JPH0466040B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION Technical Field The present invention relates to speech pattern creation in a speech recognition device.

Prior Art When performing speech recognition of words, a matching method using the DP method (dynamic programming) is currently common, but this method has the drawback of requiring a large amount of calculation. Therefore, a method has been proposed in which audio is expressed as a time frequency pattern (TSP) and matched.

FIG. 1 is a block diagram showing an example of the above TSP, in which 1 is a microphone, 2 is a switch, 3a,
3b is an amplifier, 4a and 4b are filter groups, 5
a, 5b are binarization circuits, 6 is a dictionary section, 7 is a similarity comparison section, 8 is a comparison section, and 9 is a result output section. In this method, a pattern obtained by binarizing the word TSP with a certain threshold is registered as a standard pattern, and an unknown input pattern is changed to a binary variable pattern by changing the threshold and superimposed on the registered pattern. The degree of similarity is calculated from the degree of overlap, and by doing this, fluctuations in the frequency direction are absorbed to see the superposition of a wide pattern binarized with a low threshold and a narrow pattern binarized with a high threshold. Pattern matching is possible, and this is a matching method suitable for speaker-independent recognition.

FIG. 2 is a diagram showing one sample of the TSP time of a certain word, with the horizontal axis showing frequency and the vertical axis showing level. In this case, since the vocal cords and sound source characteristics decrease as the frequency increases, the level in the high frequency range of the frequency pattern of the uttered word also decreases. this
When the threshold values of L ₁ and L ₂ are binarized to "1" and "0",
As shown in FIG. 2b and c, even if the pattern is the same, the parts that become "1" are different. Therefore, a method can be considered in which the threshold value has a slope close to the vocal fold characteristics, but according to this method, the 2
Although the difference in the value pattern is reduced, there is still a drawback that a difference appears in the pattern when the peaks of F ₂ and F ₃ are smaller than that of F ₁ as shown in FIG. 3.

Purpose The present invention has been made in view of the above-mentioned circumstances, and in particular, has been made for the purpose of preventing peaks on TSP from being omitted during binarization.

Configuration The configuration of the present invention will be described below based on Examples.

The present invention detects each convex portion or concave portion on a pattern, and determines a binarization threshold by multiplying the peak or dip of the convex portion or concave portion by a constant value or by adding or subtracting a constant value.

FIG. 4 is a configuration diagram for explaining one embodiment of the present invention. In the figure, 10 is a microphone, 11 is a filter group, 12 is a voice section extraction section, and 13 is a convex part detection part (maximum point detection part). ), 14 is a register, 15 is a binarization section, 16 is a threshold value determination section, 17 is a dictionary section, 1
8 is a similarity comparison section, 19 is a result display section, and first,
A word voice inputted through a microphone 10 is converted into a frequency characteristic pathometer by a group of bandpass filters 11, and only the voice-related portion is extracted from the voice. This signal is detected by the protrusion detection section 13 at 1 of the time.
Convex portions are detected sample by sample in the frequency direction. A convex portion can be detected, for example, by looking at the difference between adjacent filter outputs and detecting a portion where the sign is reversed. The peak value of the detected convex portion is sent to the threshold value determination unit 16, where it is determined by a straight line connecting adjacent peaks (a value lower by a certain value than the broken line in FIG. 5).
The threshold value is determined as (solid line in the figure). In addition, when changing the threshold value of binarization depending on the time of dictionary creation and the time of recognition, the above-mentioned constant value may be changed depending on each. After being registered in the dictionary in this manner, the unknown input is binarized, and the matching unit calculates the degree of similarity with each word registered in the dictionary, and outputs the result having the maximum degree of similarity as a recognition result. The degree of similarity is calculated by superimposing a binarized time-frequency pattern of an unknown voice on a similar pattern in a dictionary and determining the degree of overlap. If the overlapping patterns have different time lengths, the lengths may be made equal by linear expansion or contraction, or if the dictionary pattern consists of the addition of binary patterns in which the same word is uttered many times, the lengths may not be made equal. There are also good things. Also, the fourth
It is also possible to replace the detection of convex portions in the figure with detection of concave portions and set a value higher than a certain value as the threshold value, and the same effect can also be obtained in this way.

FIG. 6 is a configuration diagram for explaining another embodiment of the present invention. This embodiment is basically the same as the embodiment shown in FIG. 4, but as shown in the figure, a convex portion detection section 13 and a concave portion detection section (minimum point detection section) 20 are installed after cutting out the voice section. This method determines the convex portions and concave portions of the characteristic pattern, and determines the binarization threshold from the difference between the peak value of the convex portion, the peak value of the concave portion, and the dip value.

FIG. 7 is a configuration diagram for explaining still another embodiment of the present invention, and section A in the figure corresponds to section A in FIG. 4 or FIG. 6. In this embodiment, as shown in the figure, a slope correction section 21 is provided next to the voice section extraction section or before the protrusion/concavity detection section, so as to have an inclination in the frequency direction as shown in FIG. The above operation is performed after subtracting the value and making the peak level almost constant. After correcting the peak level in this manner, the signal may be transmitted to the portion corresponding to portion A in FIGS. 4 and 6. Incidentally, the tilt correction can be achieved by, for example, a least square straight line (Miwa, Kido: Speech Study Group Material S79-24).

Effects As is clear from the above description, according to the present invention, it is possible to create an accurate voice pattern without missing characteristic peaks in the time-frequency pattern of voice.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an example of the TSP matching method, FIGS. 2 and 3 are diagrams each showing one sample of TSP time, and FIG. 4 is a diagram for explaining one embodiment of the present invention. 5 is a waveform diagram for explaining the operation, and FIGS. 6 and 7 are
FIG. 7 is a configuration diagram for explaining other embodiments of the present invention. 10...Microphone, 11...Filter group, 12
...Voice section cutting section, 13...Protrusion detection section (maximum point detection section), 14...Register, 15...
Binarization unit, 16...Threshold value determination unit, 17...Dictionary unit, 18...Similarity matching unit, 19...Result display unit, 20...Concavity detection unit (minimum point detection unit), 21
...Inclination correction section.

Claims (1)

[Scope of Claims] 1. A sound pattern is created by having a band division means for dividing an audio signal into a plurality of frequency bands, and a binarization means for binarizing each band signal divided by the band division means. In the speech pattern creation device to be created, a detection means for detecting the maximum point of the band signal with respect to frequency and a straight line determined by two adjacent points among the maximum points detected by the detection means, and a line on the straight line are determined. 1. A voice pattern creation device, comprising a threshold setting means for setting a threshold value to a value lower than a predetermined value, and binarizing each of the band signals using the threshold set by the threshold setting means. 2. An audio pattern creation device that creates an audio pattern by having a band dividing means that divides an audio signal into a plurality of frequency bands, and a binarizing means that binarizes each band signal divided by the band dividing means. In this step, a straight line is determined by a detection means for detecting the minimum point of the band signal with respect to frequency and two adjacent points among the minimum points detected by the detection means, and a predetermined value is determined from the value on the straight line. 1. A voice pattern creation device, comprising a threshold setting means for setting a high value as a threshold, and binarizing each of the band signals using the threshold set by the threshold setting means.