JP2010056762A - Microphone array - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a microphone array providing a sound signal with an SNR (Signal-to-Noise Ratio) improved as compared with a conventional technique in a site such as a plant in which large noise is generated. <P>SOLUTION: The microphone array 10 includes: a microphone 1 disposed in such a way that a main axis of radiation is substantially turned to a speaker's mouth on an upper vertex out of each of vertexes of a pyramid; and a plurality of microphones 2, 3 and 4 disposed in such a way that the main axis of radiation is made substantially parallel with a speaker's mouth direction on at least two vertexes of bottom of the pyramid. The pyramid is, for instance, a triangular pyramid or a regular triangular pyramid. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a microphone array suitable for improving a speech recognition rate under, for example, a predetermined noise environment, and a speech recognition apparatus using the microphone array.

  For example, Patent Document 1 discloses a speech recognition apparatus that can improve the speech recognition rate by estimating the direction or position of a speaker.

  In the speech recognition apparatus according to this conventional example, in the speech recognition apparatus provided with the microphone array in which a plurality of microphones are juxtaposed at a predetermined interval, the direction estimation unit is configured to use the microphone array based on the electrical signal output from each microphone. The beam forming unit estimates at least one beam corresponding to the direction of the azimuth angle of at least one sound source estimated based on the electrical signal output from each microphone. Generate a signal. Next, the sound source determination unit determines whether each beam signal is speech or non-speech using the speech HMM and the noise HMM based on each beam signal, and the speech recognition unit 17 determines that it is speech. When this occurs, speech recognition is performed on the beam signal and a speech recognition result is output.

JP 2002-091469 A. JP2003-040992A. JP-A-11-327593. S. E. Boll, "Suppression of Acoustic Noise in Speech Using Spectral Subtraction", IEEE Transaction on Acoustic Speech and Signal Processing, Vol. ASSP-27, pp. 113-120, April 1979.

  However, when speech recognition is performed using a microphone array at a site that generates a large amount of noise, such as in a factory, there is a problem that the speech recognition rate is significantly reduced due to the noise.

  The object of the present invention is to solve the above-mentioned problems and obtain an audio signal having an improved signal-to-noise power ratio (hereinafter referred to as SNR) as compared with the prior art, for example, in a factory where a large noise is generated. An object of the present invention is to provide a microphone array that can be used and a speech recognition device that can obtain a speech recognition rate higher than that of the prior art by performing speech recognition using the microphone array.

The microphone array according to the present invention is:
A first microphone provided at a top vertex of each pyramid vertex such that the radiation main axis is substantially directed to the speaker's mouth;
A plurality of second microphones provided so that the main axis of radiation is substantially parallel to the direction of the speaker's mouth at at least two vertices of the bottom surface of the pyramid.

  In the microphone array, the pyramid is a triangular pyramid or a regular triangular pyramid.

  In the microphone array, three second microphones are provided at three vertices on the bottom surface of the regular triangular pyramid.

  Further, the microphone array is a voice recognition microphone array.

  According to the microphone array of the present invention, it is possible to obtain a voice signal that is improved as compared with the prior art by collecting the voice of the speaker using at least three microphones. In addition, the microphone array is used to record an audio signal, a subtractive array method is used to generate a plurality of cardioid signals, and a plurality of cardioid signals having a higher SNR are added to the sum signal. On the other hand, by performing speech recognition after removing noise using the spectral subtraction method, the speech recognition rate can be improved as compared with the prior art in a site where large noise is generated, for example, in a factory.

  Hereinafter, embodiments according to the present invention will be described with reference to the drawings. In addition, in each following embodiment, the same code | symbol is attached | subjected about the same component.

  FIG. 1 is a perspective view showing an arrangement of a microphone array 10 according to an embodiment of the present invention, FIG. 2 is a side view showing a microphone housing 11 including the microphone array 10 of FIG. 1, and FIG. 2 is a front view showing a microphone housing 11 of FIG. The microphone array 10 according to the present embodiment is characterized in that each omnidirectional microphone 1, 2, 3, 4 is provided at the position of each vertex of the regular triangular pyramid. In FIG. 1 and subsequent layout diagrams, the positions of the microphones 1, 2, 3, and 4 are shown in an XYZ three-dimensional coordinate system, and the layout coordinates of the microphones 1 to 4 are as follows.

