JP4183338B2 - Noise reduction system - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a noise reduction system for improving the S / N ratio of a speaker voice signal during speech recognition, and more particularly to a noise reduction system suitable for use in an AMNOR (Adaptive Microphone-array for Noise Reduction) system.
[0002]
[Prior art]
The current speech recognition system has reached a technical level that can achieve a recognition rate of about 95% when a signal-to-noise ratio of 15 dB or more is secured. However, when the S / N ratio is reduced due to noise present in the surrounding area, the recognition rate is rapidly lowered. Fig. 5 shows the relationship between the SN ratio and recognition performance for several types of microphones (omnidirectional, unidirectional, narrow directivity, etc.). The SNR and recognition rate of each microphone is roughly S. It is included in the band 100 showing the character characteristics. As is apparent from FIG. 5, the recognition rate rapidly decreases due to a decrease in the SN ratio, and decreases to about 50% in an environment where the SN ratio is 0 dB.
[0003]
For this reason, the above-mentioned deterioration in recognition performance is unavoidable in automobile interiors where noise generated by automobiles (engine noise, road noise, pattern noise, wind noise, etc.) is present, and a voice recognition system is mounted on the vehicle. It is one of the big problems. In view of the circumstances described above, various methods have been proposed for receiving sound with a high S / N ratio while reducing the influence of noise present in the surroundings, and the AMNOR method is one example.
[0004]
In the AMNOR noise reduction system, a plurality of microphones are provided, the target signal is delayed by a considerable amount for each microphone, and each delayed signal is added to the corresponding microphone output signal to obtain a reference signal. The adaptive signal processing unit receives each reference signal and error signal during learning, performs adaptive signal processing to minimize the power of the error signal, updates the coefficient of the adaptive filter, and performs non-learning such as during speech recognition At this time, the coefficient update of the adaptive filter is stopped, the coefficient at the time of learning is left set in the adaptive filter, and the difference between the signal output from the target response setting unit and the signal output from the adaptive filter is output as a speech signal to the speech recognition unit To do.
[0005]
FIG. 6 is a configuration example of an AMNOR type noise reduction system using two general microphones. In the figure, 11 and 12 are first and second microphones, 13 and 14 are amplifiers, 15 is a signal generator (SG) that generates random noise such as white noise as a target signal, 16 is a gain variable amplifier, and 17 and 18 are Delay units 19 and 20 add delays corresponding to signal delay times d ₁ and d ₂ from the speaker's mouth to the respective microphones to the target signal, and add signals output from the respective delay units to the output signals of the respective microphones. It is an adder.
[0006]
Reference numeral 21 denotes a 2-input / 1-output adaptive signal processing unit, which adds and outputs the outputs of the first and second adaptive signal processing units 21a and 21b and the adaptive signal processing units 21a and 21b. have. Although not shown, each of the adaptive signal processing units 21a and 21b has an LMS calculation unit and an adaptive filter having an FIR type digital filter configuration. First signal processing unit 21a, when learning a reference signal the output signal of the adder 19 to update the adaptive filter coefficients W ₁ so that the power of the error signal e is minimized, during speech recognition, adaptive filter stop coefficient update, subjected to output the filtering process on the input signal by the coefficient W ₁ that is left to set in the adaptive filter obtained at the time of learning. Second signal processing unit 21b, during training, a reference signal the output signal of the adder 20 to update the adaptive filter coefficients W ₂ so that the power of the error signal e is minimized, during speech recognition, adaptive filter The coefficient updating is stopped, the coefficient W ₂ obtained at the time of learning is set in the adaptive filter, the input signal is filtered and output. The adder 21c adds each adaptive filter output and outputs it.
