JP2007292814A - Voice recognition apparatus - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition apparatus, capable of removing audio sound which has conventionally been left to remain in an audio cancellation system, and capable of increasing the processing speed. <P>SOLUTION: An error signal is obtained by subtracting a signal of an adaptive filter 5, which is controlled by an LMS algorithm by using a step-size parameter (μ) adjusted by a step-size parameter adjusting section 7, from a mixed signal of voice of a user from a microphone 3 and audio sound, in a subtraction unit 4. In a system in which the error signal is input to a voice recognition engine, μ is decreased gradually, from a predetermined period later than voice input switch pressing, and μ is kept at a lowest value during utterance period. Voice intensity calculation of the error signal is performed, and after a predetermined period from the end of a period when a voice intensity is more than a predetermined threshold, μ is gradually increased and returned to the original value. The predetermined period and threshold which serve as reference for determining the presence of utterance, are switched over and set by personal information, and can be changed gradually by learning with voice recognition processing. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

    The present invention relates to a speech recognition device that increases a speech recognition rate by erasing audio sound input from a microphone in the speech recognition device.

  In recent years, the operation of various devices is instructed by voice, and this is recognized by a voice recognition device and the operation of the device is controlled widely in various fields such as personal computers and general household devices. The research and development is progressing rapidly. As one of the fields for controlling the operation of a device by such voice, attention has been paid to operating various on-vehicle devices by voice. That is, many of the in-vehicle devices are often operated by the driver, and on the other hand, it is not preferable for the driver to pay attention to the operation of the in-vehicle devices as much as possible for safe driving.

  In recent years, in-vehicle devices are increasingly giving various operation instructions to these devices because of advanced audio devices and diversified functions of navigation devices. As a countermeasure, the driver uses the voice recognition device, and the driver keeps his eyes on the front. For example, the navigation device instructs a nearby facility search by voice, and the navigation device displays the search result on the screen. A system that responds has been put into practical use.

  However, when the voice recognition device is mounted on a vehicle in order to control the in-vehicle device as described above, the engine sound, the tire running sound, the car wind noise, the audio sound and the voice of the surrounding people are installed in the vehicle. It is extremely difficult to recognize the contents of the operation instruction based on words spoken to the microphone in such noise. Accordingly, in the technical field of speech recognition that is widely researched and developed, speech recognition for operating instructions of in-vehicle devices can be said to be one of the most difficult fields. In order to perform speech recognition in such a noisy environment, it is necessary to remove the noise component mixed in with the speech input from the microphone and input only the user's speech as much as possible. Become.

  As a technique for that purpose, research is being conducted to obtain a sound having a desired characteristic by performing various kinds of processing on noise and voice through an adaptive filter. The control method using the adaptive filter is a well-known technique. For example, as shown in FIG. 4, the FIR filter (finite impulse response filter) in which the tap coefficient w (n) is variable for the second signal input x (n). 22 to obtain the output y (n). The output y (n) and the first signal input d (n) as the target signal are input to the subtracter 23 to obtain the error e (n). The tap coefficient w (n) of the FIR filter 22 is controlled by an adaptive algorithm (for example, LMS) 24 that changes according to the error e (n), and the power of the error e (n) is made as close to 0 as possible. Various adaptive algorithms used in this adaptive filter have been proposed. For example, a learning identification method, an LMS method, an RMS method, and a projection method are known. By using such an adaptive filter, the filter coefficient can be sequentially rewritten from an arbitrary initial state, and gradually approached to the tap coefficient w0 that minimizes the error.

In the adaptive filter that updates the tap coefficient in real time using, for example, the LMS algorithm,
wj (n + 1) = wj (n) + 2μ (n) · e (n) · xj (n)
j = 0, 1,..., N
e (n) = d (n) -y (n)
The update formula is used.

  Here, μ is called a step size parameter, and is a parameter that controls the degree of update of the tap coefficient of the adaptive filter. If this value is large, the correction amount of the tap coefficient increases, so that the convergence is accelerated. However, if there is a component that interferes with coefficient updating by the amount of correction, the effect is strongly influenced and the residual error amount increases. On the other hand, when the step size parameter is small, convergence is delayed, but the influence of the disturbing signal component is small and the residual error amount is small.

