close

Вход

Забыли?

вход по аккаунту

?

DESCRIPTION JP2007251354

код для вставкиСкачать
Patent Translate
Powered by EPO and Google
Notice
This translation is machine-generated. It cannot be guaranteed that it is intelligible, accurate,
complete, reliable or fit for specific purposes. Critical decisions, such as commercially relevant or
financial decisions, should not be based on machine-translation output.
DESCRIPTION JP2007251354
The present invention provides a microphone or the like which can improve bone conduction
speech without depending on the speaker or the contents of utterance and without the need for
noise prediction or complicated calculation. A microphone filters a bone conduction voice signal
x (n) input from a bone conduction microphone that collects bone conduction voice, an air
conduction microphone that collects air conduction voice, and a bone conduction microphone,
and derived voice A transversal filter circuit 18 that outputs the signal y (n), a derived audio
signal y (n) that is input from the transversal filter circuit 18, and an air conduction audio signal
d that is input as a desired signal from the air conduction microphone a filter coefficient
adjustment unit 22 which outputs the difference of n) as an error signal e (n) and updates the
filter coefficient so that the error signal e (n) input from the adder 20 becomes smaller; A bone
conduction voice signal x (n) correlated with the voice signal d (n) can be output as a derived
voice signal y (n). [Selected figure] Figure 2
Microphone, voice generation method
[0001]
The present invention relates to a microphone that converts air and bone conduction sounds into
electrical signals, and a method of generating the same.
[0002]
In general, under high noise environments such as traffic noise, factory noise, aircraft noise,
subway noise, construction site noise and the like, voice quality is greatly degraded and smooth
10-04-2019
1
voice communication often becomes difficult.
As a means for solving such problems, a bone conduction microphone is widely known which
converts voice (bone conduction voice) transmitted by bone conduction in the body into an
electrical signal.
[0003]
However, in the conventional bone conduction microphone, high frequency components of 1 kHz
or more are easily attenuated, and the voice quality is deteriorated due to the fact that the
relative relationship of the amplitude between the phonemes of bone conduction speech is
largely different from that of air conduction speech. There was a problem that it was easy to do.
As a result, bone conduction speech may sometimes be converted into speech that is difficult to
distinguish, and has the disadvantage that naturalness and intelligibility are lower than air
conduction speech.
[0004]
As a measure for improving the sound quality of such a bone conduction microphone, for
example, Patent Document 1 improves the quality of the sound collected by the bone conduction
microphone by utilizing the difference in the long time spectrum of air conduction speech and
bone conduction speech. An air conduction sound estimation device is disclosed that corrects the
sound to be close to natural sound. JP 2004-279768 A
[0005]
However, in the conventional bone conduction microphone including the above air conduction
sound estimation apparatus, the improvement degree of bone conduction speech fluctuates
depending on the sounding content, the speaker, the mounting position of the microphone, etc.,
and the improvement degree of bone conduction speech is unstable. There was a point.
[0006]
In addition, there are cases where complicated calculations and variance prediction of noise are
required, and there is also a problem that the degree of improvement of bone conduction speech
largely depends on the accuracy of calculation and prediction.
10-04-2019
2
[0007]
The present invention has been made to solve such problems, and does not depend on the
speaker or the content of the utterance, and does not require prediction of noise or complicated
calculations. It is an object of the present invention to provide a microphone that can be
improved and can collect clear voice even in a noisy environment and a voice generation method
thereof.
[0008]
The present invention solves the above-mentioned subject by the following means.
[0009]
(1) According to the present invention, a bone conduction voice collection unit for collecting
bone conduction voice, an air conduction voice collection unit for collecting air conduction voice,
and a filter coefficient of a bone conduction voice signal input from the bone conduction voice
collection unit The difference between the filtering unit that performs filtering and outputs as a
derived voice signal, the derived voice signal that is input from the filtering unit, and the air
conduction voice signal that is input as a desired signal from the An error signal generation unit
for outputting as a signal, and a filter coefficient adjustment unit for updating the filter
coefficient of the filtering unit so that the error signal input from the error signal generation unit
becomes smaller, A microphone characterized in that a bone conduction voice signal having a
correlation with a conduction voice signal can be outputted as the derived voice signal.
[0010]
(2) The present invention is also the microphone according to (1), wherein the filter coefficient
adjustment unit updates the filter coefficient of the filtering unit using an algorithm having a
momentum term.
