 |
Description  |
|
|
BACKGROUND OF THE INVENTION
The present invention relates to a method of detecting an acoustic signal,
and a method of detecting a period of a desired acoustic signal in a
signal including noise and the desired acoustic signal.
In recent years, although speech recognition apparatuses have been
remarkably developed, the development of a speech recognition apparatus
for recognizing speech in a noisy environment has been retarded because it
is difficult to correctly detect a speech period (i.e., to detect a period
during which speech is present on the time axis) in a signal contaminated
by noise. When a noise period is recognized as a speech period, noise is
forcibly caused to correspond to any phoneme, and it is impossible to
obtain a correct speech recognition result. Therefore, it is very
important to develop a speech period detection technique which can be used
in a noisy environment.
FIG. 1 is a timing chart for explaining the first conventional speech
period detection method. This chart shows changes in short time power as a
function of time. The short time power of a signal output from a
microphone is plotted along the ordinate, and the time is plotted along
the abscissa. In the following description, the short time power will be
referred to as a "power". A signal generally contains stationary noise 11
(noise having almost a constant power, such as air-conditioning noise or
fan noise of equipment), unstationary noise 12 (noise whose power is
greatly changed, such as a door closing sound and undesired speech), and
desired speech 13. Although the power of the stationary noise can be known
in advance, the unstationary noise power is unpredictable.
According to the first conventional method, a power of a signal is kept
monitored. When this power exceeds a threshold value Th14 determined on
the basis of the stationary noise power, the corresponding period is
recognized as a speech period. Most of the existing speech recognition
apparatuses perform speech period detection by using this method.
According to this method, although a correct speech period 16 shown in
FIG. 1 can be detected, an unstationary noise period 15 having a high
power is also erroneously detected as a speech period, resulting in
inconvenience.
The second conventional method will be described below.
According to the second conventional method, two microphones are located to
cause an S/N ratio difference between outputs from the two microphones.
The examples of microphone arrangement for the method are shown in FIGS.
2(a) and 2(b). That is, as shown in FIG. 2(a), a first microphone 1 is
located near a speaker 3, and a second microphone 2 is located away from
the speaker 3. Alternatively, as shown in FIG. 2(b), the first microphone
1 is located in front of the speaker 3, and the second microphone 2 is
located near the side of the speaker 3. In these arrangement, the speech
power level of the output from the first microphone is higher than that
from the second microphone. On the other hand, assuming that noise is
generated in a remote location, the noise power levels of the outputs from
these microphones are almost equal to each other. As a result, an S/N
ratio difference in outputs of the two microphones occurs.
FIGS. 3(a), 3(b), and 3(c) are charts for explaining an ideal operation of
the second conventional method. More specifically, FIG. 3(a) shows a time
change in power P1 of the output from the first microphone, and FIG. 3(b)
shows a time change in power P2 of the output from the second microphone.
Reference numerals 11 in FIGS. 3(a) and 3(b) as in FIG. 1 denote
stationary noise; 12, unstationary noise, and 13, speech. Since the two
microphones are arranged as shown in FIG. 2(a) or FIG. 2(b), the power of
the speech in FIG. 3(b) is lower than that in FIG. 3(a), while the noise
power levels of these outputs are equal to each other. As shown in FIG.
3(c), according to the second conventional method, a difference PD
(=P1-P2) between the short time powers P1 and P2 of the two signals is
calculated. When the power difference PD is larger than a given threshold
value Pth17, a corresponding time period 18 is detected as a speech
period. According to the second conventional method, as is apparent from
FIG. 3(c), the unstationary noise period having a high power is not
detected as a speech period, unlike in the first conventional method.
The second conventional method, however, is rarely operated in an ideal
state because the following three conditions must be satisfied to
correctly detect a speech period by utilizing a power difference in the
two signals:
Condition 1: An S/N ratio difference in two signals must be present.
Condition 2: Noise and speech periods of the two signals must be matched
with each other as a function of time.
