JPH0376473B2 - - Google Patents

Description

[Detailed description of the invention] The present invention relates to a speech recognition device.

This type of speech recognition device extracts a characteristic parameter, for example, a time series of frequency spectrum values, from an input signal including a speech signal, and on the other hand, detects the time domain of the speech signal from this input signal. This method creates an audio pattern consisting of a time series of frequency spectrum values included in a region, and performs pattern recognition on this.

Therefore, in order to reliably recognize input speech, it is necessary to create a speech pattern that accurately represents the characteristics of this input speech, and to do this, it is necessary to accurately detect the time domain of the speech signal. be.

In a conventional speech recognition device, the power value of the input signal is extracted, and the power value S is set higher than the noise signal level as shown in FIG.
The time point when the power value S exceeds the second specific value TH.2 is defined as the starting point ts of the time domain T of the audio signal, and the state in which this power value S is below the second specific value TH.2 is a certain period of time that is longer than the audio pause time. For example, the time point when 200 msec has elapsed is set as the end point te of the time domain of the audio signal. Moreover, in the time domain T of this pitch signal, the power value S exceeds the first specific value TH.1 (>TH.2), which is set lower than the peak value of the power value S of the voice. Since this was a necessary condition, when the noise signal level was low, the time domain T of the input voice could be detected quite accurately.

However, in situations where ambient noise is large, the level of the noise signal input together with the voice becomes high, so the second specific value TH.2' must be set large, as shown in Figure 1b. First, there is a problem in that the time domain T of the audio signal is not accurately detected, resulting in a time domain T' that is shorter than the actual time domain.

Due to the above-mentioned disadvantages, conventional speech recognition devices cannot obtain accurate speech patterns representing the characteristics of speech, resulting in erroneous recognition of input speech.

SUMMARY OF THE INVENTION The present invention has been made with the object of eliminating the disadvantages of speech recognition, and provides a speech recognition device equipped with means for removing the adverse effects of noise signals and obtaining accurate speech patterns.

