JPS58192095A - 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 speech recognition device.

This kind of voice recognition device is a voice signal! By extracting a time series of characteristic parameters, e.g., SIN number spectral values, from the input signal containing the input signal, and detecting the time value t$ of the audio signal from this input signal, Create an audio pattern consisting of a time series of frequency spectrum values, and this pattern is well known. Therefore, the input audio! In order to reliably recognize MJiI, it is necessary to create an audio pattern that correctly represents the characteristics of this input audio, and for this reason, it is necessary to accurately detect the time domain of the audio signal.

Traditional sound Y! ! ! ■In the identification device, the power value of the input signal is extracted and the iJ! As shown in Figure 1 1ml, this power value S was set higher than the noise signal level! The time point when J2 exceeds the specific value TH,2 is set as the starting point t8 of the time interval T of the audio signal, and this power value S becomes the second specific value TH,ZY.
For a certain period of time during which the falling state is longer than the pause time of the audio,
For example, the time point when 200 m5ec has elapsed is set as the end point of the time domain of the f-signal. Moreover, at this 94 signal j1
]@KinT A (1゛C, its power value S is set lower than the peak value of the power fa S of fN. Since the required condition is °C, when the bell is low (I and the bell is low,
The time-domain T-field detection of human power 1i4f is quite accurate.

However, when you are in a situation where there is a lot of ambient noise, the noise level input along with the voice becomes high. As shown in FIG. 1b+, the second specific value TH,
2'l: It was necessary to set it to a large value, and therefore the time mji&T of audio No. 1 was not accurate.

The conventional fm recognition device has the above-mentioned disadvantage C binary value, which is expressed in characteristic 1 of f city, and cannot obtain an accurate Kurota pattern.
The purpose of the present invention is to solve this drawback (7), which can lead to erroneous recognition of the input voice, and to eliminate the negative effects of miscellaneous sounds and create accurate voice patterns. The present invention provides a speech recognition device equipped with a means for obtaining a speech recognition device.

A block diagram of an embodiment of the ffield recognition device of the present invention is shown in the second part, and a timing diagram of the operation of this device is shown in Y! Figure J5 shows T6 In the same figure, ([) is the microphone that inputs the f field, (
2) is a microphone amplifier that amplifies the input signal including the audio signal from the microphone 1); (3) is the microphone amplifier (2);
This is a parameter extraction circuit that extracts characteristic parameters X1, for example, a time series Xty of frequency spectrum values, from an input signal from an input signal. (4) is a parameter buffer that stores the special parameters Xtv from the parameter extraction circuit (3) in time series; (5) is the time series Stv extraction of the power value S of the input signal from the microphone amplifier (2); For example, it is a power extraction circuit that calculates the sum of frequency spectrum values in all station wave number bands l:. (6)
is a $1 comparison circuit, and the puff value st from the power extraction circuit (5) is 1J! It is compared with the first specific value TH,1 similar to that shown in FIG.
A voice presence signal V1 is outputted. (71 is the second comparison circuit, which outputs the power value S from the power extraction circuit (5)
When t is compared with a second specific value TH,2 similar to that shown in Fig. 141, and S is also: > T) 1.2, the second voice presence signal ■! Output. (8) is the above parameter extraction circuit (3
Difference of parameter Xt from 1 -ΔXt, -Xt-Xt
-1) il calculation T le ff minute [a road, t9+1! @SQ
, ) is a ratio 8 circuit, which records the differential value ΔXt from the differential circuit (81).

is compared with a third specific value TH,5, ΔX t,)T)
When f, 3, the third f stomach existence reconnaissance v5 is output T.