(A) XYZ coordinates of the microphone 1 = (0, 0, 0); the position of the upper vertex of the regular triangular pyramid, which is located at the origin of the three-dimensional coordinate system of XYZ.
(B) XYZ coordinates of the microphone 2 = (0, √ (6) d / 3, √ (3) d / 3); the position of one vertex of the bottom of the regular triangular pyramid, 0 degrees on the XY plane and the XZ plane It is located in the direction of 55 degrees.
(C) XYZ coordinates of the microphone 3 = (d / 2, √ (6) d / 3, −√ (3) d / 6); the position of one vertex of the bottom surface of the regular triangular pyramid and 30 degrees on the XY plane And located at 110 degrees azimuth in the XZ plane.
(D) XYZ coordinates of the microphone 4 = (− d / 2, √ (6) d / 3, −√ (3) d / 6); the position of one vertex of the bottom surface of the regular triangular pyramid, which is 300 on the XY plane It is located at an orientation of 110 degrees on the XZ plane.

  In FIG. 1, the three-dimensional coordinate system of XYZ is arranged so that the voice emission direction of the speaker voice from the speaker's mouth tip 5 is the Y-axis direction. That is, the Y-axis direction is the direction of the normal vector from the speaker's mouth tip 5, the X-axis direction is the horizontal direction, and the Z-axis direction is the vertical direction. The main emission axes of the microphones 1 to 4 (the main axis of the radiation directivity characteristic and corresponding to the cylindrical axis of the microphone) are directed to the sound emission direction 6 and substantially parallel thereto. It is arranged to become.

  2 and 3, a microphone array 10 including four microphones 1 to 4 is accommodated in a microphone casing 11, and the microphone casing 11 is attached to the tip of a flexible arm 12 of a speaker's headphone set. . When the microphone housing 11 is viewed from the front, the radiation surfaces of the four microphones 1 to 4 can be seen as seen from FIG. 3, but only the microphone 1 at the top vertex is closer to the speaker's mouth. It has become. Moreover, the space | interval between each adjacent two microphones of the microphones 1-4 is set to 10 mm so that FIG. 2 may show. With respect to this, with respect to the sampling frequency of 16 kHz for sampling the audio signal, the maximum distance allowed between the microphones is 340000/16000 = 21.25 mm as in the case of the sampling theorem. In the case of performing signal processing using cardioid, it is necessary to further halve in order to prevent a reduction called folding, and the maximum allowable distance is determined to be 10.625 mm. Among them, a regular triangular pyramid is adopted as a shape that can obtain the maximum phase difference and angle difference.

  In the embodiment of FIGS. 1 to 3, the microphones 1 to 4 are arranged at the apexes of the regular triangular pyramid. However, the present invention is not limited to this, and the regular triangular pyramid may be a triangular pyramid, a polygonal pyramid, or a pyramid. At least two microphones 2 to 4 may be arranged at each vertex. In the case of a polygonal pyramid, the number of microphones disposed at each vertex of the bottom surface may be at least two, that is, a plurality of microphones may be disposed.

  FIG. 4 is a block diagram showing a configuration of a speech recognition apparatus using the microphone array 10 of FIG.