  On the other hand, for example, when a voice recognition device is used in a vehicle interior, there is a sound from the audio device as the loudest sound that disturbs the voice recognition most in the vehicle interior. Therefore, at the time of voice recognition, the sound of this audio device may be turned off. preferable. However, it is troublesome to turn off the audio device each time an instruction is given by voice. For example, when an operation instruction such as changing the volume is given to the audio device while using the audio, the audio device It is not appropriate to mute the sound. As a countermeasure, in the speech recognition apparatus, in order to cancel the audio sound coming from the microphone, the audio signal output from the speaker is directly input, and the audio signal is input to the adaptive filter and output from the adaptive filter. The audio adjustment signal and the audio signal input together with the audio signal from the microphone are input to the subtractor, and the adaptive filter is adjusted so that the error is minimized or in a predetermined state, and the subtraction is thereby performed. It is considered to prevent the audio signal from remaining in the output signal from the device.

As shown in FIG. 5, the basic configuration of such an audio cancellation system includes the configuration of the adaptive filter shown in FIG. 4. In particular, in this system, the tap coefficient w (n) is controlled by the LMS algorithm 24. As a second input that is an input signal to the FIR filter 22, an audio output unit that outputs to a speaker 26 in the vehicle interior is connected as a reference signal generator 27, and as a second input to the subtractor 23 The signal from the microphone 28 for the voice recognition device provided in the vehicle interior is made to correspond, and the signal from the microphone 28 is output to the subtracter 23 via the delay circuit 29. At this time, from the microphone 28, the voice CsXs to be recognized from the user 30 and the audio sound CnXn that becomes noise to be canceled while the voice recognition apparatus is operating are input. As shown by a two-dot chain line in FIG. 5, an error signal e (n) in the subtracter 23 is input to the LMS algorithm 24 as in FIG. 4, and this system outputs the signal to the speech recognition engine 36 as it is. 4 is used as it is. Here ^{y (n) = w (n} ) T · x (n) ·· (1)
e (n) = d (n) -y (n) (2)
w (n + 1) = w (n) + 2μ (n) · e (n) · x (n)
(3)
w (n) = [w (0, n) w (1, n)... w (N-1, n)] ^T (4)
x (n) = [x (n) x (n−1)... x (n−N + 1)] ^T (5)

Each of the following equations holds, and in particular, the update is performed by the adaptive update equation (3).

  In such a system, when the user 30 is listening to the audio from the speaker 26 in the vehicle interior and makes a sound to the microphone 28 in order to use the speech recognition device, the microphone 28 has a particularly loud sound in the vehicle interior. Sound from the audio is also input. These signals such as sound input from the microphone 28 are input as d (n) to the plus side of the subtractor 23 via the delay circuit 29. On the other hand, the signal of the audio output unit outputting the audio signal to the speaker 26 is input to the FIR filter 22 as the reference signal x (n), and the tap coefficient w (n) is controlled by the LMS algorithm 24 in the FIR filter 22. , An output signal y (n) is obtained.

  This output signal y (n) is input to the minus side of the subtracter to obtain the subtraction value of both, that is, the error between them, e (n) = d (n) −y (n). The error e (n) is ideally obtained by canceling the audio sound input from the speaker in the vehicle interior to the microphone by the audio signal processed by the adaptive filter. Therefore, when this is input to the voice recognition engine 36, the audio signal in the passenger compartment is canceled and the signal is almost only the voice of the user. At this time, if an error occurs between the two, the error e (n) is fed back to the LMS algorithm 24, and the tap coefficient w (n) of the FIR filter 22 is adjusted to obtain the error e (n). Control to minimize the power.

  As described above, when the audio device is operated in the passenger compartment and the sound is output from the speaker, various devices are operated by the voice recognition device. In order to cancel the mixed audio sound, an audio sound cancellation (ASC) system using an adaptive filter has been developed. In this system, the transfer function from each speaker to the microphone position is estimated, and the audio sound at the microphone position is simulated, and only this audio signal is subtracted from the microphone input signal in the speech recognition processing under music playback. Thus, it is possible to leave only the uttered voice, and as a result, it is possible to perform voice recognition under music playback.

