JP2008020872A - Voice recognition device for vehicle and navigation device for vehicle - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition device for vehicles, capable of improving a recognition rate of voice recognition processing performed in a vehicle. <P>SOLUTION: In a vehicle room, sound collecting microphones 4 to 7 are provided at a steering wheel 11, at a dashboard 5, in front of a passenger seat side speaker 14, and at the backside of the passenger seat 13. A voice separation processor 2 separates a voice signal which is input by the sound collecting microphones 4 to 7 into a driver voice signal S and the other noise signals N1 to N3 by a BSS method. A voice recognition processor 3 of the navigation device 1 for vehicles performs voice recognition processing on the separated driver voice signal. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a voice recognition device for a vehicle that performs voice recognition processing by collecting voice generated by a vehicle occupant in a vehicle interior using a sound collecting microphone, and a vehicle navigation device including the device.

  As a result of the improvement of the voice recognition processing function, recent vehicle navigation apparatuses are increasing in number that can be operated by voice emitted by a driver. For example, as a related art regarding the voice recognition processing, there are those disclosed in Patent Documents 1 to 5.

However, since the interior of the vehicle is an environment where a lot of noise is generated, such as the engine sound of the vehicle and the sound output from the speaker of the car audio device, further improvements can be made to improve the voice recognition rate. preferable. Among the prior arts described above, the blind source separation (BSS) method disclosed in Patent Documents 4 and 5 includes a speech signal to be recognized and other noise signals, even if there is no section in which only noise is generated. Can be clearly recognized.
JP 2004-069772 A JP 2004-309536 A JP 2004-317942 A JP 2002-023776 A JP 2005-227512 A

In addition, Patent Document 5 suggests that the BSS method is applied to the voice recognition in the vehicle interior. It is not disclosed at all.
The present invention has been made in view of the above circumstances, and an object of the present invention is to provide a vehicle speech recognition device capable of further improving the recognition rate of speech recognition processing performed in the vehicle interior, and the vehicle speech recognition device. It is providing the vehicle navigation apparatus comprised.

  According to the vehicle voice recognition apparatus of claim 1, the sound collecting microphone is installed at a position for collecting at least the driver sound, the engine sound, and the speaker output sound, and the sound input by these sound collecting microphones. The signal is separated into a driver voice signal and other noise signals by a blind source separation method, and voice recognition processing is performed on the separated driver voice signal. That is, as the sound source other than the sound emitted by the driver in the passenger compartment, the engine sound, the sound emitted by the car audio, etc. are mainly used. If the method is applied, it is possible to effectively separate noise other than the voice emitted by the driver.

According to the vehicle voice recognition device of claim 2, the signal y separated by the blind source separation method is subjected to nonlinear modeling of a mixing process when the signal source s is collected by the sound collecting microphone. Set as y = hs + ks ² . And when ks ² + hs−y = 0 is solved for s from the quadratic equation solution formula,
s = −h / (2k) ± (h ² / (4k ² ) + y / k) ^1/2
Is obtained. The s as a nonlinear function G, calculation z = G (y) = - α / 2 ± (α 2/4 + y / β) 1/2
(Α = h / k, β = 1 / k)
To obtain a linearized signal z. That is, when the signal source s emits an acoustic signal having periodicity such as a speaker, the mixing process is nonlinear. Therefore, if the coefficients α and β are determined so as to optimize the nonlinear function G, the separated signal y is linearized using the nonlinear function G, so that the driver voice signal and other noise signals are increased. It becomes possible to separate with accuracy.

  According to the vehicle voice recognition apparatus of the third aspect, the plurality of sound collecting microphones are respectively installed in front of the steering wheel, the dashboard, and the front passenger side speaker. That is, the above arrangement is suitable for collecting the driver's voice, engine sound, and car audio sound, respectively, and can reliably capture main noise in the passenger compartment.

According to the vehicle voice recognition apparatus of the fourth aspect, the sound collecting microphone is also installed at a position where the passenger's voice of the rear seat is collected. That is, when the vehicle has a rear seat and an occupant is present in the seat, it is also assumed that the occupant emits sound. Therefore, the driver's voice can be separated more favorably by collecting the passenger's voice in the rear seat.
According to the vehicle navigation device of the fifth aspect, in the case of the fourth aspect, since the sound collecting microphone is installed on the rear side of the passenger seat, the passenger's voice in the rear seat can be collected well.

  According to the vehicle voice recognition apparatus of the sixth aspect, the separated noise signal is subtracted from the signals input from the plurality of sound collecting microphones. In other words, the driver's voice signal obtained by separation by the BSS method also contains a slight amount of noise signal. Therefore, the separated noise signal is fed back to the input side and subtracted from the signal input from the sound collecting microphone. Accordingly, the level of the input noise signal can be reduced, and a clearer driver voice signal can be obtained.

