JPS6075900A - Word 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 [Technical Field of the Invention] The present invention relates to a word speech recognition mWA@, and particularly relates to, for example, an improvement in detecting the beginning and end of a word speech section.

[Prior Art] FIG. 1 is a schematic block diagram showing an example of a conventional word speech recognition device. In the figure, an audio signal input from a microphone 11 is amplified by a microphone amplifier 12 and then provided to an AGC circuit 13. The AGC circuit 13 is a circuit that automatically controls the gain of an internal amplifier so that a constant output can be obtained even if the magnitude of the input signal changes. The output of the AGC circuit 13 is given to an Ay'D conversion circuit 171 and converted into a digital signal. The output of the A,.../D conversion circuit 14 is given to a waveform meter 15. This waveform memory 15 is a memory that temporarily stores one frame of input waveform data. The output of the waveform memory 15 is sent to the power calculation circuit 21 and also given to the feature extraction section 3. The power needle circuit 21 is a circuit that calculates the power of a waveform. The output of the power calculation circuit 21 is given to the recognition processing section 6 as well as to the start/end detection circuit 22 . The start/end detection circuit 22 is a circuit that detects the start/end of the audio signal, and its output is given to the recognition processing section 6.

On the other hand, the feature extraction section 3 includes a digital filter and the like, and also includes a circuit for extracting feature parameters of the waveform of the input sound Fl11. The output of the feature extraction section 3 is given to the recognition processing section 6. An input pattern memory 4 and a registered pattern memory 5 are connected to this recognition processing section 6. The input cover turn memory 4 is a memory that temporarily stores minute parameters of the speech to be recognized that have been analyzed and extracted by the feature extraction section 3 in the word speech recognition mode. The registered pattern memory 5 is a memory that stores in advance characteristic parameters of registered words analyzed and extracted or characteristic parameters of standard speech in the registered mode. The recognition processing unit 6 is a circuit that includes, for example, a microprocessor, a microcombi processor, etc., and performs recognition processing using the characteristic parameters in the input pattern memory 4 and the registered pattern memory 5. In such a word speech uN@ arrangement, the speech interval is called a frame, which is a fixed period of time l1lI! 1111i
4-7t* extraction of audio is performed for each frame.

Next, the operation of the circuit shown in FIG. 1 will be explained. Microphone 1
The audio signal input from microphone amplifier 12.
AGC circuit 13. The signals A and 'D are passed through the diversion circuit 14 and stored in the -H waveform memory 15.

The feature extractor 3 receives one frame of waveform data from the waveform memory 15 and extracts feature parameters. The obtained feature parameters are stored in the tJfl record pattern memory 5 in the registration mode. On the other hand, in the recognition mode, the 1q feature parameters are temporarily stored in the input pattern memory 4, and thereafter recognition processing is performed in the recognition processing section C using techniques such as pattern matching.

On the other hand, the start/end detection circuit 22 detects the start/end of the sound @ section based on the power of the audio signal calculated by the power calculation circuit 21. The recognition processing unit 6 performs recognition processing on the audio signal in the section defined by the start/end detection circuit 22 as the audio fH that should not be recognized.

FIG. 2 is a diagram showing the power waveform of the audio signal.

Referring to FIG. 2, the operation of the start-stop detection circuit 22 shown in FIG. 1 will be described. The start/end@I detection circuit 22 detects the beginning of the audio signal when the power of the audio signal exceeds a preset threshold ps, and detects frames whose power is less than the preset threshold (i&P e). When the set interval Kth continues, the end of the audio signal is detected.In this example, 1, 1, and 2 are the start and end frames of the audio signal, respectively.As mentioned above, the recognition processing unit 6 detects the end of the audio signal. The IIN detection circuit 22 recognizes the audio signal section determined by 1.Q, that is, the section 1 to 2, as a single HH speech that should not be recognized. Detection performance has a large impact on recognition results.

FIG. 3 is a diagram showing the power waveform of the audio signal when noise is added to the IJ audio. The audio signal section of this waveform (
As shown in Fig. 2, even though 1 (1 to 2), the beginning/end detection method of the conventional device detects that 3 to 4 are audio signal sections.In this way, f, Y The conventional voice recognition device is difficult to use in a noisy environment.
This method has the disadvantage that recognition performance is degraded because the beginning and end of the voice cannot be detected accurately.