  For estimation of the transfer function in this system, FIG. 6 shows an example of the output result in the current system as described above using a normalized LMS algorithm with an audio signal as a reference. Here, the step size parameter μ (n) is used to adjust the degree of coefficient update. If the step size parameter μ (n) is large, the follow-up speed is improved. , An echoed waveform is output, and as a result, speech recognition fails. Thus, the followability and the disturbance tolerance are in a trade-off relationship.

Therefore, the present applicant proposes and discloses providing a step size parameter changing unit 31 as shown in FIG. 5 in Japanese Patent Laid-Open No. 2001-195085. That is, the normal value of μ is adaptively operated with a larger value μ1 that satisfies the stability condition, and the pressing information of the voice input switch 35 that the user presses to start the voice recognition operation is input, and the voice recognition processing is performed. The value of μ is switched to μ2, which is smaller than μ1, until the end or until a predetermined time has elapsed. By doing this, it is possible to extract only the speech that is not echoed only in the section in which data is input to the speech recognition engine 36, and to perform appropriate speech recognition. Note that after the value of the step size parameter μ was switched to μ2 in response to the pressing of the voice input switch 35, the voice recognition response was performed after a preset time without waiting for the voice input switch to be pressed. It is sometimes considered to restore the original value of μ1.
JP 2001-195085 A

  FIG. 6 shows the result of the speech recognition processing in which the above processing has been performed. As can be seen from this example, the step size parameter is decreased from μ1 to μ2 in order to perform the above-described audio sound removal processing from the pressing of the voice input switch in FIG. 6C, and then after a predetermined time set in advance. Or, when it is set to return to the original μ1 when a voice recognition response is made, as shown in FIG. The utterance will be started later. After that, speech is performed, and in this speech recognition apparatus, for example, there are cases where one word is inputted and cases where a relatively long sentence is inputted. Therefore, the speech signal inputted depending on each case at the time of speech. Therefore, the user waits for a predetermined time regardless of the user's utterance.

  When the audio signal from the microphone as shown in FIG. 6A exists, the processed signal as shown in FIG. 6B can be obtained by performing the audio sound removal process as described above. If speech recognition is performed based on a completed signal, the speech recognition rate will be much higher than that without such audio sound removal processing. Often fail.

  When the cause is examined, the main cause is that the audio sound unremaining component as shown in FIG. 6B in the time from when the user presses the voice input switch to the utterance and for a while after the actual utterance ends. It has been found that this is because it has a negative effect on speech segment detection and pattern matching processing of the speech recognition engine.

  Even when the speech recognition is successful, the response time is longer than that during normal use when music is not being played back, which hinders the usability of the audio cancellation (ASC) system. The shorter the degree of inhibition, the more conspicuous the difference is, and the delay is often about 2-3 seconds. As a countermeasure, there is a method for improving the processing inside the voice recognition engine, but the company that designs and manufactures the voice recognition engine is not disclosed, and this voice recognition engine is used for example as a navigation device. For those who are trying to apply to this, this part is a black box, so it can not be messed up, and there is no other way to deal with it.

  Therefore, the present invention is a conventional audio sound cancellation system that uses an LMS algorithm with an audio signal as a reference signal in order to reliably perform speech recognition in an environment in which audio sound is input in addition to the user's speech. Voice recognition that can almost eliminate the audio sound that could not be erased by the system of, and can improve the voice recognition rate and also improve the voice recognition processing speed The main purpose is to provide a device.

