain phoneme B, and is defined as in equation 0.

Σ, 71(Cn-μ,)]......■Cn: ath
N parameters (vector) in a frame μ: Average value (vector) of the parameters of a certain phoneme B ΣO: Covariance matrix of the parameters of a certain phoneme B The range varies depending on what the phoneme X is; for example, when X is the phoneme A (7), the range of & is 5 vowels, A,
They are E, I, O, and U. The word similarity LM obtained above is calculated for each dictionary item according to formula 0, L,
Dictionary entries with a maximum of 4 were selected as recognized words.

1・=1zo7j,/NP' ,,,,,,,
,,,, ■LM: Similarity lj of the Mth word in the dictionary: Likelihood NP of the third phoneme in the dictionary phoneme series:
The dictionary number prime number diagram is ``Land J (/1oci/)'' and the phoneme symbols (1), (0 ), (S), (
I) probability density value φ1. φ. ,φ9. It shows temporal changes in φ1 and audio power P.

Here, the phoneme symbol in the standard pattern corresponding to the phoneme /C/ is (S) representing a fricative group. In Figure 2,
Segmentation and likelihood calculation of the phoneme /C/ assuming the dictionary word /TOCI/ are performed first using the phonetic power ■)
The threshold value is set in advance, and the lower part (6-7
) is detected as a silent part immediately before the rupture, and then the rear end 8 of the phoneme /C/ is detected using the probability density value φS of the phoneme symbol (S) in the standard pattern corresponding to the phoneme /C/, section(
6-8) is the segmentation interval of the phoneme /C/. Here, if a silent part cannot be detected, segmentation of the phoneme /C/ cannot be performed and likelihood calculation is not performed. 7 Next, the likelihood of the phoneme /C/ is (S) in the interval (7-8)
Using the probability density of , calculate the likelihood according to formula (2).

Figure 3 shows the phoneme symbols (1), (0 ), (S)
, (I) probability density value, φ7. φ0. φ3. It shows temporal changes in φ1 and audio power P. Here, the phoneme symbol in the standard pattern corresponding to the phoneme /S/ is (S) representing a fricative group, which is the same as the phoneme symbol in the standard /C pattern corresponding to the phoneme /C/.

In Fig. 3, the segmentation and likelihood calculation of the phoneme /S/ assuming the dictionary word /TO8I/ are performed by first calculating the probability density of the phoneme symbol (S) in the standard tang corresponding to the phoneme /S/. The end/end of the phoneme /S/ is detected using the value, and the interval (9-10) is set as the segmentation interval of the phoneme /8/. Next, using the probability density value of (S) in the interval (9-10), likelihood calculation is performed according to equation (2). Third
In the figure, when the word /Nos i/ in the dictionary is assumed for the input speech Haos i/, the segmentation and likelihood division of the phoneme /C/ are first performed using a preset threshold value of the voice power. However, in the original phoneme /S/ interval, the voice power has a shallow dip shape as seen in the interval (9i, 0) in Figure 3. Therefore, if the audio power is lower than the threshold, that section (11-12)
is regarded as the silent part immediately before the rupture of the phoneme /C/. Now that the silent part has been detected, the rear end 10 is detected using the probability density value of the phoneme symbol (S) in the standard pattern corresponding to the phoneme /C/, and the interval (9-1,0) is converted into a phoneme. /C/ is set as the segmentation interval, and likelihood calculation is performed using the probability density value of (S) for the interval (1, 2-10) according to formula (2). Therefore, for the input phoneme /1osi/, the word /1o in the dictionary
Assuming ci/, the likelihood of the obtained phoneme /C/ is comparable to the likelihood of the original phoneme /S/, so /
It becomes difficult to distinguish S/ and /C/, /S/, /C/
The remaining words had the disadvantage that they were easily misrecognized.

OBJECTS OF THE INVENTION The present invention eliminates the drawbacks of the conventional example described above, and aims to improve the accuracy of likelihood division, thereby improving the word recognition rate.

Structure of the Invention In order to achieve the above object, the present invention calculates the likelihood of a plosive sound with respect to a segmented section (consisting of a silent section immediately before a plosive sound and a section that ruptures or ruptures and causes friction). , instead of calculating the likelihood using only the probability density value of the phoneme in the plosive interval, we also use parameters such as voice power to calculate the likelihood for the segmentation interval of the plosive, including the silent interval. This has the effect of improving the accuracy of likelihood calculation.

DESCRIPTION OF THE EMBODIMENTS The configuration of an embodiment of the present invention will be described below with reference to the drawings. In FIG. 1, the phoneme standard pattern is the same as in the conventional example. A word dictionary represents the words to be recognized as a string of phoneme symbols. Further, the parameter time series obtained by parameter extraction is the same as in the conventional example.

