JPS6310196A - Voice recognition equipment - 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] mutually The present invention relates to a speech recognition device.

Recently, as a method for easily recognizing word sounds, patterns are binarized, superimposed, and registered as standard patterns, and unknown input sounds are similarly binarized to find the similarity. A method has been proposed.

FIG. 8 shows the binarized T S P (Binary Tim
This is a configuration diagram for explaining an example of the e-5pectru+++Pattern, in which 21 is a microphone;
22 is a filter bank, 23 is a correction unit using least squares,
24 is a binarization unit, 25 is a BTSP creation unit, 26 is an addition unit for n-time occurrence patterns by linear expansion and contraction, 27 is a dictionary unit, 2
8 is a peak pattern creation section, 29 is a pattern length matching section using linear expansion/contraction, 30 is a similarity calculation section, and 31 is a result display section, which displays the input data obtained by binarizing the speech generated for each word. It is recognized by linearly matching cover turns and dictionary patterns, and as shown in the figure, in the case of speech recognition for unspecified speakers, the dictionary membership function is a combination of BTSPs without using TSP frequency characteristics. I'm trying to make a new one as a match. (Please refer to the details of BTSP and, if necessary, the Ricoh Technica
lReportNal 1. MAY, 1984. P,
P4-12; Proceedings of the Acoustical Society of Japan, 1980 1
This method is resistant to pattern fluctuations in the frequency direction, that is, differences between people, and is suitable for speaker-independent methods.
Since the absorption of time fluctuation is based on linear expansion and contraction, D
It is inferior to P matching.

Figure 9 shows the overlapping patterns of normal BTSP, and
Figure 10 shows an example in which time fluctuations are difficult to absorb. In both figures, (a) is a broad pattern, (b) is a peak pattern, and (C) is a pattern of (a) and (b). In the example shown in Figure 10, the broad pattern (a) and peak pattern (b) are linearly expanded and contracted as shown by dotted lines, but Figure 10 (c)
As shown by the circle d in the figure, there is a drawback that protruding parts occur due to time fluctuations, making it difficult to absorb 5-hour fluctuations.

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, this was done with the aim of improving the ability to absorb time fluctuations by adding only a portion of patterns that are shifted in the time direction and patterns that are not shifted in the time direction.

In order to achieve the above object, the present invention includes an audio input section, an analysis section, and a pattern holding section that holds pre-registered patterns, and binarizes input audio. In a speech recognition device that recognizes patterns, (1
) with the input audio.

After adding the pattern shifted in the time direction,
Performing binarization processing and overlapping multiple input patterns for one registered voice to create a standard pattern;
(2) Creating a standard pattern by binarizing the input audio and then adding patterns that are shifted in the time direction, overlapping multiple input patterns for one registered audio, (3) The input audio is binarized.

After superimposing a plurality of input patterns for one registered voice, a standard pattern is created by superimposing patterns obtained by shifting this pattern in the time direction (4)
) A standard pattern is created by binarizing the input audio, and when unknown audio is input during recognition, this pattern is binarized, shifted in the direction of 1 hour, and added together for recognition. It is characterized by Hereinafter, the present invention will be explained based on examples.

FIG. 1 is an electrical block diagram for explaining one embodiment of the present invention. In the figure, 1 is a microphone, 2 is a voice section detection section, 3 is a feature analysis section, 4 is a binarization section, 5 is a register, 6 is a one frame shift section, 7 is an addition section, 8 is a binarization section, 9 is a standard pattern section, 10 is a collation section, 11 is a result output section, and in this embodiment, it is an audio input section. In a speech recognition device that includes a feature analysis section, and a pattern holding section that holds pre-registered patterns, the input speech is recognized using a binary pattern. After adding the shifted patterns, a binarization process is performed, and a standard pattern is created by overlapping a plurality of input patterns for one registered voice and is recognized. Extract the audio section, perform feature analysis such as spectrum, and once import the audio feature pattern into the register 5, add the pattern shifted by one frame in the time direction and the original pattern, and then add this pattern. Binarize it and use it as a standard pattern. after that,
During recognition, the characteristic pattern of the unknown voice and the standard pattern are linearly expanded and contracted, superimposed and compared, and the degree of similarity is determined and the recognition result is output. This makes it difficult for local peaks to protrude from the broad pattern, improving the ability to absorb temporal fluctuations. Here, I explained shifting by one frame, but
It doesn't necessarily have to be one frame, just find the optimal value so that there are fewer protrusions from the broad pattern as shown in Figure 10. However, if you shift it too much, the pattern difference between words will disappear. 1 or higher, the ability to absorb even if the peak pattern is subject to time fluctuations is improved.

FIG. 2 is an electrical block diagram for explaining another embodiment of the present invention. Reference numbers are included. In this embodiment, after binarizing input audio, patterns obtained by shifting the binarized audio in the time direction are added, and a standard pattern is created by overlapping multiple input patterns for one registered audio. The method shown in Figure 1 shifts the data in the time direction and adds them together before binarization processing, whereas this method shifts the data in the time direction and adds them after the binarization process. This is how it was done. Actually used for recognition is 2 unknown inputs.
Since this method is converted into values and calculated, the effect of temporally shifting the pattern appears better. In this case, the binarized pattern will be ternarized if the signals are superimposed and added at different times, so it is necessary to pass the signals through the binarization processing section again. In addition, there is a method in which one word is uttered multiple times by the same person or by different people, and the sum or average is taken. In this case, the input voice is After performing binarization processing and superimposing a plurality of input patterns for one registered voice, patterns obtained by shifting this pattern in the time direction are superimposed to create a standard pattern, which is recognized.

