JPH05291971A - Signal processor - Google Patents

Abstract

PURPOSE: To obtain a signal processor which can reduce noise from conversation signals containing the noise. CONSTITUTION: A signal processor uses artificial intelligence 250 which can decide the adjustment of a filter subsystem 120 by distinguishing noise and conversation in the spectrum of input signals y(t) to which conversation signals s(t) and noise signals n(t) are added. The intelligence 250 decides that a part where the envelope abruptly changes indicates conversation and the remaining part indicates the frequency distribution of a noise output by checking the power of the frequency spectrum of input signals or pattern of an envelope function. The signal processor performs the above signal processing on all spectra or at every frequency band regardless of the presence/absence of the maximum value of the spectra.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a device for reducing noise, and more particularly to a device for reducing noise from a speech signal containing noise.

[0002]

2. Description of the Related Art U.S. Pat. No. 4,025,721 or 4,185,168
A conventional single microphone device for reducing noise mixed into speech as described in the specification is to check the minimum value of the envelope or average power of the input signal, which is the sum of speech and noise, and And the determination of the parameters of the input signal at its minimum value, which is considered to be the time when only interruptions or noise are present, these parameters were determined to be noise parameters.

[0003]

As mentioned above, these prior arts are limited both in the scope of their use and in the mode of realization, and they are limited to the use of analog arrays of bandpass filters. there were. The present invention has been made to solve the above problems.

[0004]

A signal processing apparatus according to the present invention selects a selection and control means for outputting a decision control parameter signal in response to an input signal in which a noise signal is added to a conversation signal, and an input signal. Filter means for reducing noise in response to the decision control parameter signal and the input signal to provide a filter output signal corresponding to the noise reduced input signal, the decision and control means comprising: Frequency subsystem means for deriving the frequency component from the input signal and supplying the output signal of each frequency component,
Energy subsystem means for deriving a power component of each frequency component in response to an output signal of the frequency component, and inputting, in response to the energy subsystem means, a rapid time variation that changes at a rate faster than a defined threshold rate. A comparator means for determining when the signal has, and in response to the comparator means, the energy subsystem means, and the input signal, determined to be changing at a rate faster than a defined threshold rate of the input signal Pattern classification subsystem means for selectively removing rapid temporal fluctuations to provide a residual output signal, the fluctuations representing fluctuations in the power component of the speech signal with respect to the frequency component over time, and the residual output signal with respect to the frequency component. The residual output signals corresponding to the power components of the noise signal and having different frequency components constitute the decision control parameter signal.

[0005]

According to the present invention, there is provided a signal processing device for reducing noise from a speech signal mixed with noise. The signal processing apparatus of the present invention uses artificial intelligence, which examines the power pattern function or envelope pattern function of the frequency spectrum of the input signal to find noise and noise in the spectrum of the input signal that is the sum of speech and noise. By distinguishing between conversations, it is decidable for the tuning of the filter subsystem, where the rapidly changing part of its envelope determines that it represents a conversation, while the rest is the frequency distribution of noise power. While ignoring where the maximum of the spectrum is, examine the entire spectrum or its frequency band.

In another embodiment of the invention, a feedback loop is provided which provides incremental adjustment to the filter by applying a gradient search process to attempt to increase certain interactive features in the device output signal. There is. The signal processing device according to the invention does not have to take into account the minimum value of the function of the input signal or the interruption of the conversation. Instead, the signal processor of the present invention uses an artificial intelligence system into which the envelope pattern of the speech and noise input signal is input. The signal processor of the present invention then filters from this envelope signal the rapidly changing variations of the envelope during a fixed time window.

[0007]

The invention will be better understood by reference to the detailed description in connection with the drawings. The signal processing device according to the invention does not have to take into account the minimum value of the function of the input signal or the interruption of the conversation. Instead, the signal processing apparatus of the present invention uses an artificial intelligence system to which the envelope patterns of the input signals of speech and noise (see FIG. 1) are input. This input or incoming signal is further described with respect to FIG. From this envelope signal, the signal processor of the present invention filters the rapidly changing variations of the envelope during a fixed time window. These rapidly changing variations are not necessarily maxima, as further explained with respect to FIG.