Note that this third specific value TH,5 is lower than the special parameter Xto)q conversion rate ΔXt specific to the 1lliB, and is the change rate ΔX of the characteristic parameter Xt of the ambient noise that does not change much.
It is set to a value higher than v. GO is a timer circuit having a Kawanta configuration, and is connected to the fifth ground restricting circuit (9) above.
Comparison of the third Onda Zonkyo signal VM and the above vIIL2 from
b) ■ of the f-voice presence signal of 弗2 from (7)! The first or second Af expressed by the logical sum 011 with
if presence signal Vz+vmy received °C1ζ・signal v
A new fourth (7) signal is added until the time when 2+■5 exists and 200 zero seconds have passed since this signal ended.
ii lli Presence signal V a 1kQtET, 0
3c No. 117 above. ) comparison circuit (the first obtained from 61
Audio presence signal ■1 time series V (first audio presence signal buffer memory for storing tv, 03 is timer circuit written in L)
Time series of f-voice presence signal V4 of wJ4 obtained from ■4℃
This is the fourth audio presence signal buffer memory that stores the rlf buffer memo! 705,031 each [%V 1t
, V41dI k Parameter Buffer Amesori (4)
is associated with the feature parameter Xt. 1114
1 is a readout control circuit for reading out the special at Based on JO3ThO:l, the first fw presence signal v
4th voice presence signal immediately before time t'R when 1 occurs ■4
Let the generation time t8 be the starting point of the time domain T of the f-voice, and the fourth f-voice presence signal v4 immediately after the time t'e when the first f-voice presence signal ■1 ends at the time t#. to 20 (lf
The existence of the sound beak at the time before s8c, that is, 1113 'pf
■The end point of s is the end point of the time domain T of t e lie f, and the characteristic parameter of the L period parameter buffer memo January 4) included in this time @VT:ts~to is
1) Time series Xc8~Xtet continues "r, as is human power 1
m pattern memory, and stores the characteristic parameters Xt successively retrieved from the parameter buffer memory (4) in the time series Xt8 to XC01 as a manual f4 pattern.

It is the old pattern recognition circuit and the input it voice pattern is seven! The human power @* at this time is recognized by comparing the 105+ human power-1w pattern with a large number of pre-stored registered VrfH patterns and selecting the registered VrfH pattern that is most similar to tjA.

In the speech recognition device having such a configuration, the temperature is 1°C. Figure 3 shows that the power of the f-voice signal reaches its peak at the time @ area T of the f-picture signal, similar to the conventional example shown in Figure 1. Audio presence signal ■t.

and miscellaneous f (I level 'IK greater than I+ level'
T[l#(1st power 11S indicates that n'C is obtained T2nd it#'[E signal V! 0) 14:15
mM T 1 , T 2 Y contains, but this it
- In order to determine the start and end points of the time domain T of the signal, the third point is that the rate of change ΔX of the characteristic parameter X of the f voice signal is obtained which is greater than the rate of change of the characteristic parameter of the noise No. The sound row exists γE (different from the conventional device in that the start φ and end point of the g-number V3 are used, and moreover, in the time domain T1 where W is used.

f? tf presence signal V 5 (b2Q 011fllieQ
The interruption of Heman.

When the audio signal is paused, the audio signal is paused.

Accordingly, °C1 such a speech recognition device is as shown in FIG.
Even if it is used in a situation where the fear f multiplier level is high, the second specific value TH,2 is greater than this axial signal level (
It is sufficient to set T to ilTH, 2', and depending on the third voice presence signal v5 that does not depend on the height of the miscellaneous if @ signal level, the start point ts and end point rig of the l'&tl region T are extremely Derived accurately. Therefore, in the C input 1w pattern memory 09, an f-voice pattern consisting of feature parameters Xt included in the time domain T of accurate f-voice is obtained. Recognition is made.

Fifth, the parameter memory 071 of another embodiment of the speech recognition device of the present invention is provided, and this koji 1 parameter memory un
Each noise parameter Xty averaging a & 54 tq value Jt151 human power frH
Each feature parameter X including the noise acid content Y that makes up the pattern T
The input ffI pattern from which #ity has been removed is stored in the human voice pattern 09', which is C1h, by the derivation 119 subtracted from t. The characteristic parameter X't of the stored *-i 4t number depends on the readout control circuit IHiE [, the first f-voice existence signal 1 and the fourth Gyoda existence (if
This is the result of reading out the parameter Xt of the parameter buffer (4) in the time domain where neither number (4) and number (4) appear.

Purchase example of a certain configuration ≦Near/' and new input 'iH
The input ff457 input turn stored in the t<turn memory Oi is.