[Summary of the Invention] The present invention was made in order to eliminate the drawbacks of the conventional device as described above. CORI~
1△ The purpose is to

[Embodiment of R-light] Figure 4 shows an embodiment of this invention! 1' is a schematic block diagram. In the figure, the embodiment of FIG. 4 is the same as the circuit of FIG. 1 except for the following points, and corresponding parts are marked with the reference image @ of 1ffiJ and their explanation will be omitted. In the embodiment of FIG. 4, a CORM A× calculation circuit 23 is provided. This CORMAX calculation circuit 23 is configured to include an autocorrelator consisting of a multiplier and an adder, for example, and
Audio old @ waveform CO input from waveform menu
Calculate RIvl AX. GL calculation CORMAX
is given to the start/end detection circuit 220 manually and also to the recognition processing section 6. The output of the power calculation circuit 21 is given to the other input of the start ii'tL edge detection circuit 220. That is, the feature of this embodiment is that the start and end ends of the audio signal are detected based on the power calculated by the power calculation circuit 23 and the C ORMA be.

Next, CORM and toX will be explained. Waveform data for one frame × (1), <1-1.2°...bow f
), the power P is expressed by the following equation (1).

The τ-order autocorrelation function COR (τ) is calculated by the following formula (2)% Formula % (2) If the interval of the autocorrelation function set to calculate CORhl A
Rh=I A X is expressed by the following equation (3).

C0Rtvl AX=klIAX [COR(r)
1. (3) Even if the U waveform has the same power level, CORM△X will be large for waveforms with strong pitch characteristics such as vowels.
Waveforms such as environmental U sounds that are close to white noise are CORM A
is small. This first step utilizes this fact to detect the beginning and end of the audio signal.

FIG. 5 is a diagram showing the CORM AX waveform when noise is added to the voice, but as shown in the figure, the influence of the noise is alleviated in the CORMAX waveform. Therefore, similar to the circuit of the first cause, threshold discrimination is performed to detect the beginning and end of the 1ζ
In this case, the audio signal section is from 5 to 30 seconds, and the start and end points can be detected more accurately than conventional devices. In this defense, even if the beginning and end of a word sound is detected based only on COR~1AX, J:0 is used. By using a combination of signals corresponding to waveform sizes such as and CORMAX
and is used.

In other words, in the example i of FIG. 4, a frame in which the power of the audio signal is greater than or equal to a preset threshold ps and the value obtained by multiplying CORMAX by a preset constant Cs is greater than or equal to the power is designated as an audio signal. CORM A
A frame in which at least one of the values multiplied by a preset constant Ce @ is less than or equal to the power is defined as a silent frame, and this silent frame is set as a frame H in which ε is determined in advance.
Although the signal continues for Kth, the end of the voice signal is detected and the first silent frame is taken as the end frame of the voice signal.

Next, a more detailed operation of the embodiment shown in FIG. 4 will be explained.

Power measurement circuit 21 and CORM A X calculation circuit 2
3 receives one frame of waveform data from the waveform memory 15, calculates the equations (1) and (2), and sends the power and CORM A X to the start/end detection circuit 220. The start/end detection circuit 220 detects power and C0.
Using the RN/IAX, the start and end of the audio signal is determined by the method described above, and the results are provided to the recognition processing section 6.

The recognition processing section 6 stores the audio signal from the start to the end detected by the start/end detection circuit 220 in the input pattern memory 4 or the registered pattern memory 5, and performs recognition processing. Note that the other operations are similar to those of the conventional device shown in FIG.

In addition, as another embodiment, the detection of the start and end edges in the start and end edge detection circuit 220 is performed using a threshold value 2 whose power is set in advance.
3 or more and CORM A KillIICM
A frame that satisfies at least one of e or less is defined as a silent frame, and when this silent frame continues for a preset number of frames Kth, the 1/'4 end is detected and the first silent frame is defined as the end frame of the audio signal. You can do it like this.

In this case too, 8! It is likely that the same effect as the embodiment shown in the T4 diagram can be achieved.

Further, as an example of the pond, when detecting the starting edge, a condition may be added that the power value is larger than the power value of one frame before, and this May conjunction can improve the ability to detect the starting edge.

Furthermore, in the above embodiment, the power of the audio signal is used as one element for detecting the beginning and end.
Instead of this power, m (waveform level, etc.) representing the size of the waveform of the audio signal may be used as an element for detecting the beginning and end.