FIG. 3 is an electrical block diagram for explaining an example of such a case. In this example, patterns shifted in the time direction are added again as shown in FIG. It is also possible to create a pattern by binarizing it, create a pattern uttered multiple times in this way, superimpose it, and take a weighted average. They are shifted in the direction and overlapped at the overlapping portion 11. In this case, since the shifted patterns are added together, the values of each element of the pattern are different, so they are normalized by the normalization unit 13. When the three utterances of the binarized pattern are superimposed, it becomes a four-value pattern from O to 3,
If you stack and add patterns that are time shifted before and after one frame, you will get a 10-value pattern from 0 to 9, so normalization here refers to converting the 10-value pattern back into a 4-value pattern. means.

FIG. 4 is an electrical block diagram for explaining another embodiment of the present invention. In this embodiment, input speech is binarized to create a standard pattern, and unknown patterns are When audio is input, this pattern is binarized and then shifted in the time direction and added together for recognition. This is a method that adds a time shift effect to the input pattern without time shifting the #j quasi-pattern, and adds a slight weight to the input pattern where the standard pattern and the input pattern did not overlap at all in the above method. This makes it easy to do.

FIG. 5 is a diagram showing an example of time-shifted superimposition (shifting one frame in a -dimensional pattern).

(a) Time-shift the pattern in the figure (b) make it as shown in the figure, then superimpose the pattern in the figure (a) and the pattern in the figure (b) to create the pattern in the figure (Q), which is then normalized. The pattern shown in Figure (d) is created by changing the way the pattern is shifted depending on the time. This is 2
When comparing two patterns, the beginning and end are matched, and the pattern lengths are matched by linear expansion and contraction between them.

Since the effect of time fluctuation is small near the beginning and end of the pattern, and large in the center of the pattern, the time shift is large for the portion where time variation is large, and is small at the start end. An example of shifting the pattern in this way is shown in Figure 6. In this case, the start and end 1/3 of the pattern is not time shifted, but only the center 1/3 is shifted by one frame. I'm letting you do it. Furthermore, as a method to obtain a similar effect, a pattern with a certain time lag and a pattern with no time lag are superimposed and only the central part is added, as shown in Figure 7. Since the number of patterns added differs, care must be taken in normalization.By adding 0 or more, it is expected that the time-varying ability of linear expansion/contraction matching will be improved.

As is clear from the above description, according to the present invention, it is possible to improve the ability to absorb time fluctuations of a matching method based on linear expansion and contraction.

[Brief explanation of drawings]

1 to 4 are electrical block diagrams for explaining the present invention in detail, FIG. 5 is a diagram for explaining the principle of time-shifted superposition, and FIGS. 6 and 7. 8 is a diagram for explaining an example of the BTSP method, and FIG. 9 and FIG. FIG. 1... Microphone, 2... Voice section detection unit, 3
...Feature analysis section, 4...Binarization section, 5...Register, 6...1 frame shifting section, 7...Addition section, 8
. . . Binarization unit, 9 . . . Standard pattern, 10 . Patent applicant Ricoh Co., Ltd. Figure 1 Figure 3 Figure 4

Claims (12)

[Claims] (1) comprises a voice input section, a feature analysis section, and a pattern holding section that holds pre-registered patterns;
In a speech recognition device that recognizes input speech using a binarized pattern, the input speech and a pattern shifted in the time direction are combined, and then binarized to create one registered speech. A speech recognition device characterized in that a standard pattern is created by overlapping a plurality of input patterns for recognition. (2) The speech recognition device according to claim (1), wherein the pattern is shifted in a different manner depending on time. (3) The speech recognition device according to claim (1), wherein only part of the pattern shifted in the time direction and the part of the pattern not shifted in the time direction are added together. (4) comprising a voice input section, a feature analysis section, and a pattern holding section that holds pre-registered patterns;
In a speech recognition device that recognizes input speech using a binarized pattern, the input speech is binarized and then patterns that are shifted in the time direction are added together to recognize multiple input speech patterns for one registered speech. A speech recognition device is characterized in that a standard pattern is created by overlapping two patterns, and the standard pattern is recognized. (5) The speech recognition device according to claim (4), wherein the manner in which the pattern is shifted in the time direction differs depending on time. (6) The speech recognition device according to claim (4), wherein only part of the pattern shifted in the time direction and the part of the pattern not shifted in the time direction are added together. (7), comprising a voice input section, a feature analysis section, and a pattern holding section that holds pre-registered patterns;
In a speech recognition device that recognizes input speech using a binarized pattern, the input speech is binarized, multiple input patterns for one registered speech are superimposed, and then this pattern is converted over time. A speech recognition device characterized in that a standard pattern is created by superimposing patterns shifted in a direction, and the standard pattern is recognized. (8) The speech recognition device according to claim (7), characterized in that only part of the pattern shifted in the time direction and the part of the pattern not shifted are added together. (9) The speech recognition device according to claim (7), wherein the pattern is shifted in a temporal direction differently depending on time. (10) A speech recognition device that includes a speech input section, a feature analysis section, and a pattern holding section that holds pre-registered patterns, and that recognizes input speech using binarized patterns. A standard pattern is created by binarizing the voice, and when an unknown voice is input during recognition, this pattern is binarized, shifted in the time direction, and added together for recognition. voice recognition device. (11) The speech recognition device according to claim (10), wherein the pattern is shifted in a temporal direction differently depending on time. (12) The speech recognition device according to claim (10), characterized in that only a part of the pattern shifted in the time direction and a part of the pattern not shifted in the time direction are added together.