Rapidly changing fluctuations are those that last less than or equal to the duration of a certain predetermined time threshold. The envelope of the input signal is evaluated in various frequency bands. Alternatively, the Discrete Fourier Transform (DFT: D
The envelope of the escrete Fourier transform) is evaluated. The duration of the predetermined time threshold is different or FFT (DFT) due to different frequencies in case of multiple bands.
The artificial intelligence system then determines the envelope level of the envelope of the input signal thus filtered,
It represents the spectral level of noise over the appropriate band or over the separation frequencies considered in the DFT.

The input signal may consist of signal envelopes or, at the same time, of multiple envelopes for multiple bands or spectral levels. Each element of speech, the phoneme, has energy at different frequencies. These frequencies are in the Baltimore (Balt
imore) Williams and Wilkins
(Wilkins) in a book titled `` Use in Hearing Aid Assessment and Auditory Reassessment, '' published in 1977 by W Hodkin and R W Skinner. As such, it is often quoted.

A low frequency (about 1.2 kHz or less in the preferred embodiment) phoneme corresponding to a voice conversation is a high frequency (about 1.2 kHz or more in the preferred embodiment) phoneme (about 3 to 30 m) corresponding to a non-voice conversation.
Due to the fact that it has a much longer duration (about 40-150 ms) than a relatively short duration of seconds)
Different predetermined time threshold durations are used in different frequency bands.

The low and high frequency transition points selected for the preferred embodiment are below 1200 Hz and above 1200 Hz, respectively. Alternatively, other transition points may be chosen, for example 800, 1000 or 1500 Hz. Also, multiple transition points or sub-transition points may be chosen, each having an individual and isolated predetermined time threshold duration.

In the preferred embodiment, the duration of the predetermined time threshold is low frequency (below 1200 Hz) corresponding to voice conversation.
It is about 120 ms for the phoneme of. The duration of this predetermined time threshold may be in the range 100-150 ms. In the preferred embodiment, the predetermined time threshold duration is about 40 m for high frequency (1200 Hz and above) phonemes corresponding to non-voice conversations.
Seconds. The duration of this predetermined time threshold can be in the range of 25-40 ms.

Accordingly, those rapidly changing variations that last less than the duration of each predetermined time threshold are considered speech by the signal processing apparatus of the present invention. On the other hand, those rapidly changing variations that last longer than the duration of each predetermined time threshold are considered noise by the signal processing device of the present invention.

The signal processing apparatus of the present invention is characterized in that the past fluctuations in the input signal envelope at different frequencies or frequency bands move rapidly in time with the progress of time from one conversation to the next. Prove the fact that it is an envelope of ingredients. Also, the input signal will differ in frequency from one phoneme to the next in any normal conversation of any human language, while the noise to be removed by the signal processor of the invention will be around that frequency location. It is considered that the frequency position and intensity do not appear at such a rate but change in a frequency band at a certain frequency or a low rate.

Once the frequency content of the noise component of the input signal has been determined by the envelope filtering described above, the artificial intelligence subsystem (see controller subsystem 250 in FIG. 1) can assume one of four states. Will recognize.
That is, (a) no noise (noise lower than a certain level 3), (b) white noise (with respect to threshold level parameters at different frequencies or frequency bands as stored in the artificial intelligence recognizer subsystem). Noise having a substantially flat spectrum), (c) bubble noise (that is, an envelope component formed by mixing the phonemes of several speakers is an envelope component of a single speaker's conversation at a certain frequency position) In the background that lasts longer than the above, noise due to simultaneous conversation by several speakers) and (d) noise other than (a) to (c) above (i.e., in one or several frequency ranges). There are four categories: noise that is peak but not bubble noise.

Distinguishing the above four categories (a)-(d), the artificial intelligence system selects respective modes for filtering the input signal by the filter subsystem 230. The mode differs for each classification (a) to (d).

For classification (a), this filter is bypassed. For classification (b), the filter is set to adapt to the average speech state such that speech intelligibility is maximized while noise effects are minimal. This will suppress (notch) the lowest or highest frequency bands or edges of the spectrum, ie below about 400 Hz and above about 2.6 kHz. For classification (c), the filter is set to suppress (notch out) the low frequencies where most of the bubble energy is concentrated.