[0007]
Reference numeral 22 denotes a target response setting unit that receives a noise signal output from the signal generator 15 as a target signal, and approximates the inverse characteristic of the acoustic system with high accuracy. When the signal delay time half of the tap length of the adaptive filter is d ′ and the average value of the delay times d ₁ and d ₂ is d ″, the target response setting unit 22 delays the delay time d (= d ′ + d ″). And has a flat characteristic (gain 1 characteristic) in the audio frequency band. A subtracting unit 23 subtracts the output signal of the adaptive signal processing unit 21 from the target response output from the target response setting unit 22 and outputs an error signal e.
[0008]
During learning, only the noises Xn ₁ (z) and Xn ₂ (z) generated by the automobile CR are input to the microphones 11 and 12. The adders 19 and 20 synthesize the random noise as the target signal and the output of each microphone, and the adaptive signal processing unit 21 uses the output signal of the adders 19 and 20 as a reference signal so that the power of the error signal e is minimized. Adaptive signal processing is performed to update the adaptive filter coefficients W ₁ and W ₂ .
At the time of speech recognition, the adaptive signal processing unit 21 does not update the filter coefficients, leaves the coefficients W ₁ and W ₂ obtained at the time of learning set to each adaptive filter, and synthesizes and subtracts the output signals of these adaptive filters. Input to the unit 23. The subtracting unit 23 subtracts the output signal of the adaptive signal processing unit 21 from the target response output from the target response setting unit 22, and inputs the difference signal as a speech signal to the speech recognition unit. According to such an AMNOR type noise reduction system, noise is minimized at the time of voice recognition, and a large speaker voice output can be obtained to improve the SN ratio.
[0009]
[Problems to be solved by the invention]
(1) The AMNOR method has less complicated sound transfer characteristics, and when used in an environment where the difference in sound transfer characteristics from the speaker's mouth to each microphone can be expressed only by a delay, such as a relatively large room. It is valid. However, in an environment having a very complicated acoustic transfer characteristic such as in a passenger compartment, the difference in transfer characteristic cannot be expressed only by a delay, and is insufficient.
(2) There are many noise sources in the passenger compartment, and the noise received by each microphone has a low correlation, so the distance between the microphones must be shortened. For this reason, the distance between the speaker and the microphone is limited to a short distance, and the time difference of arrival of the voice to each microphone is very small. The delay cannot be set accurately, and the delay alone is not sufficient. For example, when the sampling frequency fs is 11.025 (kHz), one sample is
Since it corresponds to (1 / 11.025 × 10 ³ ) (sec) × 340 m / sec) = 3.08 (cm), only a delay corresponding to a distance difference of 3 (cm) or more and 3 (cm) units can be set.
(3) The AMNOR microphone array system uses a plurality of microphones, but ideally the characteristics of each microphone should be the same, but in reality the characteristics of the microphones are not the same, For this reason, the expected noise reduction effect cannot be expected. If microphones having the same characteristics are prepared, it will be costly.
[0010]
As described above, the present invention is to provide a noise reduction system that can improve the noise reduction effect.
Another object of the present invention is to provide a noise reduction system that can improve the noise reduction effect even if the characteristics of the microphones are not the same.
[0011]
[Means for Solving the Problems]
According to the present invention, (1) a plurality of microphones, (2) a signal generator for generating a random noise signal, (3) a transfer circuit for simulating transfer characteristics from the speaker's mouth to each microphone, (4) A synthesis unit that synthesizes a random noise signal output via each transmission circuit with an output signal of each microphone. (5) During learning, adaptive signal processing is performed using the random noise signal as a target signal and each synthesis unit output as a reference signal. To update the coefficient of the adaptive filter and stop updating the filter coefficient when not learning, (6) target response setting unit to give a predetermined delay to the target signal, and (7) output of the adaptive filter A noise reduction system including a subtraction unit that obtains a difference between a signal and an output signal of a target response setting unit and outputs the difference signal as an audio signal when not learning. Ri is achieved. According to the present invention, since the transfer characteristic from the speaker mouth to the microphone is simulated by the actual transfer characteristic from the speaker mouth to the microphone, the noise reduction effect is improved. Can do [0012]
Further, according to the present invention, the present invention provides a transfer circuit for measuring transfer characteristics from a speaker's mouth to each microphone output terminal in a transfer circuit, and the system is configured to transfer characteristics measurement mode, learning mode, non-learning mode. This is achieved by measuring transfer characteristics, learning filter coefficients, and outputting audio signals. In this way, the transfer characteristic from the speaker's mouth to the output end of each microphone can be measured and set in the transfer circuit, and the noise reduction effect can be improved even if the microphone characteristics are not the same.