Condition 3: A variation in S/N ratio difference caused by various factors
is small (stability of the S/N ratio difference).
According to the second conventional method, the first condition is
satisfied, while the second and third conditions are not satisfied.
Therefore, the following problems are posed.
The first problem will be described below. FIG. 4 shows an arrangement
obtained by adding a noise source 4 to the arrangement of FIG. 2(a). At
this time, speech is input to the first microphone 1 and then the second
microphone 2. However, noise is input to the second microphone 2 and then
the first microphone 1. Therefore, the speech and noise periods of the two
microphone output signals are not matched as a function of time.
The above situation is shown in FIGS. 5(a), 5(b), and 5(c). FIG. 5(a) shows
the power P1 of the output from the first microphone 1, FIG. 5(b) shows
the power P2 of the output from the second microphone 2, and FIG. 5(c)
shows the power difference PD. Reference numeral 11 denotes stationary
noise; 12, unstationary noise; and 13, speech, as in FIGS. 3(a) to 3(c).
Relationships between the speech powers and the noise powers in FIGS. 5(a)
and 5(b) are the same as those in FIGS. 3(a) and 3(b). However, in the
relationships shown in FIGS. 5(a) and 5(b), the speech as the output from
the second microphone 2 is delayed from that as the output from the first
microphone 1 by a period .tau.S31, whereas the noise as the output from
the second microphone 2 advances from that from the output from the first
microphone by a period .tau.N32. The speech and noise periods are not
matched with each other as a function of time. As a result, the difference
PD between the two signal powers is different from that of FIG. 3(c), as
shown in FIG. 5(c). When a period during which the difference exceeds the
threshold value Pth17 is detected as a speech period, a period 33 in FIG.
5(c) is erroneously detected as a speech period, thus posing the first
problem. Because the time difference .tau.N32 in this noise period is
greatly changed depending on the position of the noise source, it is
impossible to establish matching by using a delay element.
As the second problem, there are various factors for changing an S/N ratio
difference between the two microphone outputs in a practical situation,
therefore, it is difficult to assure stability of the S/N ratio difference
between the two signals as follows.
The first variation factor is the position of the noise source. As
described above, the noise source is assumed to be located in a remote
location. When, however, the noise source is located at a relatively close
location, the position of the noise source becomes a large variation
factor for the S/N ratio difference. FIGS. 6(a) and 6(b) explain this
situation. Reference numerals 1 and 2 in FIGS. 6(a) and 6(b) denote first
and second microphones, respectively; 3, speakers; and 4, noise sources,
as in FIG. 4. When the noise source 4 is located at positions indicated in
FIGS. 6(a) or 6(b), the noise power of the output from the first
microphone 1 is higher than that from the second microphone 2, as in the
speech powers. As a result, an S/N ratio difference between the two
microphone outputs becomes fairly small.
The second variation factor is movement of the speaker. For example, when
the speaker 3 turns his head in a right 45.degree. direction in FIG. 6(b),
the speech signal is received by each microphone at almost the same level.
As a result, a speech power difference does not occur in the outputs of
the two microphones, thus an S/N ratio difference varies.
The third variation factor is an influence of room echoes. When two
microphones are located so as to cause the S/N ratio difference in their
outputs, room echoes having different time structures and magnitudes are
added to the noise and speech components of the each microphone output. As
a result, an S/N ratio is difference greatly changed as a function of
time.
In addition to the above mentioned major variation factors, there are other
factors such as electrical noise and vibration noise. Therefore, it is
very difficult to find a microphone arrangement which assure a stable S/N
ratio difference in an atmosphere where these various factors for changing
the S/N ratios are present.
As described above, the second conventional method has the above decisive
drawback and cannot be effectively utilized in practical applications.
