JPS58168100A - 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] The present invention relates to a monosyllabic speech recognition device.

In recent years, various devices for recognizing speech have been devised.

Such a speech recognition device uses an n-dimensional input feature vector series (a
a ... a 1 ), A1 la 2g P standard pattern vector sequences (b1', b2t ... b, p ) registered in the dictionary in the staking device
. . (blp, b2p, . . . bkp), the one closest in distance or the one with the highest degree of similarity is selected and recognition is completed. At this time, the analysis methods for determining the input feature vector and standard pattern vector include a filter bank consisting of a combination of multiple filters, a method of frequency analysis using Fourier f4, a method of using linear prediction coefficients, and a method of using Kegnutram coefficients. There are methods etc.

Now, when attempting to recognize speech using the method described above, Ibl, Idl, Icrl, 1ml.

Speech containing voiced consonants such as In+, 1rI, etc.
, ``da'', ``ga'', ``ma'', ``su'', and ``u'' are voiced consonants such as ``da,''``ga,''``ma,''``su,'' and ``u.'' At the beginning of utterance, only the so-called busbar vibration waveform appears as a spectral feature. Due to the similarities in the voiced consonant parts, conventional speech recognition devices tend to misrecognize voiced consonants when comparing voiced consonant patterns as described above.

On the other hand, even in the case of voiceless plosives such as IPI, Itl, and lkl, the spectral patterns showing the percent signature of Ipl, ltl, and lkl have characteristics for only a relatively short period of time compared to the earth phoneme. lpl with the same fearful time length as the phoneme
, l t l , and lkl, the proportion of the following vowel's sub-range becomes larger, and the following vowel is detected as a consonant of a voiceless plosive. This is one of the causes of lower recognition rates.

A more specific explanation will be given below with reference to FIG. 1.2.

Figure 1 (,) is a frequency analysis of the utterance of the monosyllable ``・ku'', with the vertical axis showing amplitude, the horizontal axis showing frequency, and the slope showing time. In Figure 1 (-), the waveform occurring at time T1 is a part called no1zunoku-, where the vocal cords vibrate prior to articulation, that is, the action that makes the shape of the mouth and vocal tract unique to a particular phoneme. Generally, only low frequency fireflies have a spectral peak. Next time T2
Denokuzuno (partially followed by a spectrum caused by the shape of the vocal tract and mouth unique to each phoneme, which occurs over a relatively short period of time, and then at time T3, the transition with the following vowel, the time This continues with the vowel stationary part of T=. Also, Figure 1(b),
(C) is a frequency analysis of the utterances of the monosyllables "ga" and "u", and time T1 indicates a part of each nokuzuba. As is clear from Figure 1 (a), (b), and (C), the buzzbar phenomenon is common to voiced consonants, and the spectra produced by these busbar phenomena are very similar. Recognizing speech from these spectra is extremely difficult, and is a contributing factor to speech recognition errors.This effect is especially significant in devices that recognize speech made by dividing monosyllables.

Next, we will discuss the silent rupture meter using Figure 2 (Figure 2).
-) is a frequency analysis of the utterance of the voiceless plosive "pa", and the waveform at time T6 is the characteristic part that best expresses the characteristics of the voiceless plosive "ba". T
7 is a crossover product with the following vowel.

Time T8 is the vowel stationary part. Figure 2(b) is a frequency urine analysis of the voiceless plosive sound ``me''.
-), time T6 is the silent part, time T6 is the special ia part,
Time f7 is the transition part, and time T8 is the vowel stationary part.

As is clear from Figure 2 (a) and (b), when
In the feature part 6, both have very similar spectra, and the feature part that has characteristics as a consonant occurs in a short period of time. This is often a cause of voice recognition errors.

In view of the above-mentioned drawbacks, the present invention makes it possible to perform speech recognition without being influenced by the buzz bar of voiced consonants or by the vowel part following voiceless plosives by cutting out a timed initial Peltl sequence using a pointer. The present invention provides a speech recognition device.

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

Figure 5g3 is a block diagram of a voice recognition device lid in an embodiment of the present invention. In the figure, reference numeral 1 denotes a parameter analyzer that inputs an input voice through an input signal terminal 1a and sequentially changes it into an n-dimensional input parameter vector series (a4, a2...a I J, 2 is a standard pattern memory in which a standard pattern vector sequence that becomes a standard pattern is stored in advance via a parameter analyzer 1 according to an instruction from a keyboard 3.

4 is a pointer memory that, when storing a standard pattern in the standard pattern memory 2, displays the standard pattern using a talk device 6, and indicates the starting point and ending point, which are specific head regions that are characteristic of the voice, according to instructions from the keyboard 3. I remember. ,6
Input the input vector sequence from the parameter analyzer 1 (the parameter vector series and the standard from the standard pattern memory 2) and the turn pertle series using the turn comparator, and input both vector series to the pointer in the pointer memory 4. 7 is compared by the pattern comparator 6 and is selected from among the standard pattern vector series. The device outputs the determination result through the terminal 7a.