[0011]
(3) The present invention is also characterized in that the algorithm having the momentum term
is any one of a momentum least mean square algorithm or a neural network using an error back
propagation algorithm. Microphone.
[0012]
(4) The present invention comprises the steps of collecting bone conduction voice by the bone
10-04-2019
3
conduction voice collecting unit, collecting air conduction voice simultaneously with the bone
conduction voice by the air conduction voice collecting unit, and using the trajectory voice as a
learning signal. The bone conduction voice signal input from the bone conduction voice
collecting unit is corrected by the learning system unit to be used on the basis of the strength of
the correlation with the learning signal, and the bone conduction having a strong correlation with
the learning signal is corrected. And outputting the audio signal.
[0013]
According to the microphone of the present invention, it is possible to improve bone conduction
speech without depending on the speaker or the content of the utterance and without requiring
the prediction of the noise or the complicated calculation, etc. Excellent effects such as being able
to collect various sounds.
[0014]
Hereinafter, with reference to the drawings, a microphone according to an embodiment of the
present invention, an audio generation procedure performed by the microphone, and the like will
be described in detail.
[0015]
<マイクロホン>
[0016]
FIG. 1 is a schematic configuration diagram showing the configuration of the main part of the
microphone 10 according to the present embodiment.
[0017]
As shown in the figure, this microphone 10 comprises an air conduction microphone ("air
conduction sound collecting unit" according to the present invention) 12 for collecting air
conduction sound, and a bone conduction microphone (for which And a learning system 16 for
converting the deteriorated speech S1 collected by the air conduction microphone 12 and the
bone conduction microphone 14 into a high quality speech S2. .
[0018]
<Air conduction microphone>
10-04-2019
4
[0019]
The air conduction microphone 12 is only required to be able to collect air conduction voice, and
for example, a conventionally known air conduction microphone such as a dynamic microphone
or a condenser microphone can be applied.
The "air conduction sound" according to the present invention means a sound generated from the
vocal cords and transmitted by air vibration.
[0020]
<Bone conduction microphone>
[0021]
The bone conduction microphone 14 is only required to collect bone conduction sound, and a
bone conduction microphone conventionally known can be applied.
In addition, the "bone conduction voice" which concerns on this invention means the audio |
voice transmitted by the bone conduction in the body.
[0022]
<Learning system>
[0023]
FIG. 2 is a conceptual view of a learning system 16 applied to the microphone 10, and FIG. 3 is a
block diagram of the learning system 16. As shown in FIG.
[0024]
As shown in the figure, the learning system 16 includes a transversal filter circuit ("filtering unit"
according to the present invention) 18, an adder ("error signal generating unit" according to the
present invention) 20, and filter coefficients. It has the adjustment part 22, and is comprised.
10-04-2019
5
[0025]
The transversal filter circuit 18 filters the bone conduction voice signal x (n) input from the bone
conduction microphone 14 using a filter coefficient (described later) and outputs it as a derived
voice signal y (n) .
The transversal filter circuit 18, as shown in FIG. 3, includes a plurality of delay elements 26, 28
for delaying the bone conduction voice signal x (n), the bone conduction voice signal x (n) and
the delayed bone conduction signals x (n). Multipliers 30, 32, 34 for multiplying the audio
signals x (n-1), x (n-2) by the filter coefficients set by the filter coefficient adjustment unit 22 and
the outputs of the respective multipliers 30, 32, 34 And an adder 36 that outputs the derived
voice signal y (n).
[0026]
The adder 20 adds the difference between the derived voice signal y (n) input from the
transversal filter circuit 18 and the air conduction sound signal d (n) input as the desired signal
from the air conduction microphone 12 to the error signal e ( output as n).
[0027]
The filter coefficient adjustment unit 22 updates the filter coefficient of the transversal filter
circuit 18 so that the error signal e (n) input from the adder 20 is minimized.
[0028]
Note that, as an algorithm for updating the filter coefficient by the filter coefficient adjustment
unit 22, for example, Least Mean Square (LMS), normalized LMS (Normilazed LMS: NLMS),
momentum LMS (momentum LMS: MLMS), recursion Adaptive algorithms such as LSR (Recursive
Least Squares: RLS), Back Propagation (BP), and neural networks can be mentioned. Among them,
“MLMS algorithm (hereinafter simply referred to as“ MLMS ”) may be mentioned. Update
algorithm with momentum terms such as “may be)” and “neural network using BP algorithm
(hereinafter sometimes referred to simply as“ NNET ”)” The rhythm is preferred.