The miscellaneous components are completely removed, and the pattern recognition circuit flG performs the pattern 41 ta, fil+and the pattern recognition is positive+4.

As is clear from the above 1., the i-cho Shinobu @ device of the present invention generates a third f-voice presence signal in which the difference obtained by time-differentiating the voice feature parameter time series is compared with the value of one foot. Origin:
Based on the time, the time of departure of fb'l territory weir T at the time of Y voice!
Since the determination is made, the exact input start point of the input is obtained, and the input is determined. Based on the end of the third f-voice presence signal, the end point of the time @ area T of Jin's stomach is determined, so the accurate input path r time of the input f-voice is obtained, and the time area T of this fWI is determined.
°C1 accurate voice pattern depending on the feature parameters included in! Created by Tru-tei Kadesaru. Also, there is an interruption in this fifth Soni presence signal within one foot time ° [also this! The time domain of continuous fllll is detected, eliminating the inconvenience of separating the sound mv at the pause point of fw.

Furthermore, in the noise feature parameters, [ff feature parameters! Sound patterns are created from the time series of feature parameters that are compensated! Since it is easy to create, that is, the f-voice pattern that eliminates the adverse effects of noise signals such as ambient noise is used for pattern recognition, there is no need to get used to erroneous recognition, and a significant improvement in the recognition rate can be expected.

[Brief explanation of the drawing]

Figure 1ml and lbl respectively show the operation of the conventional fw MI identification device. and 1
14 are timing diagrams showing the operation of the apparatus of the present invention, and FIG. 5 is a block diagram showing the main parts of another embodiment of the apparatus of the present invention. 131...Parameter extraction circuit, (4)...Parameter buffer memory, (5)...Power extraction circuit,
(61...first comparison circuit, (71...second comparison circuit, (8)...difference circuit, (9)...third comparison circuit, (1(1... Timer circuit, 03...
! J1 f voice presence signal buffer memory OJ... s4 fwI presence signal buffer memory, 041... Readout control circuit, as+od... Input fWI pattern memory, 061... Pattern twill weave circuit, 171... Noise Parameter Memory Applicant Sanyo Electric Co., Ltd. Agent Patent Attorney Shizuka Sano ゛・・・・・Fig. 1

Claims (1)

[Claims] 1) fFl!, which consists of time-series analysis of voice feature parameters! Pattern to pattern iJmTru sound t! ! Recognition device 1: A power extraction means for extracting the power of the input [4], a difference circuit for calculating the difference value of the time series of the characteristic parameters of the tone, and a power value from the power extraction means for the first identification. Condolence 1 that outputs a voice presence signal of 1Jls1 during the period when the value is retained
The power value from the comparator circuit and the power extracting means written in L is
a comparison circuit of T@2 which outputs a second audio presence signal during a period in which the second specific value, which is lower than the specific value of the first one, is exceeded;
a comparison circuit of Ii'S3 which outputs a voice presence signal of vIPI3 during a period in which the difference value from the difference circuit exceeds a third specific value; At least one first acoustic presence signal! A voice recognition device is characterized in that a voice pattern is created from a time series of acoustic feature parameters after the generation of a third Takeda presence signal, which is immediately before the generation of the second sound presence signal. 2) Based on the fFII presence signal from each of the above comparison circuits8,
The feature is that an acoustic pattern is created from a time series of acoustic feature parameters before the end time of the third voice presence signal after the second f-voice presence code completion time @ which includes at least one first voice presence signal. f in claim 181
-u7I professional equipment. 5) A timer means is provided in the comparison circuit of the ILS, and the timer means 1: Therefore, the above! If the time from the end point of the third acoustic presence signal from the comparison circuit of J6 to the next generation point is less than a certain time, the patent claim that makes this third sound presence signal continuous The speech recognition device according to item 1 or item 2 of scope. 4) During the period C2 in which none of the sound presence signals from the above comparisons (b) are obtained, the characteristics of the noise Para 1-IF
Create a voice pattern 21 by compensating the value of Aota's feature parameter vm depending on the value of the feature parameter of this noise.
Sections 111 and 182 of patent claims that feature a pavilion
Additionally, the voice u7j recognition device described in -X@5.