  According to the vehicle navigation device of the seventh aspect, the vehicle voice recognition device according to any one of the first to sixth aspects is provided, and more accurate based on the voice recognition processing result for the separated driver voice signal. It becomes possible to perform high operation control.

(First embodiment)
Hereinafter, a first embodiment in which the present invention is applied to a vehicle navigation apparatus will be described with reference to FIGS. FIG. 1 is a functional block diagram schematically showing the configuration of the vehicle navigation device. A vehicle navigation device (vehicle speech recognition device) 1 includes a speech separation processing unit 2 and a speech recognition processing unit 3. The voice separation processing unit 2 separates a voice signal input from the four sound collecting microphones 4 to 7 arranged in the passenger compartment into a voice signal emitted by a driver (driver) and other noise signals, The former voice signal is output to the voice recognition processing unit 3. Then, the voice recognition processing unit 3 performs voice recognition processing on the given driver voice signal.
A voice recognition result by the voice recognition processing unit 3 is output to the navigation operation control circuit 9, and the navigation operation control circuit 9 performs operation control of the navigation device 1, for example, setting of a destination, based on the voice recognition result. It is like that.

In FIG. 2, the arrangement | positioning state of the sound collection microphones 4-7 in a vehicle interior is shown. The sound collecting microphone 4 is disposed on the handle 11 for collecting the sound emitted by the driver, and the sound collecting microphone 5 is disposed on the front dashboard 12 for collecting the engine sound of the vehicle. Yes. The sound collecting microphone 6 is disposed in front of the door speaker 13 on the passenger side 13 to collect sound output by the car audio system, and the sound collecting microphone 7 collects sound emitted by the passenger in the rear seat. In order to make a sound, it is arranged behind the passenger seat 14.
As shown in FIG. 1, the sounds collected by the sound collection microphones 4 to 7 are different in the ratio of each level, but the driver sound and other noises such as engine sound, speaker sound, The sound of the rear seat occupant is mixed.

  Next, the operation of this embodiment will be described with reference to FIGS. FIG. 3 is a flowchart showing the processing by the navigation device 1 with respect to the portion according to the gist of the present invention. When the voice separation processing unit 2 starts receiving an input of the voice signal collected by the voice collecting microphones 4 to 7 (step S1), it waits until the input is completed (step S2). Here, signals collected by the respective sound collecting microphones 4 to 7 are assumed to be x1 (i) to x4 (i). The variable i (= 1 to n) indicates the number of samples of the input signal. When the number of samples reaches n, the input is completed (“YES”).

  In a succeeding step S3, it is determined whether or not the input signal includes an audio signal uttered by the driver. For example, when the input level of the sound collecting microphone 4 is somewhat higher than the input levels of the other sound collecting microphones 5 to 7, it can be determined that the input signal includes the sound uttered by the driver. ("YES"). Then, it transfers to step S4 and performs an audio | voice separation process.

  In the audio separation process, the details will be described later, but the audio signal uttered by the driver and other noises are separated from the signals input to the sound collecting microphones 4 to 7 by the BSS (Blind Source Separation) method. Then, only the separated speech signal is output to the speech recognition processing unit 3 to execute speech recognition processing (step S5). When the voice recognition can be normally performed (step S6: “YES”), the navigation operation control circuit 9 performs the operation control of the navigation device 1 corresponding to the recognition result (step S7).

  FIG. 4 is a flowchart showing the contents of the voice separation process in step S4. FIG. 5 conceptually illustrates voice separation processing by the BSS method. When sound signals (Si) output from a plurality of signal sources S1 to Sn are collected by the same number of sound collecting microphones X1 to Xn, the signals (Xi) input to the sound collecting microphones are sound signals (Si). And a linear sum (X = AS: vector expression) of the mixing coefficient Aij according to the transmission environment of the audio signal. The output signal (Yi) obtained by separating the audio signals (Si) corresponding to the signal sources S1 to Sn from the signal (Xi) input to each sound collecting microphone is also separated from the input signal (Xi). It becomes a linear sum (Y = WX) with the coefficient Wij.