  In FIG. 4, the sound input to the microphone 1 is converted into an audio signal, and then converted into a digital signal S 1 via the low frequency amplifier 21 and the A / D converter 26, and input to the subtractors 41, 42, and 43. Is done. The sound input to the microphone 2 is converted into a sound signal, and then converted into a digital sound signal S2 via the low frequency amplifier 22 and the A / D converter 27. Then, the digital sound signal S2 is delayed. Is input to the subtractor 41 via the delay unit 31, input to the subtractor 44, input to the subtractor 45 via the delay unit 35, input to the subtractor 48 via the delay unit 38, and input to the subtractor 49. Is done. The sound input to the microphone 3 is converted into a sound signal, and then converted into a digital sound signal S3 via the low frequency amplifier 23 and the A / D converter 28. Then, the digital sound signal S3 is delayed by a delay unit 32. Is input to the subtractor 42 via the delay unit 34, input to the subtractor 44 via the delay unit 34, input to the subtractor 45, input to the subtractor 46, and input to the subtractor 47 via the delay unit 37. . The sound input to the microphone 4 is converted into a sound signal, and then converted into a digital sound signal S4 via the low frequency amplifier 24 and the A / D converter 29. Then, the digital sound signal S4 is converted into a delay unit 33. Is input to the subtracter 43 via the delay unit 36, input to the subtractor 46 via the delay unit 36, input to the subtractor 47, input to the subtractor 48, and input to the subtractor 49 via the delay unit 39. . Each of the delay units 31 to 39 has a delay amount of 29.4 microseconds in the present embodiment in order to compensate for the arrival time difference of the audio signal between adjacent microphones.

  The delay type array circuit 30 includes nine delay units 31 to 39 and nine subtracters 41 to 49, and uses a known subtraction type array method (for example, see Non-Patent Document 2). Thus, as described with reference to FIGS. 5 and 6, predetermined cardioids C1 to C9 that generate zero points (minimum points of directivity gain) with respect to the noise direction are generated.

  The subtracter 41 subtracts the delayed digital audio signal S2 from the digital audio signal S1 and performs signal evaluation on the cardioid audio signal SC1 (a digital audio signal detected by a directivity characteristic of cardioid C1, which will be described later). And output to the selection circuit 50. The subtractor 42 subtracts the delayed digital audio signal S3 from the digital audio signal S1 and performs signal evaluation on a cardioid audio signal SC2 (a digital audio signal detected by a directivity characteristic of cardioid C2 described later) as a subtraction result. And output to the selection circuit 50. The subtractor 43 subtracts the delayed digital audio signal S3 from the digital audio signal S1 and performs signal evaluation on a cardioid audio signal SC3 (a digital audio signal detected by a directivity characteristic of cardioid C3 described later) as a subtraction result. And output to the selection circuit 50.

  The subtracter 44 subtracts the delayed digital audio signal S3 from the digital audio signal S2, and performs signal evaluation on a cardioid audio signal SC4 (a digital audio signal detected by a directivity characteristic of cardioid C1, which will be described later). And output to the selection circuit 50. The subtractor 45 subtracts the delayed digital audio signal S2 from the digital audio signal S3, and performs signal evaluation on a cardioid audio signal SC5 (a digital audio signal detected by a directivity characteristic of cardioid C5 described later) as a subtraction result. And output to the selection circuit 50. The subtractor 46 subtracts the delayed digital audio signal S4 from the digital audio signal S3, and performs signal evaluation on a cardioid audio signal SC6 (a digital audio signal detected by a directivity characteristic of cardioid C6 described later) as a subtraction result. And output to the selection circuit 50. The subtractor 47 subtracts the delayed digital audio signal S3 from the digital audio signal S4, and performs signal evaluation on a cardioid audio signal SC7 (a digital audio signal detected by a directivity characteristic of cardioid C7 described later) as a subtraction result. And output to the selection circuit 50. The subtracter 48 subtracts the delayed digital audio signal S2 from the digital audio signal S4, and performs signal evaluation on the cardioid audio signal SC8 (a digital audio signal detected by the directivity of cardioid C8 described later) as a result of subtraction. And output to the selection circuit 50. The subtractor 49 subtracts the delayed digital audio signal S4 from the digital audio signal S2 and performs signal evaluation on a cardioid audio signal SC9 (a digital audio signal detected by a directivity characteristic of cardioid C9 described later) as a subtraction result. And output to the selection circuit 50.