For classification (d), the filter is set so as to suppress (notch out) the frequency base where the envelope after filtering is maximum, and moderately suppress the band where the envelope is still relatively high. , Still reserve at least about 1/2 of the (logarithmic) total frequency. Furthermore, noting that speech intelligibility is highly concentrated at high frequencies (2000 kHz), when the artificial intelligence system determines that the noise to be suppressed is at frequencies below about 1500 Hz,
Bands above about 2000 Hz (up to 10-15 dB) are increased.

In one preferred embodiment, the filter subsystem 230 is an array of bandpass filters. Alternatively, the filter subsystem 230 could equally be implemented by a microcomputer system, digital signal processor, or FFT (Fast Fourier Transform) or DFT (Discrete Fourier Transform) IC or device. In fact, both the entire signal processing device of the present invention, the decision and control channel 20 and the filter channel 10, are implemented as a single microprocessor or DSP based device. Here, the microprocessor stores the envelope parameters of the input signal, analyzes each component, calculates a respective gain for each component, and responds to the stored parameters and in accordance with the teachings of the present invention, each component. Optimization is performed by adjusting the gain for.

In another embodiment of the signal processing device of the present invention (see FIG. 4), a feedback channel 30 (see FIG. 5) is incorporated in the noise reduction device, and the feedback channel 30 includes the speech component s (t ) Is divided into a high-frequency part and a low-frequency part by using a voice and non-voice discriminator based on a high-pass filter and a low-pass filter having a sharp cutoff frequency. All the output signals s ~ (t) of the noise reduction device (see Fig. 4 and Fig. 5) are input to the feedback channel 30, and the feedback channel 30 examines the output signal of the device to see if it is substantially conversational. To decide. Therefore, the presence of speech features of the speech and non-speech structure of the conversation is examined, both in the frequency content and duration of the phonetic and non-speech phonology of the conversation, respectively.

As a result, the output signal s (t) does not have the above characteristics of frequency content and associated duration over a time window (of the order of 100 to 150 msec), ie:
If the discriminator determines that the low-frequency phoneme continues for approximately 50 ms to 150 ms, then the high frequency indicated by Q
An (non-voice) internal signal is generated for a duration Tq within a predetermined time interval Tw. The ratio Tq / Tw is represented by Rq. Then, a gradient search implementation or circuit is incorporated into the feedback channel 30 to filter the main system filter subsystem (filter channel) 10 as in FIG. 4 or 5.
Of the gain parameter of s (t) is changed to a value within some predetermined compulsory range to reduce Rq, that is, to emphasize the interactive features of s (t), and thus to add more noise-free s at the output of the device.
get (t).

Returning to FIG. 1, the feedback channel 30
There is shown an electrical block diagram of the signal processing apparatus of the present invention without. The artificial intelligence pattern recognition noise reduction device for conversation processing as shown in FIG. 1 is a signal processing device that responds to an input signal y (t) 105. The input signal y (t) 105 is the conversation signal
It consists of the sum of s (t) and the noise signal n (t), and is obtained by adding the speech signal s (t) and the noise signal n (t) at the receiving source 100. The signal processor comprises a filter channel 10 and a decision and control channel 20, to each of which the input signal y (t)
105 is input.

The decision and control channel 20 receives the input signal y (t) 1
Means are provided for outputting a decision control parameter signal 260 in response to 05. The decision and control channel 20 includes a frequency subsystem 210, an energy subsystem 220, a filter subsystem 230, and a pattern classification subsystem 240.
And a pattern classification system comprising a controller subsystem 250.

The frequency subsystem 210 derives the frequency component of the input signal and outputs the output signal [y (f ₁ ), y of each frequency component.
It provides a means for supplying (f ₂ ), ... y (fn)]. The energy subsystem 220 is responsive to the output signal of the frequency components to generate energy components [|| y (f ₁ ) |
, || y (f ₂ ) ||, ... || y (fn) ||] is provided. Here, || y (fn) || indicates the absolute value of the magnitude of each frequency component. Energy subsystem 220 provides a power analyzer and may be implemented in a variety of different ways, such as a DFT power analyzer, an FFT analyzer, a squaring circuit with a smoothing circuit, and so on.

The pattern classification system includes a filter subsystem 230 for filtering time-varying peaks in || y ||
A pattern classification subsystem 240 for classifying noise from its frequency distribution and an artificial intelligence form of pattern recognition decision in accordance with the teachings of the present invention are used to determine gain adjustments at various frequencies (set gain vector, or filter Controller subsystem 2 for setting the parameters of
1 and as shown in FIG.