  In order to solve the above problems, a speech recognition engine according to the present invention inputs a microphone that collects a user's voice and audio sound input to the speech recognition device, and an audio signal that outputs the audio sound, An adaptive filter that changes a tap coefficient by an adaptive algorithm using a size parameter, a subtractor that inputs an output signal of the adaptive filter and a signal from the microphone, and an error signal of both signals output from the subtractor. In a speech recognition apparatus that is input to an algorithm and output to a speech recognition engine, speech intensity calculation means for calculating the speech intensity of the error signal, and a threshold value in which the speech intensity calculated by the speech intensity calculation means is set in advance An utterance presence / absence determining means for obtaining a predetermined time set in advance from the time point when the above is switched to the following, When it is determined that there is no speech in the speech presence determining means, characterized in that a step size parameter adjusting means for gradually increasing the step size parameter that was reduced in advance.

  In addition, in another speech recognition apparatus according to the present invention, in the speech recognition apparatus, the step size parameter adjustment unit gradually increases the step size parameter after a predetermined period after the user presses the speech input switch. It is characterized by decreasing.

  Further, another speech recognition apparatus according to the present invention is characterized in that in the speech recognition apparatus, the speech intensity threshold value is learned and changed according to a speech recognition processing result.

  In addition, another speech recognition apparatus according to the present invention, in the speech recognition apparatus, learns a preset predetermined time from the time when the voice intensity is switched from a preset threshold value to a preset threshold value based on a voice recognition processing result. It is characterized by changing.

  Another speech recognition apparatus according to the present invention is characterized in that, in the speech recognition apparatus, a predetermined period after the speech input switch is pressed is learned and changed according to a speech recognition processing result.

  In another speech recognition apparatus according to the present invention, in the speech recognition apparatus, the utterance presence / absence determination unit changes a determination criterion according to personal information of an individual utterance information setting unit preset for each user. And

  Further, another speech recognition apparatus according to the present invention comprises an output control means for outputting zero data to the speech recognition engine when the speech recognition apparatus determines that there is no speech in the speech presence / absence determination means. Features.

  Since the present invention is configured as described above, in order to perform voice recognition reliably in an environment where audio sound is input in addition to the user's uttered voice, the audio signal that outputs the audio sound is used as a reference signal as an LMS. When deleting audio with an algorithm-based audio sound cancellation system, the audio sound that could not be erased with the conventional system can be almost erased, and the speech recognition rate can be improved. In addition, the voice recognition processing speed can be improved.

  An object of the present invention is to remove the audio sound that has been left behind in the audio cancellation system and improve the processing speed. A microphone that collects the user's voice and audio sound input to the voice recognition device, and the audio An adaptive filter for inputting an audio signal for outputting sound and changing a tap coefficient by an adaptive algorithm using a step size parameter; a subtractor for inputting an output signal of the adaptive filter and a signal from the microphone; and the subtractor In the speech recognition apparatus which inputs the error signal of both signals output from the signal to the adaptive algorithm and outputs the error signal to the speech recognition engine, speech strength calculation means for calculating the speech strength of the error signal, and the speech strength calculation The voice strength calculated by the means is switched from above the preset threshold to below. An utterance presence / absence determination unit that obtains a predetermined time from the time of departure, and a step size parameter adjustment unit that gradually increases a step size parameter that has been decreased in advance when the utterance presence / absence determination unit determines that there is no utterance And realized.

  Embodiments of the present invention will be described with reference to the drawings. FIG. 1 shows an embodiment of an audio canceling apparatus for speech recognition according to the present invention. In this embodiment, an audio output x (n) from the in-vehicle audio apparatus is output, and this output is output to the speaker 1 disposed in the vehicle interior and the same signal is also output to the adaptive filter 5. ing. In addition to the sound from the speaker 1, the utterance of the user 2 is also input to the microphone 3 when performing speech recognition processing, and the input sound signal to the microphone 3 is h (n).

  The adaptive filter 5 operates based on the basic principle of FIG. 4 and operates in the same manner as the adaptive filter 22 in the audio canceling system for a speech recognition apparatus shown in FIG. The output y (n) of the adaptive filter 5 corresponding to the audio signal is subtracted by a subtracter 4 from the signal obtained by appropriately delaying the processed signal d (n) by the audio signal h (n) input to the microphone. Thus, e (n) = d (n) −y (n) is calculated to obtain an error signal e (n). The error signal e (n) obtained in the subtracter 4 is input to the LMS algorithm 6 in the same manner as the conventional one. In the LMS algorithm 6 shown in FIG. 1, in addition to this error signal, an audio sound reference signal x ( n), it can be adjusted by a signal from a step size parameter (μ) adjusting unit 7 described later.