The third conventional method for overcoming this drawback of the second
conventional method will be described with reference to FIG. 7. Referring
to FIG. 7, reference numeral 1 denotes a first microphone; 2, a second
microphone; 21, a short time power calculation unit; 22, a speech period
candidate detection unit; 23 and 24, average power calculation units for
speech period candidates; 25, a power difference detection unit; and 26, a
speech period candidate testing unit.
According to this method, as in the second conventional method, the first
microphone is located such that a ratio of speech to ambient noise is
large, whereas the second microphone is located such that an S/N ratio is
smaller than that of the first microphone. According to this method, a
short time power of an output signal from the first microphone 1 is
calculated by the short time power calculation unit 21. The short time
power of the signal is kept monitored by the speech period candidate
detection unit 22. The speech period candidate detection unit 22 detects a
speech period candidate as a period when its power exceeds a threshold
value Th. The above operations are the same as those in the first
conventional method shown in FIG. 1. The noise period 15 shown in FIG. 1
is detected as a speech period candidate. Then, average powers of the
outputs from the first and second microphones during this candidate period
are calculated by the average power calculation units 23 and 24. Next, the
difference PDL between two average powers is obtained by the power
difference detection unit 25. Finally, when the power difference PDL
exceeds a predetermined threshold value PDLt, this candidate period is
recognized as a correct speech period by the speech period candidate
testing unit 26. Otherwise, this candidate period is discarded.
According to the characteristic feature of the third conventional method, a
difference between the average powers obtained within a relatively long
time candidate period, is calculated in place of the short time power
difference. Even if the speech and noise periods of one microphone output
are not matched with those of the other microphone output, as shown in
FIGS. 5(a) and 5(b), or even time variations in S/N ratio caused by room
echoes occur, its influence on the average power difference is reratively
small. Therefore, the third conventional method seems to solve the
problems of the second conventional problem.
In the third conventional method, however, since the speech period is
determined based on the average power within the candidate period, an
incorrect discrimination result occurs when the noise and speech periods
appear continuously, as shown in FIG. 8. FIG. 8 shows an output from the
first microphone. A correct speech period is a period 34 in FIG. 8. As
shown in FIG. 8, since unstationary noise 12 is close to speech 13 along
the time axis, a period 35 which contains both the noise and speech
periods and the short time power of which exceeds a threshold value Th14
is detected as a speech period candidate. When this candidate period 35 is
discriminated as a correct speech period upon calculation of an average
power difference, a period 36 shown in FIG. 8 becomes an erroneously
detected period. When the above speech period is discarded, the correct
speech period is recognized as a non-speech period. In either case, an
erroneous discrimination result is obtained.
The third conventional method, therefore, cannot serve as a means for
solving the drawback of the second conventional method.
Various problems are present in the conventional speech period detection
methods. It is therefore difficult to correctly detect a speech period
when unstationary noise is present in an input signal.
SUMMARY OF THE INVENTION
It is therefore a principal object of the present invention to provide a
method of detecting an acoustic signal, capable of detecting an speech
period in an atmosphere of unstationary noise with higher precision than a
conventional technique.
It is another object of the present invention to provide a method of
detecting an acoustic signal, capable of detecting a speech period with
high precision even if a noise source is present at an arbitrary position
except for a position near a speaker (+30.degree. range when the speaker
is viewed from the microphone), and even if the speaker moves within an
expected range.
In order to achieve the above objects of the present invention, the
following requirements are indispensable. That is, in order to correctly
detect a speech period by using a power difference between two signals,
the following three conditions must be satisfied:
Condition 1: An S/N ratio difference in two signals must be present.
Condition 2: Noise and speech periods of the two signals must be matched
with each other as a function of time.
Condition 3: A variation in S/N ratio difference caused by various factors
is small (stability of the S/N ratio difference).
According to the first feature of the present invention, in order to
satisfy both the first and second conditions, two sound receiving units
for generating signals having different S/N ratios are located at a single
position (strictly speaking, this single position can be positions which
can be deemed to be a single position to effectively operate the present
invention), and a speech period is detected by using a power difference
between the two output signals. According to the second feature of the
present invention, one of the two sound receiving units comprises a
microphone array system having a directivity control function to satisfy
the third condition.