[0013]
DETAILED DESCRIPTION OF THE INVENTION
FIG. 1 is a block diagram of a first embodiment of the present invention. In the figure, 51 and 52 are first and second microphones, 53 and 54 are amplifiers, 55 is a signal generator (SG) that generates random noise such as white noise as a target signal, 56 is a gain variable amplifier, and 57 and 58 are Simulating transfer characteristics (propagation characteristics) CS1 'and CS2' from the speaker's mouth to the output terminal of each microphone, and transfer circuits 59 and 60 apply the propagation characteristics to the target signal. Each is an adder that adds to the output signal of each microphone.
[0014]
Reference numeral 61 denotes a 2-input / 1-output adaptive signal processing unit, which includes first and second adaptive signal processing units 61a and 61b and an adder 61c that adds and outputs the outputs of the adaptive signal processing units. ing. Although not shown, each of the adaptive signal processing units 61a and 61b includes an LMS calculation unit and an adaptive filter having a FIR type digital filter configuration.
First signal processing unit 61a, when learning a reference signal the output signal of the adder 59 to update the adaptive filter coefficients W ₁ so that the power of the error signal e is minimized, nonlearning when for example during voice recognition Then, the updating of the coefficient of the adaptive filter is stopped, the coefficient W ₁ obtained at the time of learning is left set in the adaptive filter, and the input signal is filtered and output. Second signal processing unit 21b, during training, a reference signal the output signal of the adder 60 to update the adaptive filter coefficients W ₂ so that the power of the error signal e is minimized, during speech recognition, adaptive filter The coefficient updating is stopped, the coefficient W ₂ obtained at the time of learning is set in the adaptive filter, the input signal is filtered and output. The adder 61c combines and outputs the respective adaptive filter outputs.
[0015]
Reference numeral 62 denotes a target response setting unit that receives a noise signal output from the signal generator 55 as a target signal, and is used to accurately approximate the inverse characteristics of the acoustic system. If the signal delay time that is half the tap length of the adaptive filter is d, the target response setting unit 4 has a delay characteristic of the delay time d and has a flat characteristic (gain 1 characteristic) in the audio frequency band. Reference numeral 63 denotes a subtracting unit that subtracts the output signal of the adaptive signal processing unit 71 from the target response output from the target response setting unit 62 and outputs an error signal e.
[0016]
During learning, only the noises Xn ₁ (z) and Xn ₂ (z) generated by the automobile CR are input to the microphones 51 and 52. The transfer circuits 57 and 58 give transfer characteristics CS1 ′ and CS2 ′ from the speaker mouth to each microphone output terminal to random noise as a target signal. The adders 59 and 60 synthesize the outputs of the transmission circuits 57 and 58 and the outputs of the microphones, and the adaptive signal processing unit 61 uses the output signals of the adders 59 and 60 as a reference signal so that the power of the error signal e is minimized. Then, adaptive signal processing is performed to update the adaptive filter coefficients W ₁ and W ₂ .
At the time of speech recognition, the adaptive signal processing unit 61 does not update the filter coefficient, leaves the coefficients W ₁ and W ₂ obtained at the time of learning set to each adaptive filter, and synthesizes and subtracts the output signals of these adaptive filters. Input to the unit 63. The subtracting unit 63 subtracts the output signal of the adaptive signal processing unit 61 from the target response output from the target response setting unit 62, and inputs the difference signal as a speech signal to the speech recognition unit.