  In the speech recognition apparatus in FIG. 1, the speech data after the error signal e (n) from the subtracter 4 is extracted by the band pass filter 8 is processed by the speech intensity calculation unit 9. The voice intensity is calculated and the value is input to the utterance presence / absence determination unit 10. In the utterance presence / absence determination unit 10, in addition to the signal from the voice intensity calculation unit 9, a signal obtained by pressing the voice input switch in particular from the voice input switch 11 is input. Data specific to the individual of the information setting unit 18 is input.

  The utterance presence / absence determination unit 10 appropriately determines the presence / absence of utterance based on these data and signals as described later, and performs processing for setting the step size parameter to a small value μ2 in the true utterance section. The speech / non-speech signal obtained from various data and signals in the speech presence / absence determination unit 10 is used to learn speech information when setting personal-specific information in the personal speech information setting unit 18. Also used for part 19. Further, this speech / non-speech signal is also output to the output control unit 15 for operating the changeover switch 14 constituting the silence signal generation block 13, and the output control unit 15 is connected to the s1 side by the speech signal, so that the speech recognition engine. An error signal e (n) mainly including the user's speech is output to 21. Further, the non-speech signal is connected to the s2 side, and zero data is output so as not to output the speech signal to the speech recognition engine.

  The personal setting reflection unit 16 including the personal utterance information setting unit 18 and the utterance information learning unit 19 includes a personal utterance setting information storage unit 17 for storing personal utterance setting information. The personal utterance information setting unit 18 stores this information. Use while updating as appropriate. The personal setting reflection unit 16 includes the other personal identification unit 20. For example, the individual user is identified by the telephone holder, vehicle information, selection of a pre-registered user, biometric information such as a fingerprint voiceprint, and the like. The utterance information can be set appropriately.

  The speech recognition apparatus according to the present invention having the block diagram as described above can be operated in sequence according to the operation flow shown in FIG. In the following, the operation of the present invention will be described in accordance with the operation flow of FIG. 2 which mainly explains the setting of the step size parameter, which is the main function of the present invention, with reference to the functional blocks of FIG. I will explain. In the example shown in FIG. 2, the processing after the current error signal e (i) corresponding to the error signal e (n) in FIG. 1 is output as the output of the audio sound cancellation (ASC) system. First, filter processing for extracting the audio band is performed by the band pass filter 8 of FIG. 1 (step S1). When extracting the voice band, the band is switched based on individual information on whether the current user obtained in advance is male or female, and the subsequent appropriate processing is performed.

  Next, the sound intensity calculation unit 9 in FIG. 1 performs sound intensity calculation on the filtered signal to obtain the current output signal p (i) (step S2). As a signal at this time, for example, based on the error signal e (n) obtained by the ASC processing of FIG. 3A, a signal of voice intensity information as shown in FIG. 3B is obtained. This signal is used in steps S8 and S9 described below. Next, it is determined whether or not the voice input switch has been pressed (step S3). This processing is determined by inputting the signal of the voice input switch 11 in the speech presence / absence determination unit 10 in FIG. Conventionally, when this voice input switch is pressed, for example, as shown in FIG. 3 (c), the step size parameter μ is immediately reduced at that point. However, in the present invention, the subsequent processing is further performed. And set the value of μ as appropriate according to the utterance time.

  When it is determined in step S3 that the voice input switch has been pressed, various variables used in the voice recognition process are initialized (step S4). In this state, the speech input switch is turned on, and the speech flag indicating that speech is being performed is still 0. Thereafter, the process proceeds to step S5, including when it is determined in step S3 that the voice input switch has already been pressed, and it is determined whether the start count start time has been reached with the voice input switch turned on. . That is, in the present invention, since the actual utterance start time differs depending on the individuality after the voice input switch is pressed, a fade that gradually decreases the step size parameter (μ) from when the utterance start time is approached. The in operation is performed, and it is determined whether or not the start time of the start end count for performing the fade-in operation is reached in order to keep μ as small as possible during actual speech and avoid a sudden change.