10-04-2019
6
The reason will be described later.
[0029]
Next, the operation of the microphone 10 according to the present embodiment will be described
using FIGS. 2 and 3.
[0030]
Now, the bone conduction voice signal x (n) is input from the bone conduction microphone 14,
and the filter coefficients set in the multipliers 30, 32, and 34 of the transversal filter circuit 18
at this time are c0 (n), respectively. , C1 (n), c2 (n).
[0031]
(1) First, the transversal filter circuit 18 performs product-sum operation using the filter
coefficients c0 (n), c1 (n), c2 (n), and outputs the derived voice signal y (n) shown below Do.
[0032]
[0033]
(2) Next, the adder 20 generates an error signal indicating the difference between the derived
sound signal y (n) and the air conduction sound signal d (n) input from the air conduction
microphone 12 as a desired signal. Output as e (n).
[0034]
[0035]
(3) Then, the filter coefficient adjustment unit 22 calculates filter coefficients c0 (n + 1), c1 (n +
1), c2 (n + 1) that minimize the error signal e (n) using the update algorithm, and performs
transversal The filter coefficients of the filter circuit 18 are updated.
[0036]
When the next bone conduction speech signal x (n + 1) is input, the transversal filter circuit 18
uses the filter coefficients c0 (n + 1), c1 (n + 1), c2 (n + 1) updated by the filter coefficient
adjustment unit 22. The product-sum operation is performed, and thereafter, the adaptive
10-04-2019
7
processing of (1) to (3) is repeatedly performed.
[0037]
The adaptive processing by such a learning system 16 can be described as the following
equation.
[0038]
The bone conduction speech signal vector X (n) is expressed as follows.
[0039]
[0040]
M is the filter order, and T is transpose.
[0041]
Also, the filter coefficient vector C (n) is expressed as follows.
[0042]
[0043]
At this time, the derived voice signal y (n) and the error signal e (n) are respectively expressed as
follows.
[0044]
(Equation 1)
[0045]
(Formula 2)
[0046]
The learning filter 16 updates the filter coefficient vector C (n) so that the error signal e (n) of
10-04-2019
8
(Expression 2) becomes smaller.
[0047]
Next, the principle of the present invention will be described.
[0048]
For example, when the air conduction voice signal d (n) on which noise is superimposed is
collected in the air conduction microphone 12, the adaptation process of the learning system 16
refers to a past signal and the air conduction voice signal which is a desired signal. The bone
conduction speech signal x (n) will be adapted to d (n).
And since the derived voice signal y (n) based on the bone conduction voice signal x (n) and the
error signal e (n) are in the relationship of the above (Equation 1) and (Eq. 2), the air conduction
voice signal A bone conduction speech signal x (n) (a real speech component other than noise)
correlated with d (n) is well adapted to the air conduction speech signal d (n) but correlated with
the air conduction speech signal d (n) The noise component without the noise component can not
be adapted, so the noise component is mainly output to the error signal e (n).
Therefore, if the filter coefficients are determined so that the error signal e (n) becomes small, it
is possible to derive noise approximating air conduction voice by removing noise.
[0049]
Further, in the present embodiment, the above-described MLMS algorithm or NNET is applied as
an update algorithm of the filter coefficient vector C (n).
[0050]
The “MLMS algorithm” is an algorithm in which a momentum term is introduced to the LMS
algorithm, and an update equation of the filter coefficient vector C (n) is given as follows.
[0051]
(Equation 3)
10-04-2019
9
[0052]
Here, β is a momentum coefficient.
[0053]
[0054]
In the LMS algorithm, if a large error occurs suddenly in the adaptation process, the filter
coefficient vector C (n) fluctuates greatly, but in the MLMS algorithm, the momentum term (the
term including the momentum coefficient) of (Equation 3) works. The change of the filter
coefficient vector C (n) becomes smooth.
Therefore, it is possible to suppress a rapid change of the derived voice signal y (n) and to reduce
distortion.
[0055]
FIG. 4 is a conceptual view of a Multi-Layer Perceptron (MLP) network.
[0056]
"NNET" calculates the derived voice signal y (n) from the bone conduction voice signal x (n) by
this MLP network, and also derives the derived voice signal y (n) and the air conducted voice
signal (desired signal) d (n) The unit connection weights and unit threshold values are updated
based on the error signal (learning signal) of
[0057]
The first layer is the input layer, the Mth layer is the output layer, the layer between them is the
hidden layer, and the weight of the connection from the jth neuron in the m-1th layer to the ith
neuron in the mth layer is wij Assuming that <[m]> and the learning signal of the output layer are
δi <[m]>, the update amount Δwij <[m]> of the connection weight between units in NNET is
expressed by the following equation.