The pattern classification system selectively removes fast (or rapidly changing) time variations that are determined to be changing at a rate that is faster than a defined threshold rate of the input signal, so that the variation is at each frequency. Means are provided for providing a residual output signal representative of variations in speech signal power with respect to the components. Here, the residual output signal corresponds to the noise signal power for each frequency component, and the output signals at different frequency components constitute the decision control parameter signal 260.

The filter channel 10 further comprises a frequency subsystem 110 and a gain vector subsystem 120 that provides separate gain control over multiple frequency bands. Gain vector subsystem 120 acts as a filter subsystem within filter channel 10.

The frequency subsystem 110 derives the frequency component of the input signal and outputs the output signal [y (f ₁ ), y of each frequency component.
It provides a means for supplying (f ₂ ), ... y (fn)].

The filter channel 10 includes a decision control parameter signal by selectively filtering the input signal y (t) 105.
Noise reduction in response to 260 and the input signal 105 and corresponding to the noise reduced input signal
(t) Provide a means for supplying 140.

The gain vector subsystem 120 of the filter channel 10 is responsive to the decision control parameter signal 260 to selectively change the frequency response for each frequency component, and the gain of the output signal y (fn) of the frequency subsystem 110. It provides a means for adjusting the parameters.

Rapid time variations can be determined over a frequency range that includes all or part of the speech frequency spectrum. Rapid time variations can be determined over each frequency range, including frequency bands within the frequency spectrum of speech. The defined threshold rate is associated with the particular frequency component being processed. The energy function may be determined as the sample variation of each frequency component.

The frequency component of the input signal can be a discrete Fourier transform (DFT) parameter of the input signal. The decision and control channel 20 may also include a DFT analyzer subsystem or frequency subsystem 210 that selectively outputs DFT parameters for the input signal in response to the input signal.

Alternatively, the frequency components of the input signal can be determined by the subsystem with an array of bandpass filters responsive to the input signal. This array of bandpass filters simultaneously produces the frequency component output signals of the decision and control channel 20. Where frequency subsystem 110
Instead of, the output signal from each bandpass filter is
The filter channel 10 is then routed through each gain element of the gain vector subsystem 120 for each frequency band.
Is also passed. Here, the gain value is determined in response to the decision control parameter signal 260.

The gain of the gain vector subsystem 120 of the filter channel 10 is determined in the preferred embodiment in response to an artificial intelligence controller subsystem 250 in the decision and control channel 20. In one mode, the controller subsystem 250 causes the noise power to substantially respond to the decision to operate the white noise control mode where the gains at the high and low ends of the frequency range considered are suppressed. Decide In the preferred embodiment, the gains at the highest and lowest ends of the frequency range considered are suppressed to gain settings below 0.1 (-20 dB).

In another mode, the controller subsystem 250 operates in bubble noise mode. Where the noise power determined by the decision and control channel 20 is substantially higher at the low end of the frequency range for frequencies up to approximately 1000 Hz, and at the same time the noise power at the high end of the frequency range is determined to be non-zero, In response to the determination that power fluctuations in the high frequency range are determined to occur at a rate much higher than the predetermined rate normally associated with speech, the low frequency range of the filter is strongly suppressed, while the high frequency range is suppressed. Is slightly emphasized.

The decision and control channel 20 outputs a decision control parameter signal 260 via the controller subsystem 250. At this time, for example, it is determined that most of the noise power in the frequency range located below the predetermined highest frequency is substantially high, and there is a small amount of noise power above that frequency that is lower than the predetermined threshold level. And in response to the decision by the control channel 20, the high frequency gain is substantially increased, whereas the low frequency range of the noisy filter is strongly suppressed. Here, the controller subsystem 250 of the decision and control channel 20 determines that the noise is low frequency noise.

FIG. 2 shows the input signal and its component parts.
A voice receiver 100, such as the human ear or microphone, provides the sum of the incoming speech signal s (t) and the noise signal n (t). The output signal from the voice receiver 100 is the incoming signal or input signal y (t) 1.
05, and y (t) = s (t) + n (t).