FIG. 2 shows a block diagram of one embodiment of the speech recognition device of the present invention, and the timing diagram of this device is shown in FIG.
As shown in the figure. In the figure, 1 is a microphone that inputs audio, 2 is a microphone amplifier that amplifies the input signal including the audio signal from the microphone 1, and 3 is a characteristic parameter X of the input signal from the microphone amplifier 2;
For example, it is a parameter extraction circuit that extracts a time series Xt of frequency spectrum values. Reference numeral 4 denotes a parameter buffer memory that stores the characteristic parameters Xt from the parameter extraction circuit 3 in time series. 5 is the power value S of the input signal from the microphone amplifier 2
This is a power extraction circuit that extracts the time series St, and for example, calculates the sum of frequency spectrum values in all frequency bands. 6 is a first comparison circuit, which receives the power value from the power extraction circuit 5.
St is the first specific value TH.1 similar to that shown in Figure 1.
When St > TH.1, the first voice presence signal
Output V ₁ . 7 is a second comparison circuit, in which the power value St from the power extraction circuit 5 is compared with a second specific value TH.2 similar to that shown in FIG.
When St>TH.2, the second voice presence signal _V2 is output.
8 is the parameter from the parameter extraction circuit 3 above.
A difference circuit 9 is a third comparison circuit that calculates the difference value △Xt=Xt−Xt _-1 of Xt, and the difference circuit 8
The difference value ΔXt from ΔXt is compared with a third specific value TH.3, and when ΔXt>TH.3, a third voice presence signal V ₃ is output. Note that this third specific value TH.3 is lower than the change rate △Xt of the characteristic parameter Xt specific to the voice signal, and is a characteristic parameter of the ambient noise that does not change much.
The rate of change of X′t is set to a value higher than ΔX′t.
Reference numeral 10 denotes a timer circuit having a counter configuration, which is a logical sum 11 of the third voice presence signal V ₃ from the third comparison circuit 9 and the second voice presence signal V ₂ from the second comparison circuit 7. A new fourth voice presence signal V ₂ +V ₃ is received, and a new fourth voice presence signal V 2 +V 3 is received, and a new fourth voice presence signal V 2 +V 3 is received, and a new fourth voice presence signal V ₂ +V ₃ is received, and a new fourth voice presence signal V 2 +V 3 is received, and a new fourth voice presence signal V 2 +V 3 is detected for a period of time during which this signal V 2 +V 3 is present, and further 200 msec after this signal ends.
generates a voice presence signal _V4 . 12 is the first above
The first voice presence signal obtained from the comparison circuit 6 of
A _first voice presence signal buffer memory 13 stores the time series V ₁ t of the voice presence signal V 1 , and a fourth voice presence signal buffer memory 13 stores the time series V ₄ t of the fourth voice presence signal V ₄ obtained from the timer circuit 10 . and each signal V ₁ t of these buffer memories 12 and 13,
V ₄ t is associated with the characteristic parameter Xt of the parameter buffer memory 4. Reference numeral 14 denotes a readout control circuit for reading out the feature parameters Xt stored in the parameter buffer memory 4, which reads out the first voice presence signal V based on both the first and fourth voice presence signal buffer memories 12 and 13. The fourth voice presence signal V immediately before the time t′s when ₁ occurs _{. 4} Let the generation time ts be the starting point of the time domain T of the voice,
On the other hand, the fourth voice presence signal V ₄ immediately after the time t′e when the first voice presence signal V ₁ ends is a time point 200 msec earlier than the end time t″e, that is, the third voice presence signal
Let the end point te of V ₃ be the end point of the audio time domain T,
The time series ts to Xte of the characteristic parameters Xt in the parameter buffer memory 4 included in this time domain T:ts to te is read out. 15 is an input voice pattern memory, and the parameter buffer memory 4
Time series Xts of feature parameters Xt read from
~Storing Xte as an input audio pattern. Reference numeral 16 denotes a pattern recognition circuit, which compares the input voice pattern in the input voice pattern memory 15 with a large number of registered voice patterns stored in advance, and selects the most similar registered voice pattern. Input speech is recognized.

In the speech recognition device having such a configuration, as shown in FIG. 3, the power value S peculiar to the speech signal has a peak in the time domain T of the speech signal, similar to the conventional example shown in FIG. a first voice presence signal V ₁ indicating;
and a second voice presence signal V ₂ indicating that the power value S of the voice signal having a level greater than the noise level is obtained, including both time regions T ₁ and T ₂ of the voice signal. a third voice presence signal indicating that a rate of change ΔX of the feature parameter X of the voice signal that is greater than the rate of change of the feature parameter of the noise signal is obtained in determining the start and end points of the time domain T of
This device differs from the conventional device in that the start and end points of V ₃ are used, and furthermore, in the time domain T ₁ , interruptions of less than 200 msec of the voice presence signal V ₃ are recognized as pauses in the voice signal.

Therefore, as shown in FIG. 4, even if this speech recognition device is used in a situation where the noise level is high, the second specific value TH.2 can be set to TH.2' which is higher than the noise signal level. The start point ts and the end point te of the time domain T of the voice can be derived very accurately by the third voice presence signal _V3 , which does not depend on the height of the noise signal level. Therefore, the input speech pattern memory 15 obtains a speech pattern consisting of the feature parameters Xt included in the time domain T of accurate speech, and the pattern recognition circuit 16 performs reliable recognition of the input speech.

FIG. 5 shows the main parts of another embodiment of the speech recognition device of the present invention. This embodiment device differs from the embodiment device shown in FIG. 2 in that it includes a noise parameter memory 17 that stores noise characteristic parameters By subtracting 19 from each feature parameter Xt containing the noise component constituting the input speech pattern in the input speech pattern memory 15, the input speech pattern from which the noise component has been removed is reconstructed into a new input speech pattern 15'. It's in the points I saved up. In this case, the noise parameter memory 1
_The characteristic parameter X′t _of the noise signal stored in The characteristic parameter X't of the parameter buffer memory 4 is read out.