[0058]
10-04-2019
10
(Equation 4)
[0059]
Here, μ is a step size, and β is a momentum coefficient.
[0060]
Also in this NNET, the change of the filter coefficient vector becomes smooth due to the function
of the momentum term in the second term of (Expression 4).
Therefore, it is possible to suppress a rapid change of the derived voice signal y (n) and to reduce
distortion.
[0061]
<Outline of Experiment>
[0062]
Next, an outline of an experiment conducted by the inventor of the present invention will be
described.
[0063]
First of all, in this experiment, the bone conduction microphone signal and the air conduction
microphone 12 collect the voices of the subjects (2 men and 2 women around 20 years old) by
using the bone conduction microphone 14 and the air conduction microphone 12, and then the
bone conduction voice signal and the air conduction voice signal are amplified. Amplified and
simultaneously recorded as audio signals of two channels.
[0064]
In this experiment, the subjects asked three vocalizations ("Who is a black dress?", "Are you taller
than you?"
10-04-2019
11
"I have a table of various sizes," I asked them each and recorded the music signal.
The sampling frequency of the music signal is 11.025 kHz, and the number of quantization bits is
16 bits.
[0065]
Next, in this experiment, assuming use in a noisy environment, white noise calculated by a
computer, in-vehicle noise (electronic co-noise database No. 1 <2000 cc class in a traveling
vehicle>), and crowded noise (electronic Common noise database No. 10 (people crowded) was
added to each of the recorded air conduction voices to generate three types of sound signals,
which were used as desired signals (air conduction voice signals).
[0066]
Then, LMS, NLMS, MLMS, RLS, and NNET algorithms were applied as the filter coefficient
updating algorithm, and the above-described adaptive processing was repeatedly performed
based on each algorithm to observe the derived voice signal.
[0067]
As parameters of each algorithm, LMS: step size μ = 0.1, filter order M = 10, NLMS: step size μ
= 0.1, stabilization parameter β = 0.01, filter order M = 10, MLMS Step size μ = 0.02,
momentum coefficient β = 0.1, filter order M = 10, RLS forgetting factor λ = 0.95, filter order M
= 10, NNET step size μ = 0.1, momentum The coefficients β = 0.1 and M = 4-4-4-1 (the number
of neurons in the input layer is 4, the number of neurons in the middle layer consisting of 2
layers is 4 each, and the number of neurons in the output layer is 1).
In addition, each parameter was changed using a logarithmic cross section ratio (LAR) as an
index in each algorithm, and finally, it was listened to the ear to determine an optimum
parameter.
[0068]
<Result of experiment>
10-04-2019
12
[0069]
Next, the results of this experiment will be described.
[0070]
In this experiment, “Log-Area Ratio (LAR)”, which is an evaluation of a frequency domain, and
“segmental SNR”, which is an evaluation of a time domain, are used as indices for evaluating
the quality of speech.
Note that "LAR" is a scale for evaluating the degree of distortion of speech from the reflection
coefficient using linear prediction (LPC) coefficients, and "segmental SNR" is the difference in
time waveform compared to the original signal It is a scale to measure
[0071]
5 to 7 are graphs showing the relationship between the SNR (horizontal axis) of noise added in
this experiment and LAR (vertical axis), and FIG. 5 is a graph showing the relationship between
white noise and LAR. FIG. 6 is a graph showing the relationship between in-vehicle noise and
LAR, and FIG. 7 is a graph showing the relationship between human noise and LAR.
[0072]
Note that “AIR” in the same figure is a comparison of air conduction speech with noise added
and clean air conduction speech without noise added, “BONE” is bone conduction speech with
noise added and clean Air conduction voices are compared.
Furthermore, LMS, NLMS, MLMS, RLS, and NNET are comparisons of the derived speech by each
algorithm and the clean air conduction speech.
[0073]
As shown in the figure, LMS, MLMS, and NNET algorithms were able to obtain derived voices that
10-04-2019
13
were better than air conduction voice (AIR) and bone conduction voice (BONE) in a condition
where the SNR was 0 dB or less. Among them, the speech derived by the MLMS algorithm
showed particularly good results.