As described above, adaptive signal processing is performed using the signal obtained by adding the transfer characteristics CS1 ′ and CS2 ′ to the target signal and the output signal of each microphone as a reference signal, so that the power of the noise output is minimized during learning. Thus, the adaptive filter coefficient can be accurately determined. As a result, noise can be reduced during speech recognition, and a speech signal having a large SN ratio can be output.
[0017]
FIG. 2 is a block diagram of a measuring apparatus for measuring transfer characteristics CS1 ′ and CS2 ′ from the mouth of the speaker DR to the output terminals of the microphones. The same parts as those in FIG. In the figure, 70 is a speaker provided near the speaker's mouth, and outputs white noise output from the signal generator 55 toward the microphones 51 and 52.
71 is an adaptive signal processing unit that uses white noise as a reference signal and the output of the microphone 51 as a target signal, performs adaptive signal processing so as to minimize the power of the error e ₁ , and updates the coefficient Wcs _{1 of the} adaptive filter. It is a subtraction unit that outputs a difference (error) e ₁ between the output of the microphone 51 and the output of the adaptive signal processing unit 71.
73 is an adaptive signal processing unit 74 that uses white noise as a reference signal and the output of the microphone 52 as a target signal, performs adaptive signal processing so that the power of the error e ₂ is minimized, and updates the coefficient Wcs _{2 of the} adaptive filter. It is a subtraction unit that outputs a difference (error) e ₂ between the output of the microphone 52 and the output of the adaptive signal processing unit 73.
[0018]
When the adaptive signal processing units 71 and 73 continuously perform adaptive signal processing to update the coefficients Wcs ₁ and Wcs ₂ of the adaptive filter (not shown), the coefficients converge to a constant value. When the coefficient value Wcs ₁ converges to a constant value, the characteristic of the adaptive filter of the adaptive signal processing unit 71 indicates the transfer function CS 1 ′ from the speaker's mouth to the output terminal of the microphone 51. Further, when the coefficient value Wcs ₂ has converged to a constant value, the characteristics of the adaptive filter of the adaptive signal processing unit 73, showing the transfer function CS2 'until the output end of the microphone 52 from the speaker mouth.
Therefore, if the transmission circuits 57 and 58 in FIG. 1 are constituted by FIR type digital filters and Wcs ₁ and Wcs ₂ are set as the coefficients of these filters, the transmission circuits 57 and 58 output the microphones 51 and 52 from the speaker's mouth. Transfer functions CS1 'and CS2' to the end can be simulated.
[0019]
As described above, if Wcs ₁ and Wcs ₂ are determined for each vehicle, the transfer characteristics from the speaker mouth to the microphone output end including the characteristics of the microphone can be measured. However, it is complicated to determine Wcs ₁ and Wcs ₂ for each vehicle. Therefore, when the vehicle type and the microphone position on which the voice recognition device is mounted can be specified, the values of the coefficients Wcs ₁ and Wcs ₂ are determined in advance for one vehicle, and the values are transferred to the transmission circuits 57 and 58 in FIG. Set. However, this method does not have a microphone characteristic correction effect.
[0020]
FIG. 3 is a configuration diagram of another AMNOR type noise reduction system having a function of measuring a transfer function from the speaker's mouth to the output end of each microphone. The same parts as those in FIG. Yes. The system shown in FIG. 3 has the configuration shown in FIG. 2 when the transfer function is measured, and the configuration shown in FIG. 1 when learning / speech recognition is performed. 3 differs from FIG. 1 in that
(1) The changeover switches SW1 to SW4 are provided,
(2) A speaker 80 for measuring transfer characteristics is provided in the vicinity of the speaker's mouth,
(3) The transfer circuits 57 and 58 are configured to be capable of adaptive signal processing, and the transfer function from the speaker mouth to each microphone output end can be measured.