Next, the operation of this embodiment will be explained. First, parameters for each frame are obtained from the input audio, and then the probability density obtained from each phoneme standard pattern is calculated using the value of the parameter, and for each dictionary item, the phoneme The segmentation is performed on the phoneme X and the segmented interval I! corresponding to the phoneme X! Calculate 8. Here, when calculating the likelihood of a plosive, in the segmented interval, we not only calculate the likelihood obtained from the probability density value of the phoneme in the part that ruptures or ruptures and rubs, but also calculate the likelihood for the silent part immediately before the plosive. 7
However, the likelihood is determined using the value of the voice power and its temporal change, and the likelihood of the plosive is determined from the obtained likelihood of each part.

In Figure 2, if we assume that the input speech /1oci/ is the original word /TOCI/, then the plosive /C/
As in the conventional example, the segmentation is performed by detecting the silent part (6-7) just before the rupture and the part (7-8) that ruptures and rubs, and converts the interval (6-8) into the segmentation interval of the plosive sound /C/. shall be. Next, the likelihood of the phoneme /C/ is 7! To calculate o, first, the minimum value (b) of the voice/lower P in the silent part (6-7), and the adjacent 7 frames near the rear end 7 of the silent part shown in equation 0. cepstral distance CD
Using the value of , calculate the likelihood of the silent part according to the formula 0. Next, the likelihood 1cl of the part (7-8”) that ruptures and rubs is found using the probability density value, and finally, from JQ and lcl, the phoneme /
The likelihood 1c of C/ is determined according to formula 0.

l! Q:1QIIll! Q2 1c = j'cl-1゜...Song...
・■In Figure 3, for input voice /1osi/,
Assuming the word /1oci/ in the dictionary, the likelihood calculation for the phoneme /C/ first calculates the silent part (11
-12>, but in this case,
The minimum value e→ of the voice power P of the silent part is sharper than that in the case of FIG. It is smaller than the case shown in the figure. Therefore, the likelihood value of the phoneme /C/ is also smaller, so when the input speech /1osi/ is assumed to be the word /1osi/ in the dictionary, the likelihood of the phoneme /S/ is lower than that of the phoneme /C/. The likelihood also increases, /S/
The ability to discriminate between and /C/ is improved. Therefore, the phoneme /87.
This has the advantage of improving the recognition rate for words containing /C/.
゛Effects of the Invention The present invention has the above-mentioned configuration, and when calculating the likelihood of a plosive, it calculates the standard of the phoneme or phoneme group in the part of the segmented interval where there is a plosive or friction accompanied by a plosive. The likelihood of a plosive is calculated by combining the above two types of likelihood: the likelihood obtained from the distance between the pattern and its phoneme, and the likelihood obtained from the value of the speech power in the silent part immediately before the plosive and the magnitude of its temporal change. By performing the degree calculation, there is an advantage that the likelihood can be determined more accurately than in the conventional example.

[Brief explanation of drawings]

, Fig. 1 is a block diagram of a device implementing the word speech recognition method in the conventional method and an embodiment of the present invention, and Fig. 2 shows the probability density of each phoneme when it occurs in the land. Figure 3 is a diagram showing the temporal changes in voice power and cepstral distance between adjacent frames. Figure 3 is a diagram showing temporal changes in the probability density of each phoneme, voice power, and cepstral distance between adjacent frames when uttered ``city''. 7 used when calculating the likelihood of silent parts? , and j'Q2. 1... Parameter extraction section, 2... Probability density calculation section,
3... Word recognition section, 4... Phoneme standard pattern section, 5.
...Word dictionary department. Name of agent: Patent attorney Toshio Nakao (1st person)
Illustrated η chorus-related sheep measure seam No. 3171 Fig. 4

Claims (3)

[Claims] (1) Word recognition of input speech is performed using a word dictionary that describes the word to be recognized as a symbol string for each phoneme and a standard pattern for each phoneme or phoneme group expressed by the acoustic parameters of each phoneme or phoneme group. To do this, the input speech is checked against each dictionary entry in the word dictionary, and the input speech is segmented for each phoneme according to the dictionary phoneme sequence that constitutes each dictionary entry.
For this segmented speech interval, use the standard pattern of that phoneme or phoneme group and the distance between that phoneme and find the likelihood of the phoneme in the dictionary entry and the input voice, and use this likelihood value. In calculating the likelihood of plosives including affricates, when recognizing words by calculating the similarity between dictionary entries and input speech, we calculate the likelihood of plosives including affricates. The likelihood obtained from the distance between the standard pattern of the phoneme or group of phonemes and the phoneme in the part where the sound occurs, and the likelihood obtained from the value of the speech power in the silent part immediately before the rupture and the magnitude of its temporal change are used together. A word speech recognition method featuring: (2) As a standard pattern for each phoneme or phoneme group, use a standard pattern expressed by the distribution of acoustic parameters of each phoneme or phoneme group, and as a distance measure between the standard pattern of a phoneme or phoneme group and that phoneme, 2. The word speech recognition method according to claim 1, which uses a probability density that a segmented speech section is generated from its phonemes. (3) The word speech recognition method according to claim 1, in which the cepstral distance between adjacent frames is used as the magnitude of the temporal change in speech power in the silent part immediately before bursting.