3A to 3D show input signals at different times.
The frequency distribution of the envelope of y (t) is shown, and speech and noise are identified according to the power pattern of the input signal. 3A to 3D respectively show the frequency distribution of the envelope of the input signal y (t) at the successive times t ₁ , t ₂ , t ₃ and t ₄ .
3A to 3D show fluctuations that change rapidly at the position X ₁ .
(Peak) is stable at all times t _{1 to} t ₄ and therefore shows noise power, which, by contrast,
The peaks at X ₂ , X ₃ and X ₄ remain short (do not repeat over time samples), indicating the power of speech phonology.

FIG. 4 is an electrical block diagram of the signal processing apparatus of FIG. 1, which is an audio receiver which provides an input signal y (t) 105 applied to the input terminals of the decision and control channel 20 and the filter channel 10. Shown as 100 is the decision control parameter signal 260 of the decision and control channel 20 applying the gain control setpoint Gi to the filter channel 10 and the addition of the feedback channel 30.

The feedback channel 30 has a device output signal s˜ (t) 140 applied to its input terminal and provides an adaptive variation with respect to the gain setting of the filter channel 10 and the feedback channel 30. An output signal applied as a feedback signal to both
supply i.

FIG. 5 is an electrical block diagram of the feedback channel 30 of FIG. Feedback channel 30
Includes a bandpass filter subsystem 410, a decision subsystem 440, and a gradient search subsystem 450. The bandpass filter subsystem 410 is a highpass
An (HP) filter 420 and a low pass (LP) filter 430 are provided. The device output signals s to (t) 140 are applied to the respective input terminals of the high pass filter 420 and the low pass filter 430.

As mentioned above, the high pass filter 420 is
The low pass filter 430 provides an output signal responsive to the detection of unvoiced speech phoneme (UV), while the low pass filter 430 provides a voiced speech phoneme.
An output signal is supplied in response to the detection of (V). UV and V
Output signal is applied to an input terminal of the decision subsystem 440, which, in accordance with the teachings of the present invention, responds to the determination of the duration of each V and UV output signal corresponding to voiced and unvoiced phonemes. And output signal Q is supplied.
The output signal Q is applied to the input terminals of the slope search subsystem 450, which in accordance with the teachings of the present invention, provides an output signal ~ G i460 that provides a signal for varying the gain setting of the filter channel 10. To supply.

The output signal ~ G i460 is also applied to the gradient search subsystem 450 as a feedback signal. or,
Random initialization parameters ~ Gi (o) 452 initialization values are used as another initial input signal for the gradient search subsystem 450.
Is supplied to.

Although various specific embodiments have been described herein, it will be appreciated by those skilled in the art that other different embodiments can be implemented in accordance with the teachings of the present invention. Accordingly, the scope of the invention is not meant to be limited to the disclosed embodiments, but rather is defined by the appended claims.

[0045]

As described above, according to the present invention, the noise in the spectrum of the input signal in which the speech and the noise are added and the speech are distinguished from each other by using the artificial intelligence capable of adjusting and deciding the filter subsystem to determine the frequency spectrum. Of the power or envelope function of the envelope to determine that the rapidly changing part of the envelope is indicative of speech, and the remaining part is the frequency distribution of the noise power, with or without the maximum of the spectrum. Since the processing is performed for the entire spectrum or for each frequency band, it is possible to obtain a signal processing device that reduces noise from a conversation signal containing noise.

The invention also incorporates a feedback loop that provides incremental adjustments to the filter by using a gradient search process to augment certain interactive features in the output signal of the device. Artificial intelligence into which the envelope pattern of the input signal is input allows to filter out rapid fluctuations during a fixed time window without having to consider the minimum of the function of the signal, i.e. the interruption of the conversation. Therefore, it is possible to obtain a signal processing device that reduces noise from a conversation signal mixed with noise.

[Brief description of drawings]

FIG. 1 is an electrical block diagram showing a signal processing device of the present invention without feedback.

FIG. 2 is an explanatory diagram showing an input signal and its component parts.

FIG. 3 is a waveform diagram showing an envelope of an input signal at successive times.

FIG. 4 is an electrical block diagram showing the signal processing apparatus of FIG. 1 with a feedback channel added.

5 is an electrical block diagram illustrating the feedback channel of FIG.