It is.
[0021]
The transmission circuit 57 includes an adaptive signal processing unit 57a and a subtraction unit 57b. Adaptive signal processing unit 57a includes the reference signal white noise, the target signal output of microphone 51, and updates the coefficient Wcs ₁ of the adaptive filter by performing adaptive signal processing so that the power error e ₁ is minimized, subtraction The unit 57b outputs a difference (error) e ₁ between the output of the microphone 51 and the output of the adaptive signal processing unit 57a.
The transmission circuit 58 includes an adaptive signal processing unit 58a and a subtraction unit 58b. Adaptive signal processing unit 58a includes the reference signal white noise, the target signal output of microphone 52, and updates the coefficient Wcs ₂ of the adaptive filter by performing adaptive signal processing so that the power error e ₂ is minimized, subtraction The unit 58b outputs a difference (error) e ₂ between the output of the microphone 52 and the output of the adaptive signal processing unit 58a.
[0022]
In order to measure the transfer function from the speaker's mouth to the output terminal of each microphone, the switches SW1, SW2, and SW3 are turned on, the switch SW4 is turned off, and the system is configured as shown in FIG. Thereafter, if the coefficients of the adaptive filters (not shown) of the adaptive signal processing units 57a and 58a converge to the constant values Wcs ₁ and Wcs ₂ by the method described in FIG. 2, each adaptive filter is connected to the microphone from the speaker's mouth. The transfer functions CS1 'and CS2' up to the output terminals 51 and 52 are simulated. Note that the speaker 80 is removed upon completion of the measurement.
For learning / recognition, the switches SW1, SW2, and SW3 are turned off, the switch SW4 is turned on, and the system is configured as shown in FIG. Thereafter, learning and speech recognition are performed by the method described in FIG.
[0023]
FIG. 4 is an overall control flow of the noise reduction system of FIG.
First, it is checked whether the transfer function from the speaker's mouth to the output terminal of each microphone is identified (step 101). If the mode is the identification mode, the switches SW1, SW2, SW3 are turned on and the switch SW4 is turned off. The system is configured as shown in FIG. 2, and the coefficients Wcs ₁ and Wcs ₂ are updated by adaptive signal processing (step 102). Next, it is checked whether the coefficients Wcs ₁ and Wcs ₂ have converged to a constant value (step 103), and the update process of step 102 is performed until the coefficients Wcs ₁ and Wcs ₂ have converged to a constant value. When the coefficients Wcs ₁ and Wcs ₂ converge to a constant value, the adaptive filters of the adaptive signal processing units 57a and 58a simulate transfer functions CS1 ′ and CS2 ′ from the speaker mouth to the output ends of the microphones 51 and 52, respectively. It will be. When it converges to a certain value, the switches SW1, SW2, and SW3 are turned off, the switch SW4 is turned on, and the system is configured as shown in FIG.
[0024]
Next, it is checked whether or not the talk switch for instructing the start of voice recognition is operated to enter the voice recognition state (step 104). If the talk switch is not turned on, the learning mode is set, so that the 2-input / 1-output adaptive signal processing unit 61 uses the output signals of the adders 59 and 60 as reference signals so that the power of the error signal e is minimized. Then, adaptive signal processing is performed to update the coefficients W ₁ and W ₂ of the adaptive filter (step 105). Thereafter, the coefficients W ₁ and W ₂ are updated until the talk switch is turned on. As a result, the coefficients W ₁ and W ₂ converge to a constant value.
In step 104, when the talk switch is turned on, the voice recognition mode is set, and the updating operation of the coefficients W ₁ and W ₂ is terminated (step 106). At the time of speech recognition, the adaptive signal processing unit 61 does not update the filter coefficients, but keeps the filter coefficients W ₁ and W ₂ determined at the time of learning set in the adaptive filter, and the subtraction unit 63 outputs from the target response setting unit 62. A signal obtained by subtracting the output signal of the adaptive signal processing unit 61 from the target signal, that is, a speech signal with reduced noise and an improved S / N ratio is output to a speech recognition unit (not shown).