Regarding the speech recognition device configured as described above.

The operation of voice recognition will be explained below.

First, the parameter analyzer 1 inputs the input voice to the human input signal terminal 1a.
, and an n-dimensional input parameter vector series (all, a2゜...a) as shown in Figure 4(a).
i) and output to the no-turn comparator 6.

As shown in FIG. 4(b), the edge standard pattern memory 2 stores in advance one of the standard patterns β, a standard pattern vector series (bl, b2, . . .
・b,,'[-Output to pattern comparator 6.

The pattern comparator 6 inputs the input parameter vector series (al, a2...d, l of the parameter analyzer 1 and the standard pattern vector series! column (1b, 1b'',...1b-) of the standard pattern memory 2. A), 1 2 and the standard pattern vector series (b1β, b
2β.

... Pointer (S!, εl) corresponding to bm'()
and input the pointers (Sβ, εβ) with both input parameter vector series and standard pattern vector series, and the fcd area 4 which is sandwiched between the pointer Sβ and the pointer
．． At L2, mutual vector sequences are compared.

That is, the input parameter vector sequence (al s a2
. . . ai j to pointer 81. The vector sequence when cut out in the region L1 given by εβ is (aBlm
"sIl, +1..., a,), and the standard pattern vector 1212, Il torque series (1b, 1b,... 1bm) was cut out in the area L2 given by the pointer 2sIl,, ε! The actual vector sequence is (b,, bJ+1 ,..., b
If gjβ, the pattern comparator 6 finds the distance 0 between both vectors using the following equation.

In the above manner, the distance 40"f between both vectors is determined by sequentially comparing all the standard turns stored in the turn memory 2. Then, the minimum distance Pa determining unit 7
Then, only the standard/-C turn vector series with the smallest distance 4D is selected, and its standard) (turn information) is outputted via the terminal 7a.

As described above, according to this embodiment, when storing a standard turn in the pattern memory 2, the keyboard 3 and the display 4
By registering in the pointer memory 4 the area in which features occur according to the standard pattern in 6, the input parameter vector series sent from the parameter analyzer 1 and the standard pattern vector series sent out from the pattern memory 2 can be combined. When calculating the distance between both vector series, the distance between both vector series is calculated in the area designated by the pointer registered in the pointer memory 4, so that very accurate speech recognition can be performed.

In this embodiment, the parameter analyzer 1 is configured by a filter puncture, but the parameter analyzer 1 may be an analysis means such as a Fourier transformer, an analyzer with flJl decimal coefficients, or an analyzer with cepstral coefficients.

Further, in this embodiment, the pointer S1. Although both the input parameter vector series and the standard pattern vector series were cut out in a characteristic region using εβ, the pointer Sμ, ε1
Either the standard pattern vector series or the input parameter vector series may be cut out in a specific region.

Furthermore, although the pattern comparator 6 and the distance minimum determination unit 7 are designed to find the minimum distance between the input parameter vector sequence and the standard turn vector sequence, they may be designed to find the maximum similarity. Needless to say.

As described above, the present invention provides a point storage means that stores in advance the starting point of a region having characteristics in each voice as a point, and stores the input feature vector series which is the output of the parameter analysis means and the standard pattern. The area of at least one of the four pattern vector series output from the means is determined by the points of the point storage means, and the distance between both vector series is calculated. By determining 4, it is possible to prevent the effects of the busbar of voiced consonants and the following vowels of voiceless plosives, and improve the recognition rate during speech recognition. There is.

[Brief Description of the Drawings] Fig. 1 is a characteristic diagram of voiced consonants, Fig. 2 is a characteristic diagram of voiceless plosives, Fig. 3 is a block diagram of a speech recognition device in an embodiment of the present invention, and Fig. 4 is a characteristic diagram of voiced consonants. It is a characteristic diagram of the voice when cutting out by evening. 1...Parameter analyzer, 2...Standard) turn memory, 4...Pointer memory, 6...
...Pattern comparator, 7...Distance minimum determination unit. Name of agent: Patent attorney Toshio Nakao and 1 other person No. 1
Figure 1 Figure 4 (α)

Claims (1)

[Scope of Claims] Parameter analysis means for performing parameter analysis of input speech and outputting an n-dimensional input feature vector sequence (a 1s a 2...k) (where i is an integer); Parameter-analyzed two sets (where P is an integer) of standard pattern vector sequences (b', b y,...
...b,')...[blp, b2p...1
2] A drift pattern storage means that stores ``bkp)'' (where j is an integer) and a point that stores in advance the start and end points of regions having characteristics for each voice. a storage means, and an input feature vector sequence (a 1 . a 2 ,
...at) and the standard pattern vector string (b 1.b 2...
bmn) (where m and n are integers m and m≦k); a comparison means for determining the distance between both vector sequences by determining an area using the points of the point storage means; Standard pattern vector sequence (blr, b2r...-b9rJ (however, "s
q is an integer) is a speech recognition device.