It is considered that this is a result of improvement in the intelligibility due to the distortion of
bone conduction speech being improved by the adaptive processing of the learning system 16.
In addition, when the inventor actually auditioned each derived voice by ear, it is confirmed that
the derived voice by LMS, MLMS and NNET algorithm has reduced the feeling peculiar to bone
conduction voice and the clarity is increased. It was done.
[0074]
In addition, it is considered that the reason why the MLMS and NNET algorithms show better
results than other algorithms is that the momentum term suppresses abrupt changes in the
derived voice and distortion is reduced.
[0075]
8 to 10 are graphs showing the relationship between the SNR (horizontal axis) of noise added in
this experiment and the segmental SNR (vertical axis), and FIG. 8 shows the relationship between
white noise and segmental SNR. FIG. 9 is a graph showing the relationship between in-vehicle
noise and segmental SNR, and FIG. 10 is a graph showing the relationship between human
crowding noise and segmental SNR.
[0076]
As shown in the figure, in the case of white noise, the LMS, MLMS, NNET, and the speech derived
by the NLMS algorithm showed good results.
In addition, in the noise in the car and the crowded noise, the result was not much different from
the bone conduction speech.
[0077]
10-04-2019
14
Figures 11 to 14 by male subjects "Who is a person in black dress?
Is a diagram showing an audio waveform of
[0078]
Note that FIG. 11 shows, from the top, waveforms of clean air conduction speech, air conduction
speech with noise (white noise) added, bone conduction speech, and speech derived by the LMS
algorithm, respectively. FIG. It is the figure which showed the waveform of the derivation |
leading-out speech by NLMS, MLMS, RLS, and NNET algorithm from the top, respectively.
Further, FIG. 13 is a diagram showing, from the top, waveforms of clean air conduction speech,
air conduction speech to which noise (people crowd noise) is added, bone conduction speech, and
derived speech by the LMS algorithm, respectively. It is the figure which showed the waveform of
the derivation | leading-out speech by NLMS, MLMS, RLS, and NNET algorithm from the top,
respectively.
[0079]
As shown in the figure, in the voice derived by the LMS algorithm, amplification of noise is
suppressed, and it can be confirmed that the real voice part approaches clean air conduction
voice.
In addition, although human noise is slightly amplified compared with bone conduction speech, it
can be confirmed that the real speech part is similar to clean air conduction speech and the
intelligibility is improved, in particular, MLMS The algorithm derived speech is similar to clean
air conduction speech.
This is also considered to be because the momentum term of the MLMS algorithm suppresses
rapid changes in the derived voice and distortion is reduced.
10-04-2019
15
[0080]
Figures 15-18 by male subjects "Who is the person in black dress?
Is a diagram showing a spectrogram of
[0081]
FIG. 15 shows, from the top, a spectrogram of clean air conduction speech, air conduction speech
with noise (white noise) added, bone conduction speech, and speech derived by the LMS
algorithm, and FIG. It is a spectrogram of speech derived by MLMS, RLS, and NNET algorithm.
FIG. 17 shows, from the top, a spectrogram of clean air conduction speech, air conduction speech
added with noise (people crowded noise), bone conduction speech, and speech derived by the
LMS algorithm, and FIG. 18 shows, from above, NLMS, MLMS, It is a spectrogram of the voice
derived by RLS and NNET algorithm.
[0082]
As shown in the figure, in the voice derived by each algorithm, the high frequency region is
emphasized as compared to the bone conduction voice, and it can be confirmed that the voice
approach speech is approached.
[0083]
As described above, the microphone 10 according to the present embodiment includes the bone
conduction voice collecting unit (bone conduction microphone 14 in the present embodiment)
that collects bone conduction voice, and the air conduction voice collection unit that collects air
conduction voice ( In the present embodiment, the air conduction microphone 12) and a filtering
unit that filters the bone conduction voice signal x (n) input from the bone conduction voice
collection unit using a filter coefficient and outputs it as a derived voice signal y (n) In this
embodiment, the difference between the transversal filter circuit 18), the derived audio signal y
(n) input from the filtering unit, and the air-conduction audio signal d (n) input as the desired
signal from the air-conduction audio collection unit And an error signal generating unit (in the
present embodiment, the adder 20) that outputs the error signal e (n), and filtering so that the
error signal e (n) input from the error signal generating unit becomes smaller And a filter
10-04-2019
16
coefficient adjustment unit 22 which updates the filter coefficients of the above, and outputs a
bone conduction voice signal x (n) correlated with the air conduction voice signal d (n) as a
derived voice signal y (n) Since this is possible, it is possible to remove noise components that are
not correlated with the air conduction speech signal d (n) by performing an adaptive process on
the air conduction speech for the bone conduction speech.