The signal evaluation and selection circuit 50 detects a voice interval and a noise interval using the VAD (Voice Activity Detection) function for the nine input cardioid audio signals SC1 to SC9, and calculates an SNR based on the detected voice interval and noise interval. The top two (three in the modified example) cardioid audio signals with the highest SNR are selected, the selected cardioid audio signals are added, and the resulting cardioid audio signal is output to the noise removal circuit 51. Here, the VAD function detects a voice section under the following conditions.
(1) The signal level is not less than a predetermined threshold value.
(2) There is no cardioid signal that is more than a predetermined power level. This is because the three cardioid signals corresponding to the mouth direction and the cardioid signal corresponding to the face plane direction are not limited to the former three cardioid sound signals and the latter six carousoid signals for the sound from the mouth direction. While the power of the geoid audio signal is slightly increased, the audio signal from directions other than the mouth is likely to enter the blind spot of one or more cardioids, and the power difference tends to open relatively among the nine. It is intended to use what is in.
(3) Handles 500 milliseconds before and after the frame detected as the speech section as the speech section.

  Next, the noise removal circuit 51 removes noise in the speech signal from the input cardioid speech signal using a known spectral subtraction method (hereinafter referred to as SS method), and the processed digital speech signal Is output to the voice recognition circuit 52. Here, the SS method has been conventionally used as a noise removal method in the frequency domain, and the power spectrum of the separately estimated noise is subtracted from the power spectrum of the speech signal to which the noise is added, and the power spectrum is subjected to inverse Fourier transform. Thus, the audio signal from which noise has been removed is restored (see, for example, Patent Document 3 and Non-Patent Document 1). Here, the spectrum component X (f) after calculation using the SS method is expressed by the following equation.

[Equation 1]
X ² (f) = max {x (f) −αN (f), βN (f)} (1)

  Here, α and β are predetermined constants, for example, α = 2.0 and β = 0.001. Further, X (f) is a spectral component obtained as a result of spectral subtraction of noise, x (f) is a spectral component of recorded voice data (voice + noise), and N (f) is a noise spectral component.

  The speech recognition circuit 52 executes speech recognition processing on the input digital speech signal using, for example, a predetermined speech dictionary or speech model (for example, HMM), and text data of the speech recognition result is displayed on a liquid crystal display (LCD). The data is output to 53 or output to an external device such as a personal computer.

  Next, the cardioids C1 to C9 formed in the speech recognition apparatus of FIG. 3 will be described below with reference to FIGS.

  FIG. 5 is a perspective view showing three cardioids C1, C2, and C3 corresponding to the mouth orientations realized in the speech recognition apparatus of FIG. In FIG. 5, the cardioid C <b> 1 is formed by the digital audio signals S <b> 1 and S <b> 2 and has a zero point in the direction toward the microphone 2. The cardioid C2 is formed by the digital audio signals S1 and S3 and has a zero point in the direction toward the microphone 3. Further, the cardioid C3 is formed by the digital audio signals S1 and S3, and has a zero point in the direction toward the microphone 3.

  FIG. 6 is a perspective view showing six cardioids C4, C5, C6, C7, C8, and C9 corresponding to the horizontal face orientation realized in the speech recognition apparatus of FIG. In FIG. 6, cardioids C4 and C5 are formed by the digital audio signals S2 and S3. The cardioid C4 has a zero point in the direction toward the microphone 3, and the cardioid C5 is in the direction toward the microphone 2. Has zero point. The cardioids C6 and C7 are formed by the digital audio signals S3 and S4. The cardioid C6 has a zero point in the direction toward the microphone 4, and the cardioid C7 has a zero point in the direction toward the microphone 3. Have Further, the cardioids C8 and C9 are formed by the digital audio signals S4 and S2. The cardioid C8 has a zero point in the direction toward the microphone 2, and the cardioid C9 has a zero point in the direction toward the microphone 4. Have

  FIG. 7 is a perspective view illustrating a noise arrangement in a simulation experiment (three stationary noises Nst11, Nst12, and Nst13) according to the first embodiment performed by the present inventors. In FIG. 7, the symbol of the speaker indicates the arrangement position and radiation direction of the three stationary noises Nst11, Nst12, Nst13. Here, the stationary noise Nst11 is radiated from the orientation of the XY plane 60 degrees and the XZ plane 90 degrees, the stationary noise Nst12 is radiated in the direction from the + Y axis toward the origin, and the stationary noise Nst13 is radiated from the XY plane 300 degrees and XZ. Radiated from a 90-degree plane. At this time, the SNR (Cn) for each cardioid Cn (n = 1, 2,..., 9) evaluated by the speech recognition apparatus of FIG. 4 is as follows.