According to the first feature of the present invention, since noise and
speech reach both the sound receiving units at the identical time, the
noise and speech periods of an output from one sound receiving unit are
matched with those from the other sound receiving unit as a function of
time, thus satisfying the second condition and solving the first problem
of the second conventional method.
When the two sound receiving units are located at the single position, the
time structures of the echoes added to the signals are equal to each
other. Therefore, the influence of the echoes which causes variations in
S/N ratio difference between the two sound receiving unit outputs, as
pointed as the second problem of the second conventional method, can be
greatly reduced by the first feature of the present invention.
According to the second feature of the present invention, variations in S/N
ratio difference between the two sound receiving unit outputs caused by
the position of the noise source and movement of the speaker, as pointed
out as the second problem of the second conventional problem, can be
decreased. This will be described in detail later.
The present invention will be described in detail with reference to
preferred embodiments in conjunction with the accompanying drawings.
BRIEF DESCRIPTION OF THE DRAWINGS
FIG. 1 is a chart showing the first conventional speech period detecting
method;
FIGS. 2(a) and 2(b) are views showing microphone arrangements for
explaining the second conventional speech period detecting method;
FIGS. 3(a), 3(b), and 3(c) are charts for explaining an ideal operation of
the second conventional method;
FIG. 4 is a view showing a positional relationship between microphones and
a noise source;
FIGS. 5(a), 5(b), and 5(c) are charts for explaining problems of the second
conventional method;
FIGS. 6(a) and 6(b) are views each showing a relationship between
microphones and a noise source;
FIG. 7 is a block diagram showing a third conventional speech period
detecting method;
FIG. 8 is a chart for explaining a problem of the third conventional method
described in FIG. 7;
FIG. 9 is a block diagram for explaining an embodiment of a method of
detecting an acoustic signal according to the present invention;
FIGS. 10(a) and 10(b) are views for explaining problems posed when
unidirectional and omnidirectional microphones are used;
FIG. 11 is a view for explaining a problem posed when a superdirectional
sound receiving unit is used;
FIG. 12 is a block diagram of a detailed arrangement of a first sound
receiving unit shown in FIG. 9;
FIG. 13 is a view showing directivity characteristics of an adaptive
microphone array;
FIGS. 14(a) and 14(b) are charts showing waveforms of reception signals of
impulsive noise with room echoes when an omnidirectional microphone and an
adaptive microphone array are used;
FIG. 15 is a block diagram showing a detailed arrangement of the embodiment
shown in FIG. 9;
FIGS. 16(a), 16(b), and 16(c) are charts for explaining an operation of a
speech period detection unit shown in FIG. 15;
FIGS. 17(a), 17(b), 17(c), and 17(d) are charts showing experimental
results to confirm effectiveness of the present invention; and
FIGS. 18, 19 and 20 are block diagrams showing other embodiments of the
present invention.
FIG. 21 is an alternative, yet equivalent, illustration of the diagram of
FIG. 12.
DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
An arrangement of the present invention is shown in FIG. 9. Referring to
FIG. 9, reference numeral 41 denotes a first sound receiving unit (i.e., a
microphone array system) for outputting a signal having a high S/N ratio.
The first sound receiving unit 41 comprises a microphone array 51
consisting of a plurality of microphone elements and a directivity
controller 52. Reference numeral 42 denotes a second sound receiving unit
for outputting a signal having an S/N ratio lower than that of the output
from the first sound receiving unit 41. These two sound receiving units 41
and 42 are located at the same position. Reference numerals 43 and 44
denote short time power calculation units; and 45, a speech period
detection unit based on a short time power difference.
In order to describe the effectiveness of the microphone array system in
the present invention, assume that a unidirectional microphone is used as
the first sound receiving unit 41 in place of the microphone array system,
and that an omnidirectional microphone is used as the second sound
receiving unit 42. With this arrangement, an S/N ratio of an output from
the first sound receiving unit directed toward the speaker is larger than
that of the output from the omnidirectional second sound receiving unit.