If the above two expressions are combined, Y = WAS, and if T = WA, Y = TS. When this matrix T becomes a unit matrix as shown in FIG. 5, it can be said that each signal was completely separated. In the BSS, a natural gradient method is used to perform the separation in this way. The natural gradient method is an algorithm that learns to minimize the mutual information amount, and the learning update formula for obtaining the separation coefficient W is:
W (t + 1) = W (t) + η [Λ (t) −f (Y (t)) Y ^T (t)] W (t)
... (1)
Where η: learning rate, Λ (t): diagonal matrix,
f (Y) = − dp (Y) / dY / p (Y), p (Y) is a probability density function of the output signal Y. T is a variable indicating a time series of processing. If the BSS method is used, the signal can be separated even if the mixing coefficient Aij is unknown.

  In FIG. 4, the speech separation processing unit 2 first initializes a variable i to “0” (step S11), and increments the variable i (step S12). Until the variable i reaches [n] indicating the total number of samples (step S13: “NO”), the subsequent processing is repeatedly executed.

  In step S14, the speech separation processing unit 2 subtracts the sum of the noise components N1 (i-1) to N3 (i-1) separated in the previous process from the input signals x1 (i) to x4 (i). X'1 (i) to x4 '(i) are obtained. This process is described in FIG. That is, the audio signal obtained by separation by the BSS method contains a small amount of noise components. Therefore, the noise level in the output signal is further reduced by performing a sound separation process on the separated noise component fed back on the input side and subtracted from the input signal.

  Reference is again made to FIG. In subsequent step S15, the sound separation processing unit 2 obtains a separated output signal Y (i) by multiplying the input signal X '(i) by the separation coefficient W (i) obtained in the previous process. In addition, what was written with the capital letter of an alphabet shall show a matrix. Then, the above equation (1) is calculated to obtain the separation coefficient W (i + 1) used for the next processing (step S16). In step S16, the variable t in equation (1) is replaced with i.

  In subsequent steps S17 to S20, a voice specifying process is performed. In other words, in the BSS method, the separated output signal obtained as shown in FIG. 5 cannot be specified for which output the driver's audio signal is obtained. That is, in the configuration of FIG. 1, four outputs from which the sound source is separated are obtained corresponding to the four inputs (sound collecting microphones 4 to 7), but it is not known which of the four is the driver sound. . Therefore, the technique disclosed in Patent Document 4 is used to specify which separated output signal is the audio signal of the driver. Therefore, the variable i is replaced with the variable j, and the voice specifying process is executed in step S18 until the variable j becomes the current value of the variable i.

The speech specifying process will be briefly described. The kurtosis of the probability distribution is calculated and compared for each separated output signal. The signal having the highest kurtosis is the driver's voice signal S (i), and the other signals are noises N1 (i) to N3 (i) (step S21). Since the noises N1 (i) to N3 (i) are used for subtracting the sum in step S14 as described above, it is not necessary to dare to specify the generated sound source. When the audio signal is specified, the process proceeds to step S12, the variable i is incremented, and the process is continued.
When the voice separation processing unit 2 obtains the voice signal S (i) issued by the driver as described above, the voice separation processing unit 2 outputs the voice signal S (i) to the voice recognition processing unit 3 to execute the voice recognition processing in step S6.

As described above, according to the present embodiment, in the vehicle interior, the sound collection microphones 4 to 7 collect the sound emitted by the driver, the engine sound, the speaker output sound, and the sound emitted by the passenger in the rear seat, respectively. The steering wheel 11, the dashboard 5, the front passenger side speaker 14 and the rear side of the passenger seat 13 are installed. Then, the voice separation processing unit 2 separates the voice signal input from the sound collecting microphones 4 to 7 into the driver voice signal S and the other noise signals N1 to N3 by the BSS method, and The voice recognition processing unit 3 performs voice recognition processing on the separated driver voice signal.
That is, as the sound source other than the sound emitted by the driver in the vehicle interior, mainly the engine sound, the sound emitted by the car audio, the sound emitted by the occupant of the rear seat, etc., so that at least the sound emitted by those sound sources is collected. If the sound collecting microphones 4 to 7 are disposed in the BSS method and the BSS method is applied, noise can be effectively separated.

  Further, since the sound separation processing unit 2 subtracts the separated noise signal from the signal input from the sound collection microphones 4 to 7, the level of the noise signal included in the signal input from the sound collection microphones 4 to 7 is set. Therefore, a clearer driver voice signal can be obtained. Then, the navigation device 1 can perform operation control with higher accuracy based on the voice recognition processing result of the separated driver voice signal.

(Second embodiment)
7 to 9 show a second embodiment of the present invention. The same parts as those of the first embodiment are denoted by the same reference numerals and the description thereof will be omitted. Hereinafter, different parts will be described. In the first embodiment, it is assumed that the process (mixing process) in which the audio signal and the noise signal are mixed and input to the microphone is linear. That is,
xi = ai1.s1 + ai2.s2 + (2)
(I = 1 to 4).