In addition, in the embodiment of (above) E, a frame in which the power is greater than or equal to a preset threshold value 1) S and the value obtained by multiplying CORMAX by a preset constant C3 is greater than or equal to the power is considered to be the starting frame of the audio signal. However, the starting frame is around 1
Even if it is a frame of thought, the difference L7 does not support it.

Further, in the above-described embodiment, the last frame of the audio signal is the first silent frame, but the last 71 norm may be a frame in the vicinity of this frame.

Furthermore, in the above-mentioned embodiment C, for convenience of explanation, the word speech recognition device is intended to be registered to a specific speaker, but it is also possible to use a single speech recognition device for unspecified speakers whose characteristics of standard speech are registered in advance in the registration pattern memory. Of course, it may be a recognition device.

[Effects of the Invention] As described above, according to the present invention, since the beginning and end of word sounds are detected based on CORM AX, detection of the beginning and end of sound FtT [No. 8] is possible even in a noisy environment. can be performed accurately, and the recognition performance of speech recognition can be improved.

[Brief explanation of drawings]

FIG. 1 is a schematic block diagram showing an example of a conventional single-tube speech recognition device. , FIG. 2 is a diagram showing the power waveform of the audio signal. FIG. 3 is a diagram showing a power waveform when noise is added to voice. , FIG. 4 is a schematic block diagram showing one embodiment of the present invention. Figure 5 shows how noise is added to audio.
It is a figure which shows the COR1LIAX waveform of 'Iel:Ara-n'. In the figure, 3 is a feature extraction unit, 4 is an input pattern memory, 5 is a registered pattern memory, 6 is a recognition processing unit, 11 is a microphone, 211t power unit circuit, 23 is a CORM
AX calculation circuit; 220 indicates a start/end detection circuit; Representative: Masu Oiwa

Claims (1)

[Scope of Claims] (1) Extract the feature parameters of the input speech signal, and determine the word speech based on the degree of similarity between the extracted feature parameters and the feature parameters of a plurality of pre-registered word sounds. In a word speech recognition device that performs a recognition process, in order to define nine sections of a voice message in which the recognition process is performed,
A start/end detecting means for detecting the start/end of the input audio signal is provided, and the start/end @ detecting means measures the maximum value CORMAX within a predetermined range of the autocorrelation function of the input audio signal. A word speech WX recognition device comprising: CORMΔX calculation means for calculating CORMΔX; and start nm determining means for determining the start and end of a speech signal based on the CORMAX. (2) The start/end end determining means determines the start end of the audio signal section on the condition that the CORMAX exceeds a preset CORMAX threshold. Word speech recognition device, equipment. (3) The start/end determining means determines the end of the audio signal section on the condition that the input waveform in which the CORMAX is smaller than a preset threshold value of CORNIA X continues for a preset section. The word speech recognition device according to claim 1 or 2, characterized in that the word speech recognition device determines the number of words. (4) The start/end detection means further includes a level calculation means for calculating n by 1 corresponding to the size of the waveform of the input audio signal, and the start/end determination means calculates the value of the CORMAX and the level calculation means. Word acoustic recognition 1i according to claim 1, which determines the beginning and end of an audio signal based on the output.
ii Haze. (5) The word speech recognition device according to claim 1, wherein the level measuring means includes means for calculating the power of the input speech signal. (6) The start/end determining means determines that the power calculated by the power calculating means exceeds a preset power threshold and that the CORM A
6. The word speech recognition device according to claim 5, wherein the start end of the speech recognition section is determined on the condition that a threshold value of R to 1ΔX is exceeded. (7) The four stages of the start and end determiners are arranged so that the power calculated by the power calculation means exceeds a preset power threshold, and the value obtained by multiplying the CORMAX by a preset constant is less than the power. Provided that it is large, the voice (
The word speech recognition device according to claim 5, characterized in that the start end of section No. 3 is determined. (8) The start/end determining means at least determines that the CORMAX is smaller than a preset CORMAX threshold, and that the power calculated by the power calculation means is smaller than a preset power threshold. 6. The word speech mWA device according to claim 5, wherein the end of an audio signal section is determined on the condition that an input waveform satisfying either one of the conditions continues for a preset section. (9) The start/end determining means determines that the power calculated by the power calculating means is smaller than a preset power threshold and that the CORMAX is a preset number. The word according to claim 5, characterized in that the end of the audio signal section is determined on the condition that the input waveform satisfying at least one of the conditions of being smaller continues for a preset section. Audio mtA@M.