[0025]
Thereafter, the operation of step 106 is performed until the talk switch is turned off and the voice recognition is released. When the talk switch is turned off (step 107), the learning mode is returned to and the coefficients W ₁ and W after step 105 are set. The update operation of ₂ resumes.
According to the noise reduction system of FIG. 3, it is only necessary to identify the transmission characteristics CS1 ′ and CS2 ′ once in a production line or at a dealer before the vehicle is shipped. Since the transmission characteristics CS1 ′ and CS2 ′ including the characteristics can be identified, no problem occurs even if the microphones have irregular characteristics. Further, even when an uncertain element is included such as when the vehicle type cannot be specified, the transfer characteristics CS1 ′ and CS2 ′ can be identified and set in advance.
In the above, the case where the speech signal output from the system of the present invention is input to the speech recognition apparatus at the time of non-learning has been described. However, the present invention is not limited thereto, and the speech signal is input to the hands-free telephone at the time of non-learning, This can be applied when inputting to a device.
The present invention has been described with reference to the embodiments. However, the present invention can be variously modified in accordance with the gist of the present invention described in the claims, and the present invention does not exclude these.
[0026]
【The invention's effect】
As described above, according to the present invention, since the transfer characteristic from the mouth to the microphone is simulated with only the delay characteristic, the transfer characteristic from the actual mouth to the microphone is simulated, so that the noise reduction effect can be improved. In addition, according to the present invention, it is possible to remove vehicle interior noise that cannot be handled by a general AMNOR system.
Further, according to the present invention, by incorporating a mechanism for simulating the microphone-speaker transfer characteristic into the noise reduction device, it is possible to correct the variation in the microphone characteristic.
Further, according to the present invention, adaptive processing for noise reduction can be performed in real time.
[Brief description of the drawings]
FIG. 1 is a configuration diagram of a noise reduction system of the present invention.
FIG. 2 is a configuration diagram of a transfer characteristic measuring apparatus from a speaker's mouth to each microphone output terminal according to the present invention.
FIG. 3 is a configuration diagram of a noise reduction system having a transfer characteristic measurement function.
FIG. 4 is an overall control flow of the present invention.
FIG. 5 is a relationship diagram between an SN ratio and a recognition rate.
FIG. 6 is a conventional AMNOR type noise reduction system.
[Explanation of symbols]
51, 52... First and second microphones 55... Signal generators 57 and 58 for generating random noise as a target signal... Transfer circuit 59 and 60 for simulating transfer characteristics. 1-output adaptive signal processing units 61a, 61b, first and second adaptive signal processing units 62, target response setting unit 63, subtracting unit

Claims (1)

In the noise reduction system AMNOR method for improving the SN ratio of the speaker's speech signal,
Multiple microphones,
A signal generator that generates random noise signals,
A transfer circuit that simulates the transfer characteristics from the speaker's mouth to the output end of each microphone,
A synthesis unit that synthesizes a random noise signal output via each transmission circuit with an output signal of each microphone,
An adaptive signal processing unit that performs adaptive signal processing by using the random noise signal as a target signal and learning by using each synthesized unit output as a reference signal during learning, and updates the coefficient of the adaptive filter during non-learning;
A target response setting unit for giving a predetermined delay to the target signal;
A calculation unit that outputs a difference between the output signal of the adaptive filter and the output signal of the target response setting unit as an audio signal;
The transfer circuit measures transfer characteristics from the speaker mouth to the output end of each microphone, and simulates the transfer characteristics from the speaker mouth to the output end of each microphone by the measured transfer characteristics. And switching the system to transfer characteristic measurement mode, learning mode, and non-learning mode to measure transfer characteristics, learn filter coefficients, and output audio signals.
Noise reduction system characterized by

Images