  This process is performed by counting time from time t1 when the voice input switch is pressed in FIG. 3D and determining whether or not time T2 in the figure, that is, time t2, has been reached. Here, when it is determined that the voice input switch is on and the start end count start time is reached, the speech flag is set to 1, and it is determined that the current state is fading in (step S6). The starting end count value at this time can be used for learning the start time. If it is determined in step S5 that the start time has already passed rather than the start end count start time, the start end count is continued (step S7).

  After the operations of step S6 and step S7, it is determined whether or not the utterance flag is 1 and the current utterance volume p (i) is less than or equal to the utterance volume threshold (step S8). Here, when it is determined that the current utterance volume is not lower than the threshold value, for example, in FIG. 3B, it is determined that the voice intensity information as the utterance volume is larger than the threshold value p1, and as shown in step S14, Speaking. Here, the threshold of the speech volume can be learned by utilizing the processing result of the speech recognition engine.

  If it is determined in step S8 that the current utterance volume is equal to or lower than the threshold value, it is subsequently determined whether or not the previous utterance volume p (i-1) is equal to or higher than the threshold value (step S9). That is, when the voice input switch is determined to be on and the start count start time is determined in step S5, the fade-in process of μ is performed in step S6 and the speech flag is set to 1. It is determined whether or not p (i-1) that is the utterance volume value is equal to or greater than the threshold value of the utterance volume. That is, it is checked whether the utterance volume is continuously below the threshold value. Here, whether or not the utterance volume is equal to or higher than a predetermined value is determined based on the current value p based on the data as shown in FIG. This is performed by detecting whether (i) and the previous value p (i-1) are equal to or greater than a predetermined threshold value p1.

  When it is determined in step S8 that the current utterance volume is equal to or lower than the threshold value and in step S9 that the previous utterance volume is equal to or higher than the threshold value, that is, at time t5 in FIG. (Step S10). If it is determined in step S9 that the previous utterance volume is not greater than or equal to the threshold, the termination count is continued (step S11).

  After the operations of step S10 and step S11, it is determined whether the end count is equal to or longer than a predetermined interval length (step S12). This predetermined interval length is time T6 in FIG. 3D, and can be set to about 1 second, for example. This period is set in consideration of the fact that the input speech does not end with one word. Further, the end count value at this time can be used for interval length learning.

  If it is determined in step S12 that the termination count is not greater than or equal to the predetermined interval length, it is determined that speech is still being performed (step S14), and the process proceeds to the next step S15. If it is determined in step S12 that the end count has become equal to or longer than the predetermined interval length, it is determined in step S13 that the current fade-out state is present. This state is performed at time T7 after time t6 in FIG.

  In the example of FIG. 2, after performing the above operation, various processes of the step size parameter μ are performed after step S <b> 15. That is, in step S15, it is determined whether or not the current speech is being performed. If it is determined that the current speech is being performed, μ is set to the minimum value. This is a period T5 in FIG. 3D or a period including T4 and T6. If it is determined in step S15 that the current speech is not being spoken, it is determined in step S17 whether or not the current speech is fading out. If it is determined that the time corresponds to the period during the fade-out, the fade-out process of μ is performed in step S18. This process is a period T7 in FIG. At this time, the fade-out process is gradually performed by linear interpolation in a predetermined section.

  If it is determined in step S15 that it is not currently speaking, and it is determined in step S17 that it is not currently fading out, it is determined in step S19 whether or not it is a time corresponding to the current fade-in period. Here, when it is determined that the current time corresponds to the period during the fade-in, the fade-in process of μ is performed (step S20). This process is a period T3 in FIG. In this process, as in the fade-out process, the fade-in process is gradually performed by linear interpolation of the determined section. In step S19, when it is determined that the current fade-in is not being performed, that is, it is determined in step S15 that it is not currently speaking, it is determined in step S17 that it is not currently fade-out, and it is determined in step S19 that it is not currently fade-in. Sometimes, the value of μ is set to a predetermined maximum value by not performing the voice recognition process (step S21).