[0084]
Therefore, the quality of bone conduction speech can be improved, and clear speech can be
collected even in a noisy environment.
Also, unlike the conventional bone conduction microphone, it does not depend on the speaker or
the content of the utterance, and does not require prediction of noise or complicated calculation.
Furthermore, the airway voice and bone conduction voice are collected in real time
simultaneously with vocalization, and the bone conduction voice side is used as an output voice,
and the airway voice side is used as a learning signal (desired signal / teaching signal). Since it
becomes possible to make the bone conduction voice approach the ideal voice by learning the
ideal sound inherent to this airway voice, it is possible to achieve both efficient removal of noise
and clarification of voice.
In addition, by using adaptive filters and neural networks, learning is made as needed in the
sound generation process, so that preparation processes etc. become unnecessary, and it is
possible to cope with any changes in the external sound environment over time, and A userfriendly microphone that can follow changes in vocal cords can be obtained.
[0085]
In particular, since the microphone 10 according to the present embodiment updates the filter
coefficient of the filtering unit using an algorithm having a momentum term, it is possible to
suppress a rapid change in the derived voice and to further reduce distortion. .
As the “algorithm with momentum term” is preferably a momentum least mean square
10-04-2019
17
algorithm (MLMS) or a neural network (NNET) using an error back propagation algorithm, as
shown in the experimental results. The invention is not limited to this.
[0086]
The microphone according to the present invention can collect clear voice even under a noisy
environment, so it can be widely used in, for example, station platforms, pachinko parlors, sports
stadiums, etc., and mobile phones and voice recognition The present invention can be widely
applied as a microphone of communication equipment and audio equipment represented by a
system and the like.
[0087]
FIG. 2 is a schematic configuration view showing the configuration of main parts of a microphone
according to an embodiment of the present invention.
It is a conceptual diagram of a learning system applied to the microphone.
It is a block diagram of the learning system.
It is a conceptual diagram of the multilayer perceptron network applied to the learning system.
It is the graph which showed the relationship between white noise and LAR.
It is the graph which showed the relationship between in-vehicle noise and LAR.
It is the graph which showed the relationship between crowd noise and LAR.
It is the graph which showed the relationship between white noise and segmental SNR.
It is the graph which showed the relationship between in-vehicle noise and segmental SNR.
10-04-2019
18
It is the graph which showed the relationship between human congestion noise and segmental
SNR.
It is the figure which showed the waveform of the induction | guidance | derivation audio | voice
which added the clean air conduction speech, the noise (white noise), the bone conduction
speech, and the derivation | leading-out speech by LMS algorithm, respectively.
It is the figure which showed the waveform of the derivation | leading-out speech by NLMS,
MLMS, RLS, and NNET algorithm in the case of adding white noise, respectively.
It is the figure which showed the waveform of the induction | guidance | derivation audio | voice
which added the clean air conduction speech, the noise (people crowd noise) addition, the bone
conduction speech, and the derivation | leading-out speech by LMS algorithm, respectively.
It is the figure which showed the waveform of the derivation | leading-out speech by NLMS,
MLMS, RLS, and NNET algorithm in the case of adding human congestion noise, respectively.
These are spectrograms of clean air conduction speech, air conduction speech with noise (white
noise) added, bone conduction speech, and speech derived by the LMS algorithm.
It is a spectrogram of the derivation | leading-out speech by NLMS, MLMS, RLS, and NNET
algorithm in the case of adding white noise.
These are spectrograms of clean air conduction speech, air conduction speech added with noise
(people crowd noise), bone conduction speech, and speech derived by the LMS algorithm.
It is a spectrogram of the derivation | leading-out speech by NLMS, MLMS, RLS, and NNET
algorithm in the case of adding crowd noise.
Explanation of sign
10-04-2019
19
[0088]
Reference Signs List 10 microphone 12 air conduction microphone 14 bone conduction
microphone 16 learning system 18 transversal filter circuit 20, 36 adder 22 filter coefficient
adjustment unit 26, 28 delay element 30, 32, 34 multiplier
10-04-2019
20
Документ
Категория
Без категории
Просмотров
0
Размер файла
30 Кб
Теги
jp2007251354, description
1/--страниц
Пожаловаться на содержимое документа