The above method is not always operated well, as will be described with
reference to FIGS. 10(a) and 10(b). Referring to FIGS. 10(a) and 10(b),
reference numeral 61 denotes a directivity pattern of a unidirectional
microphone; and 62, a directivity pattern of an omnidirectional
microphone. Reference numerals 3 denote speakers; and 63 and 64, positions
of the noise sources. As shown in FIG. 10(a), the unidirectional
microphone has a high sensitivity in the speaker side and a low
sensitivity in the opposite side. FIG. 10(b) shows the omnidirectional
microphone has equal sensitivity levels in all directions. When the noise
source is located at the position 63 in each of FIGS. 10(a) and 10(b), an
S/N ratio of an output from the unidirectional microphone is larger than
that of an output from the omnidirectional microphone. However, when the
noise source is located at the position 64 (or moved to the position 64)
in FIGS. 10(a) and 10(b), the sensitivity of the unidirectional microphone
for noise is much increased, and a difference between the S/N ratios of
the outputs from the unidirectional and omnidirectional microphones
becomes fairly small. In this manner, by the method using the
unidirectional microphone as the first sound receiving unit, the S/N
ratios are greatly changed depending on the position of the noise source.
The problem posed by use of the unidirectional microphone may be solved by
using a so-called "superdirectional sound receiving unit" as the first
sound receiving unit 41 of FIG. 9. However the directivity characteristics
of the "superdirectional sound receiving unit" generally vary depending on
frequencies. The directivity characteristics have almost omnidirectivity
in a low-frequency range and very sharp directivity as shown in FIG. 11 in
a high-frequency range. As a result, the S/N ratios are changed depending
on the position of the noise source in the low-frequency range, and the
S/N ratios are changed depending on slight movement of the speaker in the
high-frequency range.
As described above, in order to obtain good speech period detection
results, it is difficult to use a general-purpose directional sound
receiving unit as the first sound receiving unit 41 in the arrangement of
the present invention shown in FIG. 9.
In the present invention using the microphone array system having a
directivity control function, the variations in S/N ratio can be kept
small for changes in noise source position and movement of the speaker.
This will be described in detail below.
A typical example of a microphone array system having a directivity control
function is a sound receiving unit called an adaptive microphone array. An
arrangement of the adaptive microphone array is shown in FIG. 12.
Referring to FIG. 12, reference numeral 51 denotes a microphone array
consisting of M microphone elements 56.sub.1 to 56.sub.M ; and 52, a
directivity controller. The directivity controller 52 comprises filters
53.sub.1 to 53.sub.M respectively connected to microphone outputs, an
adder 55 for adding filter outputs, and a filter controller 54.
The filter controller 54 receives each microphone output signal and an
output x.sub.1 from the adder 55 and controls the characteristics of the
filters 53.sub.1 to 53.sub.M to reduce a noise component contained in the
output x.sub.1.
The principle of operation of the filter controller 54 will be described
below. The output signal x.sub.1 from the adder 55 can be expressed as a
sum of a speech component s and a noise component n as follows:
x.sub.1 =s+n (1)
When filter characteristics for minimizing a power n.sup.2 of the noise
component are unconditionally obtained, all the filters 53.sub.1 to
53.sub.M become filters having zero gain. As a result, although the noise
component n becomes minimized to zero, the speech component s is not
output either. Therefore, a constraint is imposed on the speech component
s contained in the signal x.sub.1 obtained as a result of a filtering
operation. Then, filter characteristics for minimizing the noise component
n contained in the output signal x.sub.1 under this constraint are
obtained. The constraint may be s=s.sub.0 where S.sub.0 is a speech
component contained in a microphone output signal (i.e., a filter input
signal) or a condition in which a mean value of .vertline.s-s.sub.0
.vertline..sup.2 is kept to be a threshold value or less.