[Explanation of symbols]

 10 Filter channel 20 Decision and control channel 30 Feedback channel 105, y (t) input signal 110, 210 Frequency subsystem 120 Gain vector subsystem 140 Filter output signal 220 Energy subsystem 250 Controller subsystem 260 Decision control parameter signal 410 Bandpass Filter subsystem 420 High-pass filter 430 Low-pass filter 440 Decision subsystem 450 Gradient search subsystem 460, Gi Gain control setpoint s (t) Speech signal n (t) Noise signal

Claims (22)

[Claims] 1. A decision and control means for outputting a decision control parameter signal in response to an input signal in which a noise signal is added to a conversation signal, and the decision control parameter signal and the decision control parameter signal for selectively filtering the input signal. Filter means for reducing the noise in response to an input signal and supplying a filter output signal corresponding to the noise-reduced input signal, wherein the determining and controlling means determines the frequency from the input signal. Frequency subsystem means for deriving a component to supply an output signal of each frequency component; energy subsystem means for deriving a power component of each frequency component in response to the output signal of the frequency component; and the energy subsystem. Responsive to the means for determining when the input signal has rapid time variations that change at a rate faster than a defined threshold rate A comparator means, the comparator means, the energy subsystem means, and the rapid rate determined to be responsive to the input signal at a rate faster than the defined threshold rate of the input signal. Pattern classification subsystem means for selectively removing a time variation to provide a residual output signal, said variation representing a variation over time of a power component of said speech signal with respect to said frequency component, said residual output The signal corresponds to the power component of the noise signal with respect to the frequency component, and the residual output signal at different frequency components is
A signal processing device that constitutes the decision control parameter signal. 2. The filter means further comprises adjusting means for adjusting a gain parameter of the filter means in response to the decision control parameter signal to selectively change the frequency response of the filter means for each frequency component. The signal processing device according to claim 1, wherein the adjusting unit adjusts the gain parameter for each frequency component in response to the residual output signal for each frequency component. 3. The signal processing device according to claim 1, wherein the rapid time variation is determined over a frequency range including a frequency spectrum of a conversation including all frequency components. 4. The signal processing apparatus of claim 1, wherein the power component is determined at each frequency component as a finite sum of squared individual time samples of the input signal. 5. The frequency component of the input signal is a transform parameter of a discrete Fourier transform (DFT) of the input signal, and the determining and controlling means is responsive to the input signal, the DFT parameter for the input signal. The signal processing apparatus according to claim 1, further comprising a DFT analyzer subsystem that selectively outputs the signal. 6. The signal processing apparatus of claim 1, wherein the frequency subsystem means comprises an array of bandpass filters responsive to the input signal No. 3. 7. The array of bandpass filters simultaneously generates output signals of the frequency components of the decision and control means, the output signals from each bandpass filter comprising:
7. A signal processing apparatus as claimed in claim 6, which is subsequently passed through the filter means via respective gain elements for each frequency band, the gain value being determined in response to the decision control parameter signal. 8. The signal processing apparatus according to claim 1, wherein the rapid time variation is determined over a frequency range each including a frequency band of 100 Hz to 10,000 Hz. 9. The gain parameter of the filter means is determined in response to an artificial intelligence subsystem means in the decision and control means, the decision and control means determining when the noise power is over the entire frequency range considered. 3. The signal processing apparatus according to claim 2, wherein the gain parameters at the highest end and the lowest end of the considered frequency range are suppressed by determining whether they are substantially equal and operating a white noise control mode in response to the determination. .. 10. The noise power determined by the determining and controlling means is substantially high at the low end of the frequency range for frequencies up to about 1000 Hz, and at the same time the noise power is determined to be non-zero at the high end of the frequency range. And in response to the change in the power component in the high frequency range being determined to be significantly faster than a predetermined rate of change associated with normal speech, the decision control means activates a bubble noise mode, At least one low frequency range of the filter is strongly suppressed,