[Table 1]
―――――――――――――――――
SNR (C1) = 25.8 dB
SNR (C2) = 24.4 dB
SNR (C3) = 24.1 dB
SNR (C4) = 15.0 dB
SNR (C5) = 14.8 dB
SNR (C6) = 13.6 dB
SNR (C7) = 13.8 dB
SNR (C8) = 14.9 dB
SNR (C9) = 14.9 dB
―――――――――――――――――

SNR _ADD (Tm) when the upper m (m = 2, 3,..., 9) cardioid audio signals of the SNR (Cn) in Table 1 are added is shown below.

[Table 2]
―――――――――――――――――
SNR _ADD (T2) = 25.3 dB
SNR _ADD (T3) = 25.9 dB
SNR _ADD (T4) = 23.3 dB
SNR _ADD (T5) = 21.6 dB
SNR _ADD (T6) = 20.7 dB
SNR _ADD (T7) = 20.0 dB
SNR _ADD (T8) = 19.4 dB
SNR _ADD (T9) = 18.7 dB
――――――――――――――――――

  As is apparent from Table 2, the highest three SNR audio signals are obtained by adding the top three cardioid audio signals.

  FIG. 8 is a perspective view showing a noise arrangement in a simulation experiment (one sudden noise Nsu21) according to Example 2 performed by the present inventors. In FIG. 8, the symbol of the speaker indicates the arrangement position and radiation direction of the sudden noise Nsu21. Here, the sudden noise Nsu21 is radiated from the orientations of 60 degrees on the XY plane and 90 degrees on the XZ plane. At this time, the SNR (Cn) for each cardioid Cn (n = 1, 2,..., 9) evaluated by the speech recognition apparatus of FIG. 4 is as follows.

[Table 3]
―――――――――――――――――
SNR (C1) = 5.2 dB
SNR (C2) = 0.8 dB
SNR (C3) = 16.4 dB
SNR (C4) = − 6.5 dB
SNR (C5) = 1.3 dB
SNR (C6) = 16.0 dB
SNR (C7) = − 8.6 dB
SNR (C8) = − 6.6 dB
SNR (C9) = 1.6 dB
―――――――――――――――――

SNR _ADD (Tm) when the upper m (m = 2, 3,..., 9) cardioid audio signals of the SNR (Cn) in Table 3 are added is shown below.

[Table 4]
―――――――――――――――――
SNR _ADD (T2) = 16.2 dB
SNR _ADD (T3) = 9.5 dB
SNR _ADD (T4) = 7.1 dB
SNR _ADD (T5) = 6.5 dB
SNR _ADD (T6) = 5.0 dB
SNR _ADD (T7) = 2.7 dB
SNR _ADD (T8) = 1.3 dB
SNR _ADD (T9) = − 0.5 dB
――――――――――――――――――

  As is apparent from Table 4, the highest two SNR audio signals are obtained by adding the top two cardioid audio signals.

  FIG. 9 is a perspective view showing a noise arrangement in a simulation experiment (one sudden noise Nsu31 and one stationary noise Nst32) according to Example 3 performed by the present inventors. In FIG. 9, the symbol of the speaker indicates the arrangement position and the radiation direction of one sudden noise Nsu31 and one stationary noise Nst32. Here, the sudden noise Nsu31 is radiated from the azimuth of 60 degrees of the XY plane and 90 degrees of the XZ plane, and the stationary noise Nst32 is radiated from the azimuth of 300 degrees of the XY plane and 90 degrees of the XZ plane. At this time, the SNR (Cn) for each cardioid Cn (n = 1, 2,..., 9) evaluated by the speech recognition apparatus of FIG. 4 is as follows.