However, if the sound source that generates noise, such as a speaker, generates an acoustic signal with periodicity such as a speaker, the mixing process may be nonlinear, and the signal input to the microphone is nonlinear. Will also be included. That is,
xi = ai1 · s1 + ai2 · s2 + ai1 2 · s1 2 + ai2 2 · s2 2 ···
... (3)
Model. Actually, the third and higher terms are extremely small in value and can be ignored, so there is no problem. Therefore, the nonlinear function Fi (u) acting on the signal u mixed with the coefficient a by each signal source s is expressed as follows:
Fi (ui) = ai · ui + bi · ui ² (4)
Assuming that Note that the coefficient ai in the equation (4) is not a mixing coefficient shown in FIG. 7, but a coefficient for describing the nonlinear function Fi (ui).
Therefore, in the second embodiment, the signal Yi is linearized by applying a nonlinear function Gi to the signal Yi separated by using the BSS method as in the first embodiment to obtain the signal Zi. .

FIG. 7 is a diagram corresponding to FIG. 5 modeled including the linearization process when the mixing process is nonlinear. However, in order to further simplify the nonlinear function Gi, the signal source is divided into a speech signal and other noises that are combined into one, and S1 and S2 are obtained (therefore, the signal S, the function). I for G and signal Z is i = 1, 2). For example, if Y1 is the audio signal S in the first embodiment, Y2 is the sum of all noises 1-3, ie,
Y2 = N1 + N2 + N3 (5)
It has become.

FIG. 8 is a diagram corresponding to FIG. 3, but step 8 of “linearization processing” is inserted between steps S4 and S5.
Hereinafter, the linearization process performed in step S8 will be described in more detail. Each signal is written in lower case. The mixed signal xi obtained through the nonlinear function Fi (ui) in the mixing process is
xi = ci · s1 + di · s2 + ei · s1 ² + fi · s1 · s2 + gi · s2 ²
... (6)
It becomes. Note that c, d, e, f, and g are coefficients of each term.

It can be generally said that two independent signals (s1, s2) are independent even if their higher-order terms (s1 ² , s2 ² ) are included. Therefore, the signal yi obtained by separating the mixed signal xi into an independent form is
y1 = h1 · s1 + k1 · s1 ² (7)
y2 = h2 · s2 + k2 · s2 ² (8)
It becomes. However, h and k are arbitrary provisional coefficients. Here, for example, when the equation (7) is derived, it is necessary to eliminate the three terms s2, s1, s2, and s2 ² from the equation (6), but since there are four sound collecting microphones, four simultaneous equations are required. Therefore, unnecessary three terms can be deleted based on them.

And the quadratic equation of si ki · si ² + hi · si−yi = 0 (9)
From the solution formula, the original signal si is
si = −hi / (2ki) ± (hi ² / (4ki ² ) + yi / ki) ^1/2
... (10)
Is obtained. The (10) as a nonlinear function Gi (yi), and operation zi = Gi (yi) = - αi / 2 ± (αi 2/4 + yi / βi) 1/2 ··· (11)
(Αi = hi / ki, βi = 1 / ki)
By performing the above, a linearized signal zi is obtained. That is, the nonlinear function Gi is an inverse function of the nonlinear function Fi.

また、係数α，βの更新式は、以下のようになる。

Here, the signal yi separated by the BSS method includes a first-order term si and a second-order term si ^{2 as} shown in equations (7) to (9). In this case, the average of the primary term is “0”, and the average of the secondary term has a certain value. Accordingly, when the average value of the separation signal yi becomes “0”, the second-order term also becomes “0”. Therefore, the second-order term can be eliminated and only the first-order term can be extracted. Therefore, an average value of the final output signal zi is set as an error function Ei (t). Note that t is a variable indicating a time series of calculation.

The update formulas for the coefficients α and β are as follows.

  In the equations (13) and (14), η is a learning rate (step function) and sets the degree of updating. When the learning rate η is small, learning (update) takes time, but the optimum value can be surely reached. On the other hand, when the learning rate η is large, learning (updating) does not take much time, but it diverges in the vicinity of the optimum value and may not reach the optimum value. M is the number of samples used in one calculation, and is about 500 to 1000, for example.

  FIG. 9 is a flowchart showing details of the “linearization process” in step S8 of FIG. That is, an error function Ei (t) that is an average of the output signals zi (t) is calculated (step S31), and it is determined whether or not the result is an optimum value (= 0) (step S32). If it is not the optimum value (“NO”), the coefficients α and β are updated by the equations (13) and (14) (step S33), and the process returns to step S31 to calculate the error function Ei (t) again.