When outputs from the M microphone elements are denoted as U.sub.1 to
U.sub.M, and characteristics of the filters 53.sub.1 to 53.sub.M are given
as h.sub.1 to h.sub.M, a power x.sub.1.sup.2 of the signal x.sub.1 is
represented as follows:
##EQU1##
Assuming that the speech and the noise are mutually uncorrelated, the
following equation is derived from equation (1):
x.sub.1.sup.2 =s.sup.2 +n.sup.2 (3)
Judging from equations (2) and (3), the power n.sup.2 of the noise
component contained in the output signal x.sub.1 is a second order
function of the filter characteristics h.sub.1 to h.sub.M. Therefore,
filter control for minimizing the power n.sup.2 of the noise component
under the constraint results in well-known minimization problem of the
second order function with a constraint.
Various solutions for various constraints, and practical algorithms are
described in detail in "Introduction to Adaptive Arrays", R. A. Monzingo
et al., John Wiley & Sons, New York, 1980 and U.S. Pat. No. 4,536,887.
A specific example of the method of realizing an adaptive microphone array
will be described. FIG. 21 shows a method proposed by Griffiths and Jim.
In FIG. 21, parts corresponding to those in FIG. 12 are denoted by like
refernce numerals, and corresponding signals are shown as like signals.
Reference numeral 51 denotes a microphone array consisting of M microphone
elements 56.sub.1 to 56.sub.M and a directivity controller 52. The
directivity controller 52 comprises subtracting units 57.sub.1 to
57.sub.M-1, adaptive filters 58.sub.1 to 58.sub.M-1, and a subtracting
unit 59. The subtracting units 57.sub.i (i being 1, 2, . . . , M-1)
receive microphone output signals u.sub.i and u.sub.i+1 and output
subtraction results v.sub.i. The adaptive filters 58.sub.1 to 58.sub.M-1
receive the subtraction results v.sub.1 to v.sub.M-1, and their outputs
are subtracted in the subtracting unit 59 from the first microphone
element output u.sub.1 to produce an output signal x.sub.1. The output
signal x.sub.1 is fed back to each adaptive filter.
The operation of this method is as follows. It is now assumed that the
microphone elements 56.sub.1 to 56.sub.M are arranged on a line, and a
voice arrives as a plane wave in a direction perpendicular to the line. At
this time, all the voice components contained in the microphone outputs
u.sub.1 to u.sub.M are in phase. Thus, by operations of taking the
difference between two microphone outputs in the subtracting units
57.sub.1 to 57.sub.M-1, the voice components are cancelled, that is, the
subtracting unit outputs v.sub.1 to v.sub.M-1 do not contain voice
components. If noise arrives in a direction different from the direction
of arrival of the noise, the noise components contained in signals u.sub.1
to u.sub.M are not in phase and thus are not cancelled throughthe
subtracting operation. Thus, the signals v.sub.1 to v.sub.M-1 contain the
sole noise components.
The adaptive filters correct the filter characteristics as a result of
subtraction of each filter output from the first microphone element output
u.sub.1 so as to minimize the power of the signal x.sub.1. These adaptive
filters are usually realized as digital filters, and the well-known LMS
algorithm or the like is used for the correction of the coefficients of
the digital filters. Details of the algorithm of the adaptive filters are
described in, for instance, B. Widrow and S. Samuel, "Adaptive Signal
Processing", Prentice-Hall, 1985. Also, various commercially available LSI
chips for realizing the function of the adaptive filter may be utilized.
Since the signal v.sub.1 to v.sub.M-1 contains the sole noise components,
the noise component contained in signal u.sub.1 is not affected by the
subtracting operation in the subtracting unit 59. This means that the
operation of the adaptive filters for minimizing the power of the output
x.sub.1 minimizes the power of the noise component contained in the output
x.sub.1. Thus, it is to be understood that the adaptive microphone array
structure shown in FIG. 21 is a method for minimizing the noise component
under a condition of x.sub.1 =s.