In this embodiment, the input speech pattern stored in the new input speech pattern 15' is
Noise components are completely removed, and the pattern recognition circuit 16 performs pattern recognition, that is, speech recognition in a square manner.

As is clear from the above description, the speech recognition device of the present invention recognizes speech based on the generation point of the third speech presence signal, at which the difference value obtained by time-differentiating the speech feature parameter time series is compared with a constant value. time domain T
Since the occurrence point of the input voice is determined, the accurate input start time of the input voice can be obtained. Also, since the end point of the voice time domain T is determined based on the end of the third voice existence, the input voice An accurate input end point can be obtained, and an accurate voice pattern can be created depending on the characteristic parameters included in the time domain T of this voice. Furthermore, even if the third voice presence signal is interrupted within a certain period of time, the time domain of the continuous voice is detected, eliminating the inconvenience of dividing the voice at the pause point of the voice.

Furthermore, since a speech pattern is created from a time series of feature parameters obtained by compensating the speech feature parameters with the noise feature parameters, the pattern recognizes the speech parameters that eliminate the negative effects of noise signals such as ambient noise. Therefore, there is no risk of misrecognition, and a significant improvement in the recognition rate can be expected.

[Brief explanation of drawings]

1a and 1b are timing diagrams showing the operation of a conventional speech recognition device, FIG. 2 is a block diagram showing the configuration of an embodiment of the speech recognition device of the present invention, and FIGS. 3 and 4 are timing diagrams showing the operation of a conventional speech recognition device, respectively. FIG. 5 is a timing diagram 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. 3...Parameter extraction circuit, 4...Parameter buffer memory, 5...Power extraction circuit, 6...
First comparison circuit, 7...Second comparison circuit, 8...
Difference circuit, 9... Third comparison circuit, 10... Timer circuit, 12... First voice presence signal buffer memory, 13... Fourth voice presence signal buffer memory, 14... Readout control circuit, 15, 15 ′
...Input voice pattern memory, 16...Pattern recognition circuit, 17...Noise parameter memory.

Claims (1)

[Claims] 1. In a speech recognition device that recognizes a speech pattern consisting of a time series of speech feature parameters, a power extraction means for extracting the power of input speech and a difference value of the time series of speech feature parameters are provided. a first comparison circuit that outputs a first voice presence signal during a period in which the power value from the power extraction means exceeds a first specific value; a second comparison circuit that outputs a second voice presence signal during a period in which the second specific value, which is lower than the first specific value, is exceeded; and a third comparison circuit that outputs a third voice presence signal during the exceeded period, and based on the voice presence signals from these comparison circuits, a second voice presence signal that includes at least one first voice presence signal. A speech recognition device characterized in that a speech pattern is created from a time series of speech characteristic parameters after a third speech presence signal generation point immediately before the signal generation point. 2 Based on the voice presence signals from each of the comparison circuits, a time series of voice characteristic parameters immediately after the end of the second voice presence signal and before the end of the third voice presence signal including at least one first voice presence signal. 2. The speech recognition device according to claim 1, wherein a speech pattern is created from a speech recognition device. 3. A timer means is provided in the third comparison circuit, and the timer means detects when the time from the end of the third voice presence signal from the third comparison circuit to the next occurrence is less than a certain time. The speech recognition device according to claim 1 or 2, wherein the third speech presence signal is made continuous. 4 During the period when none of the voice presence signals from the above-mentioned comparison circuits is obtained, the noise characteristic parameter is obtained, and the value of the voice characteristic parameter is compensated based on the value of this noise characteristic parameter. A speech recognition device according to claim 1, 2, or 3, characterized in that a speech pattern is created.