[Table 5]
―――――――――――――――――
SNR (C1) = 9.3 dB
SNR (C2) = 6.4 dB
SNR (C3) = 9.4 dB
SNR (C4) = − 1.5 dB
SNR (C5) = 0.8 dB
SNR (C6) = − 0.2 dB
SNR (C7) = − 2.9 dB
SNR (C8) = − 1.2 dB
SNR (C9) = 1.0 dB
―――――――――――――――――

The SNR _ADD (Tm) when the upper m (m = 2, 3,..., 9) cardioid audio signals of the SNR (Cn) in Table 5 are added is shown below.

[Table 6]
―――――――――――――――――
SNR _ADD (T2) = 10.0 dB
SNR _ADD (T3) = 7.6 dB
SNR _ADD (T4) = 7.0 dB
SNR _ADD (T5) = 6.4 dB
SNR _ADD (T6) = 5.6 dB
SNR _ADD (T7) = 4.9 dB
SNR _ADD (T8) = 4.3 dB
SNR _ADD (T9) =-3.4 dB
――――――――――――――――――

  As is apparent from Table 6, the highest two SNR audio signals are obtained by adding the top two cardioid audio signals.

FIG. 10 is a perspective view showing a noise arrangement in a simulation experiment (one stationary noise Nst41) according to Example 4 performed by the present inventors. In FIG. 10, the symbol of the speaker indicates the arrangement position and radiation direction of one stationary noise Nst41. Here, the stationary noise Nst41 is radiated at a background noise level of 90 dBA from directions of 30 degrees on the XY plane and 90 degrees on the XZ plane. At this time, based on each cardioid speech signal SCn (n = 1, 2,..., 9) evaluated by the speech recognition apparatus of FIG. _ADD (Tm) is shown below.
[Table 7]
―――――――――――――――――
SNR _ADD (T2) = 8.0 dB
SNR _ADD (T3) = 7.3 dB
――――――――――――――――――

  Here, SNRss when the noise removal circuit 51 of FIG. 4 using the SS method is used when the top two cardioid audio signals having higher SNR are added is shown below.

[Table 8]
―――――――――――――――――――――――――――
SNR _SS (α = 1.0; β = 0.001) = 8.0 dB
SNR _SS (α = 2.0; β = 0.001) = 10.3 dB
―――――――――――――――――――――――――――

  As apparent from Table 8, it can be seen that the SNR is greatly improved by using the noise removal circuit 51 using the SS method.

  In the first to fourth embodiments described above, the stationary noise is white noise generated from, for example, a belt conveyor, and the sudden noise is sudden noise generated from, for example, punching of a metal material.

  In Example 5, the present inventors read 100 different numbers of four digits by the speaker under various severe noise environments (at the applicant's Inuyama Factory) under the following experimental conditions. The speech recognition rate was measured.

[Table 9]
―――――――――――――――――――――――――――――――――――――――
(A) Voice recognition software: NEC voice recognition test application (B) Recognition dictionary: Number recognition dictionary 4 digits (C) Microphone used:
(C1) NEC Headset Microphone (Comparative Example 1; comprising a single voice microphone and a non-directional noise microphone)
(C2) Sennheiser HMD-25 type microphone (Comparative Example 2)
(C3) Microphone array according to the present embodiment (embodiment; as shown in FIGS. 1 to 3, one omnidirectional audio microphone 1 and three omnidirectional noise microphones 2, 3, 4 And configured with.)
―――――――――――――――――――――――――――――――――――――――

  FIG. 11 is a table showing experimental results (speech recognition rate) of a speech recognition experiment under noise according to Example 5 performed by the present inventors. As is clear from FIG. 11, the SNR greatly improved as compared with the prior art by collecting sound using the microphone array 10 according to the present embodiment in a very severe noise environment where the noise level is 80 dBA. Can be obtained.