The signal processing device according to claim 1, wherein at least one high frequency range of the filter is amplified. 11. The decision and control means outputs the decision control parameter signal so that a gain parameter in a high frequency component is substantially amplified, and most of power components of noise fall below a predetermined maximum frequency. 3. In response to the decision by the determining and controlling means to be located, the gain parameter at low frequency components is strongly suppressed and the determining and controlling means determines that the noise is low frequency noise. Signal processing device. 12. The decision and control channel determines that the noise is high frequency noise and, in response to the determination that the noise power component is above a predetermined high frequency range, determines an appropriate frequency range in which the noise is present. The signal processing device according to claim 1, wherein the signal processing device is strongly suppressed. 13. The reduction of the gain parameter is reduced to less than 1, and when the gain parameter of the filter means over a frequency range is to be suppressed, the suppression is gradual and over a time interval of 1 second or less. 11. The signal processing device of claim 10 that occurs smoothly. 14. The signal processing apparatus of claim 11, wherein the amplification is performed gradually over a time interval of 1 second or less when the gain parameter of the filter means over a frequency range is to be increased. 15. The determination and control means determines a frequency range in which the noise power has a maximum value, and the reduction of the filter output signal is the highest with respect to the determined maximum frequency range. Signal processing equipment. 16. An audio signal coupled to receive an output signal of a filter channel, the audio signal including a high-pass sub-filter and a low-pass sub-filter having a sharp cutoff for measuring an output level at a frequency above a predetermined threshold frequency. And a feedback channel including a non-voice discrimination circuit, and signal power at each output terminal of the high-pass sub-filter and the low-pass sub-filter in response to the feedback channel during a predetermined time window (Tw) of the order of 300 msec. Provides an output signal Q in response to a determination that it is mostly within the frequency range of the high-pass sub-filter at a level above a predetermined level for a time longer than a second predetermined time interval, and said first signal Until it is determined that the power of the signal has dropped below the predetermined level during the time window Tw of A decision subsystem that continues to provide the output signal until the end of the time window Tw of, and the power at the low-pass sub-filter is at a second predetermined time period that is longer than the second predetermined time interval. An output signal Q is output in response to a determination that the power level of both the high-pass sub-filter and the low-pass sub-filter overlaps and simultaneously exceeds a threshold level, An output signal Q is output for the duration of the overlap of the threshold level, the power levels of both the high-pass sub-filter and the low-pass sub-filter in a time window, and the duration of the output signal of level Q, denoted by Tq, The ratio to the window length indicated by Tw, that is, the ratio Tq / Tw = Rq, is repeatedly calculated for each window Tw, The gain parameter in each range of frequency of the filter means is slightly changed, whereby the slope ratio between the variation of Rq and the variation of each parameter is calculated so that the slope search serves as slope search feedback and reduces Rq. In order to modify the gain of the filter means in order to provide a recursive slope search in the direction of reducing Rq, the latter variation of the gain of the filter channel is due to the decision and control means without consideration of the feedback channel. The influence of feedback correction is limited by limiting the gain value as determined within a predetermined percentage ratio, and the gain Gi for the i-th frequency range (where i is a continuous integer, i = 1, 2, ... N, where N is the total number of frequency ranges considered) and the slope relationship between Rq and various gays over very small increments over a given time interval Tq. By applying the, also, before such time interval T
Updated by comparing Rq over q and its variations, the time interval Tq is not necessarily equal to Tw, and the slope function is represented by Rq / Gi; i = 1,2, ... N and the time interval Tq (j ),
The signal processing device according to claim 1, wherein the j-th integer time interval j is represented by j = 0, 2 ,. 17. The j-th time interval T, denoted by Gi (j)
The correction variation Gi (j) of Gi between q (j) and the preceding time interval Tq (j−1) is a recursive relationship: Gi (j) = B · Rq / Gii = 1,2 , ... N, i is the frequency range to be considered, B is a predetermined coefficient, and Gi (j) representing the sum over j is not considered when ignoring the feedback channel,
17. The signal processing device according to claim 16, wherein a predetermined threshold ratio for Gj as determined by the determination and control means is not exceeded. 18. The reduction of the filter output signal in the low frequency range is responsive to the determination that the noise power component is present in both the predetermined high frequency range and the low frequency range and the determination of the control means. 21. The signal processing apparatus of claim 20, wherein the signal output apparatus is made larger than the filter output signal of. 19. The signal processing device according to claim 15, wherein the reduction of the filter output signal is smaller than the maximum reduction for a frequency range other than the determined range. 20. The signal processing device of claim 15, wherein the maximum reduction of the filter output signal is a greater value for low frequencies. 21. Means for reducing the filter output signal only in the high frequency range and the low frequency range in response to the speech signal in response to determining that the distribution of the noise component is white noise. 19. The signal processing device according to claim 18. 22. The signal of claim 18, further comprising means responsive to the determination that the distribution of the noise component is a bubble to reduce the filter output signal only in the low frequency range in response to the speech signal. Processing equipment.