  After performing the various processes as described above, in step S22, the output to the speech recognition engine is obtained by multiplying the current error signal e (i) by the utterance flag, that is, when the utterance flag is 0, the output is 0. When the utterance flag exists, the error signal after the above processing is output to the speech recognition engine. This processing is performed in the no-signal generating unit 13 in FIG. After the above processing, filter count update processing is performed, and thereafter the same operation is repeated.

  In the processing as described above, particularly as shown in the personal setting reflection unit 16 of FIG. 1, the utterance information of various setting times and setting values in each processing is gradually learned, and appropriate setting values corresponding to each individual are set. It can be. That is, the personal setting reflection unit 16 is customized by learning on the assumption that the utterance start time, the utterance interval, and the utterance volume, which are parameters used in the present invention, are substantially the same for each individual. The performance is further improved by reflecting the set value. Specifically, the parameter a is loaded from a personal setting database registered in advance using the biometric information such as the telephone holder, vehicle information, and voiceprint as a key. The utterance information learning unit calculates parameter information t when the recognition engine is actually used, and updates the information by an update formula of a ′ = a + k (at). Here, k is a minute coefficient reflecting the learning result. Here, when the user switches, the parameter a 'is stored in the database and the parameter b is newly loaded.

It is a functional block diagram of the Example of this invention. It is an operation | movement flowchart of the Example of this invention. It is a figure explaining the signal processing of this invention. It is a figure explaining the action | operation of the adaptive filter used by this invention. It is a functional block diagram of the conventional speech recognition apparatus. It is a figure explaining the signal processing in the conventional apparatus.

Explanation of symbols

1 Speaker
2 User 3 Microphone 4 Subtractor 5 Adaptive filter 6 LMS algorithm 7 Step size parameter (μ) adjustment unit 8 Bandpass filter 9 Voice intensity calculation unit 10 Speech presence / absence determination unit 11 Voice input switch 12 Timer 13 Silent signal generation unit 14 Switching Switch 15 Output control unit 16 Personal setting reflection unit 17 Personal utterance setting information storage unit 18 Personal utterance information setting unit 19 Utterance information learning unit 20 Personal identification unit 21 Voice recognition engine

Claims (7)

A microphone that collects the user's voice and audio input to the voice recognition device;
An adaptive filter that inputs an audio signal that outputs the audio sound and changes a tap coefficient by an adaptive algorithm using a step size parameter;
A subtractor for inputting the output signal of the adaptive filter and the signal from the microphone;
In the speech recognition apparatus, the error signal of both signals output from the subtracter is input to the adaptive algorithm and output to the speech recognition engine.
Voice intensity calculation means for calculating the voice intensity of the error signal;
Utterance presence / absence determining means for obtaining a predetermined time from the time when the voice intensity calculated by the voice intensity calculating means is switched from a threshold value to a preset threshold value to
A speech recognition apparatus comprising: a step size parameter adjusting unit that gradually increases a step size parameter that has been decreased in advance when the speech presence / absence determining unit determines that there is no speech.   2. The speech recognition apparatus according to claim 1, wherein the step size parameter adjusting means gradually decreases the step size parameter after a predetermined period after the user presses the voice input switch.   The speech recognition apparatus according to claim 1, wherein the speech intensity threshold is learned and changed according to a speech recognition processing result.   2. The voice recognition apparatus according to claim 1, wherein a predetermined time set in advance from a time when the voice intensity is switched from a preset threshold value to a preset threshold value is learned and changed based on a voice recognition process result.   3. The voice recognition apparatus according to claim 2, wherein a predetermined period after the voice input switch is pressed is learned and changed according to a voice recognition processing result.   2. The speech recognition apparatus according to claim 1, wherein the utterance presence / absence determination unit changes the determination criterion according to personal information of a personal utterance information setting unit preset for each user. 2. The speech recognition engine according to claim 1, further comprising output control means for outputting zero data to the speech recognition engine when the speech presence / absence determining means determines that there is no speech.

Images