  Further, as is clear from the results of Examples 1 to 4, the voice is collected by using the microphone array 10 according to the present embodiment and the voice is recognized by using the voice recognition apparatus of FIG. 4 according to the present embodiment. The recognition rate can be greatly improved.

  In the above embodiment, the subtractive array method and the SS method are used together. However, the present invention is not limited to this, and speech recognition may be performed after signal processing using only the former.

  As described above in detail, according to the microphone array of the present invention, it is possible to obtain an improved audio signal as compared with the prior art by collecting the voice of the speaker using at least three microphones. . In addition, the microphone array is used to record an audio signal, a subtractive array method is used to generate a plurality of cardioid signals, and a plurality of cardioid signals having a higher SNR are added to the sum signal. On the other hand, by performing speech recognition after removing noise using the SS method, the speech recognition rate can be improved as compared with the prior art in a site where large noise is generated, such as in a factory.

It is a perspective view which shows arrangement | positioning of the microphone array 10 which concerns on one Embodiment of this invention. It is a side view which shows the microphone housing | casing 11 provided with the microphone array 10 of FIG. It is a front view which shows the microphone housing | casing 11 of FIG. It is a block diagram which shows the structure of the speech recognition apparatus using the microphone array 10 of FIG. FIG. 5 is a perspective view showing three cardioids C1, C2, and C3 corresponding to the mouth orientation realized in the speech recognition apparatus of FIG. FIG. 5 is a perspective view showing six cardioids C4, C5, C6, C7, C8, and C9 corresponding to the face horizontal orientation realized in the speech recognition apparatus of FIG. 4. It is a perspective view which shows the noise arrangement | positioning in the simulation experiment (three stationary noises Nst11, Nst12, Nst13) based on Example 1 performed by the present inventors. It is a perspective view which shows the noise arrangement | positioning in the simulation experiment (one sudden noise Nsu21) based on Example 2 performed by the present inventors. It is a perspective view which shows the noise arrangement | positioning in the simulation experiment (one sudden noise Nsu31 and one stationary noise Nst32) based on Example 3 performed by the present inventors. It is a perspective view which shows the noise arrangement | positioning in the simulation experiment (one stationary noise Nst41) based on Example 4 performed by the present inventors. It is a table | surface which shows the experimental result (voice recognition rate) of the speech recognition experiment under noise based on Example 5 performed by the present inventors.

Explanation of symbols

1, 2, 3, 4 ... microphones,
5 ... Mouth tip,
6 ... Sound radiation direction,
10 ... Microphone array,
11 ... Microphone housing,
12 ... Flexible arm,
21, 22, 23, 24 ... low frequency amplifiers,
26, 27, 28, 29 ... A / D converter,
30 ... Delay type array circuit,
31, 32, 33, 34, 35, 36, 37, 38, 39 ... delay devices,
41, 42, 43, 44, 45, 46, 47, 48, 49 ... subtractor,
50. Signal evaluation and selection circuit,
51. Noise removal circuit,
52. Voice recognition circuit,
53 ... Liquid crystal display (LCD),
C1, C2, C3, C4, C5, C6, C7, C8, C9 ... cardioid,
Nst11, Nst12, Nst13, Nst32, Nst41 ... stationary noise,
Nsu21, Nsu31 ... sudden noise.

Claims (5)

A first microphone provided at a top vertex of each pyramid vertex such that the radiation main axis is substantially directed to the speaker's mouth;
A microphone array comprising: a plurality of second microphones provided so that a principal axis of radiation is substantially parallel to a direction of a speaker's mouth at at least two vertices of the bottom surface of the pyramid.   The microphone array according to claim 1, wherein the pyramid is a triangular pyramid.   The microphone array according to claim 1, wherein the pyramid is a regular triangular pyramid.   4. The microphone array according to claim 3, wherein three second microphones are provided at three vertices of the bottom surface of the regular triangular pyramid.   5. The microphone array according to claim 1, wherein the microphone array is a voice recognition microphone array.

Images