As a result of repeating the above processing, when Ei (t) = 0 in Step S32 (“YES”), the coefficients α and β are determined at that time, and the nonlinear function Gi (yi) is obtained (Step S34). Then, by calculating the nonlinear function Gi (yi), the output signal zi in the form in which the primary term si (original signal) is multiplied by, for example, a predetermined coefficient q without including the secondary term si ² in the separated signal yi. That is,
zi = Gi (yi) = q · si (17)
Can be obtained (step S35).

As described above, according to the second embodiment, the signal y separated by the BSS method is converted into a non-linear model by mixing the signal source s when the signal source s is collected by the sound collecting microphones 4 to 7. hs + ks ² is set, the nonlinear function Gi is determined as shown in the equation (10) based on the model, the coefficients α and β are optimized by learning, and the signal zi obtained by linearizing the separated signal y is obtained. did. Therefore, the driver's voice signal and other noise signals can be separated from each other with higher accuracy.

The present invention is not limited to the embodiments described above or shown in the drawings, and the following modifications are possible.
The subtraction processing of the noises N1 (i) to N3 (i) in step S14 may be performed as necessary. In that case, x′1 (i) to x4 ′ (i) may be replaced with x1 (i) to x4 (i).
The arrangement of the sound collection microphones 4 to 7 is an example, and if the driver sound, engine sound, speaker output sound, and rear seat occupant sound can be collected satisfactorily, the sound collecting microphones 4 to 7 are appropriately changed according to the structure of the different passenger compartment. Just do it.

The sound collection microphone 7 for collecting the occupant voice of the rear seat may be arranged as necessary. For example, in the case of a vehicle with few passengers in the rear seat, the vehicle may be deleted. Also, it is not necessary for a two-theta vehicle.
Further, the number of sound collecting microphones arranged in the vehicle compartment may be five or more.
For example, the present invention may be applied to a system such as a car audio system or a car air conditioner, and the operation control thereof may be performed by voice recognition.

1 is a functional block diagram schematically showing a configuration of a navigation device according to an embodiment when the present invention is applied to a vehicle navigation device. The figure which shows the arrangement | positioning state of each sound collection microphone in a vehicle interior The flowchart which shows the process by a navigation apparatus about the part which concerns on the summary of this invention The flowchart which shows the content of the audio | voice separation process in step S4 of FIG. The figure explaining the audio | voice separation process by BSS method The figure explaining the feedback process of the separated noise signal FIG. 5 equivalent diagram showing a second embodiment of the present invention. 3 equivalent figure The flowchart which shows the detail of the "linearization process" in step S8 of FIG.

Explanation of symbols

  In the drawings, 1 is a vehicle navigation device (vehicle speech recognition device), 2 is a speech separation processing unit, 3 is a speech recognition processing unit, 4 to 7 are sound collecting microphones, and 9 is a navigation operation control circuit.

Claims (7)

In a vehicle voice recognition device that collects voice generated by a vehicle occupant in a passenger compartment with a microphone and performs voice recognition processing.
The sound collecting microphone is installed at a position for collecting at least the driver sound, engine sound, and speaker output sound, and the sound signal input by the plurality of sound collecting microphones is converted into the driver sound signal by the blind source separation method. And a noise signal, and a voice recognition process is performed on the driver voice signal. The signal y separated by the blind source separation method is set as shown in the following equation by nonlinear modeling the mixing process when the signal source s is collected by the sound collecting microphone,
y = hs + ks ² (h and k are arbitrary provisional coefficients)
The s obtained by solving the above equation, calculation z = G was used as the nonlinear function G (y) = - α / 2 ± (α 2/4 + y / β) 1/2
(Α = h / k, β = 1 / k)
The vehicle speech recognition apparatus according to claim 1, wherein a linearized signal z is obtained by performing the step.   The vehicle voice recognition apparatus according to claim 1 or 2, wherein the plurality of sound collecting microphones are respectively installed in front of a handle, a dashboard, and a passenger seat side speaker.   The vehicular voice recognition device according to any one of claims 1 to 3, wherein the sound collecting microphone is also installed at a position for collecting occupant voice in a rear seat.   The vehicle sound recognition apparatus according to claim 4, wherein the sound collecting microphone is installed on the rear side of the passenger seat.   The vehicular speech recognition apparatus according to any one of claims 1 to 5, wherein the separated noise signal is subtracted from signals input from the plurality of sound collecting microphones.   A vehicle navigation system comprising the vehicle voice recognition device according to any one of claims 1 to 6, wherein the vehicle navigation system is configured to perform operation control based on a voice recognition processing result for the driver voice signal. apparatus.

Images