The structure shown in FIG. 21 may seem to be different from that shown in
FIG. 12. However, FIG. 21 is produced from FIG. 12 for facilitating the
understanding of the description, and these two Figures are equivalent.
Actually, the function of the filter controller 54 shown in FIG. 12 is
provided by the adaptive filters 58.sub.1 to 58.sub.M-1 shown in FIG. 21.
Further, considering characteristics between the input and output sides of
the directivity controller 52, there are correspondence relations h.sub.1
=1-g.sub.1, h.sub.i =g.sub.i-1 -g.sub.i (for i=2, 3, . . . , M), g.sub.i
being the filter characteristic of the i-th adaptive filter 58.sub.i.
To reduce the noise component contained in the output signal x.sub.1 is to
reduce the sensitivity of the array system in noise arrival directions. As
a result, this array system has a high sensitivity for a target direction
and a low sensitivity in unknown noise arrival directions.
FIG. 13 shows typical directivity characteristics 66 formed by the adaptive
array. Reference numeral 3 in FIG. 13 denotes a speaker as in the previous
embodiments; and 63 and 64, noise sources. As can be apparent from FIG.
13, although the adaptive array does not have sharp directivity, but has
directivity having a low sensitivity in the noise source directions. A
portion having this low sensitivity in the directivity is called a "dead
angle". When the microphone array consists of M elements, (M-1) dead
angles can be formed by the array system.
When noise reflected indoors reaches the adaptive array having such
directivity from many directions in addition to the noise source
direction, the resultant S/N ratio is small as compared with that of the
superdirectional sound receiving unit. However, adaptive array has a
feature capable of obtaining almost a constant S/N ratio for all noise
source locations except the neighborhood of a speaker (about+30.degree.
range when the speaker is viewed from the adaptive array), and it has a
feature of small variations in the S/N ratio upon movement of the speaker
3 since adaptive array does not have sharp directivity in the speaker
direction. According to these features, the adaptive microphone array is
very suitable for assuring stability in an S/N ratio difference for
detecting a speech period by using a difference between the two signal
power levels.
The adaptive microphone array has an additional feature capable of reducing
variations in noise power as a function of time.
Noise components reflected by walls, a floor, and a ceiling in addition to
noise directly from the noise source are input to the sound receiving unit
indoors. It is impossible for the adaptive microphone array to form dead
angles in all direct and reflected noise directions. When the microphone
array consists of M microphone elements, (M-1) dead angles are formed in
the directions where the sound is directly input or an echo having a high
energy is input, thereby improving the S/N ratio.
This effect will be described with reference to FIGS. 14(a) and 14(b). FIG.
14(a) shows impulsive noise with room echoes received by an
omnidirectional microphone, and FIG. 14(b) shows the one received by an
adaptive microphone array. Reference numeral 71 in FIG. 14(a) denotes
noise directly input from a noise source; and 72, 73, and 74, echoes of
noise reflected once or a plurality of times by the walls or floor and
then received. The energy levels of the echoes 72, 73, and 74 are
exponentially decreased as a function of time as compared with the energy
level of the direct noise 71. If the number of microphone elements
constituting the array is 4, three dead angles are formed in the noise
source direction and the directions of the echoes 72 and 73. An echo power
74 of the output (FIG. 14(b)) from the adaptive microphone array does not
have a large difference with that of the output (FIG. 14(a)) from the
omnidirectional microphone. However, the power levels of the direct noise
component and the echoes 72 and 73 are greatly decreased in FIG. 14(b). As
a result, variations in noise power as a function of time can be
apparently reduced by adaptive microphone array.
As previously described, the major factor for a detection error of a speech
period is large variations in noise power as a function of time, or in
other words, unstationary noise with high power causes incorrect
detection. In order to cope with these noise power variations, a speech
period is detected by utilizing a difference between two signal powers in
the present invention. It is, however, impossible to perfectly eliminate
various S/N ratio variation factors, i.e., eliminate detection errors by
100%. Therefore, the feature of the adaptive microphone array for reducing
the variations in noise power, or misdetection factor, is very effective
to reduce detection errors of speech periods.
There are many other choices for the second sound receiving unit 42 in FIG.
9 in addition to an omnidirectional microphone. The only requirement for
the second sound receiving unit is to output a signal which satisfies the
above-mentioned conditions 1 to 3 for the detection based on power
difference in cooperation with the first sound receiving unit 41.
One of the microphone elements constituting the microphone array 51 may be
used as the second sound receiving unit 42 in the arrangement of the
present invention of FIG. 9 according to the simplest way, which will be
shown in FIG. 15 (to be described later).
The second sound receiving unit 42 may be arranged, as shown in FIG. 18.
Referring to FIG. 18, the second sound receiving unit 42 comprises some of
a plurality of microphones as constituent elements of the first sound
receiving unit 41, i.e., a microphone array (which may sometimes be called
a sub-array when compared to the overall microphone array 51 in the first
sound receiving unit) and directivity synthesizer 52A. The output of the
microphone array is supplied to the directivity synthesizer 52A, and a
second signal x.sub.2 is output from the directivity synthesizer 52A. In
this specification, however, the "directivity synthesizer" is defined such
that is synthesizes the directivity through the simple operations of
delaying and addition on a plurality of signals. For example, in the case
where the microphone array in FIG. 18 is linear and the directivity
synthesizer is an adder for adding all the inputs, a high sensitivity
directivity is synthesized with respect to the direction perpendicular to
the line of the microphone array.
Another arrangement of a microphone array system having a directivity
control function for the first sound receiving unit 41 is exemplified as a
sound receiving system, as described in U.S. Pat. No. 791,418. In this
system, speech signals having clear arrival directions are preserved, and
signal processing is performed to suppress noise uniformly input form the
ambient atmosphere. In order to properly operate this system, a condition
in which a speaker position does not coincide with a noise source position
must be satisfied (in this condition, the direction of the speaker
position may be the same as the direction of the noise source position
when viewed from the microphone). A method in this system can be deemed as
a kind of directivity control in a sense that only sounds from a sound
source located at a desired position are extracted.
FIG. 15 is a block diagram showing a detailed arrangement of the first
embodiment (FIG. 9) of the present invention. Reference numeral 51 in FIG.
15 denotes a microphone array; 52, a directivity controller; 43, a first
short time power calculation unit; 44, a second short time power
calculation unit; and 45, a speech period detection unit, as in the
previous embodiment. Reference numeral 81 denotes a first amplifier,
connected to the output of the directivity controller 52, for receiving a
signal x.sub.1 and sending an output to the first short time power
calculation unit 43; 82, a second amplifier, connected to the second sound
receiving unit 42 (one of the microphone elements of the microphone array
51 is used in this embodiment), for receiving the signal x and sending an
output to the second short time power calculation unit 44; 83, a
subtracter for receiving outputs pl and p2 from the first and second short
time power calculation units 43 and 44; 84, a detection unit based on the
power for receiving the output pl from the first short time power
calculation unit 43 and detecting a short time period having a possibility
for constituting part of the speech period; 85, a detection unit based on
the power difference for receiving an output from the subtracter 83; and
86, a speech period determination unit for receiving an output Sl from the
detection unit 84 based on the power and an output S2 from the detection
unit 85 based on the power difference.
The sequence of this method will be described below.
A speech input containing noise is received by the microphone array 51. An
output signal from the microphone array 51 is input to the directivity
controller 52, and the directivity controller 52 generates the first
signal x.sub.1. An output from one of the microphone elements constituting
the microphone array 51 is given as x.sub.2. At this time, as a result of
directivity control by the directivity controller 52, an S/N ratio of the
signal x.sub.1 is larger than that of t | | |
