KR20040030638A - Detecting voiced and unvoiced speech using both acoustic and nonacoustic sensors - Google Patents

Abstract

A system and method are provided for detecting voiced and unvoiced speech in an acoustic signal having several levels of background noise. This system (FIG. 3) receives acoustic signals at two microphones Mic1 and Mic2, generating a difference parameter between the acoustic signals received at each of the two microphones Mic1 and Mic2. The difference parameter represents the relative difference in signal gain between portions of the received acoustic signal. The system classifies the acoustic signal information as unvoiced speech when the difference parameter exceeds the first limit, and when the difference parameter exceeds the second limit, the system classifies the acoustic signal information as voiced speech. Moreover, embodiments of the system according to the invention include a non-acoustic sensor 20 that receives physiological information to assist in voiced speech classification.

Description

System and method for detecting voiced and unvoiced sound using acoustic and non-acoustic sensors {DETECTING VOICED AND UNVOICED SPEECH USING BOTH ACOUSTIC AND NONACOUSTIC SENSORS}

The ability to accurately distinguish between voiced and unvoiced speech is important for many speech applications, including speech recognition, speaker identification, noise suppression, and more. In a typical acoustic device, speech from the speaker is captured and sent to a receiver in another location. One or more noise sources may be present in the speaker environment that contaminate the speech signal or the signal of interest with unwanted acoustic noise. This may make it difficult or impossible for the receiver or receiver to find the user speech.

Typical methods for classifying voiced and unvoiced speech rely primarily on the acoustic content of microphone data. This approach suffers from noise problems, and there is uncertainty in the signal content. This becomes even more problematic as mobile communication devices such as cellular telephones and PDAs become widespread. This is because, in most cases, the quality of service provided by these devices depends on the quality of voice service provided by this device. Although methods for suppressing noise present in speech signals are known in the art, these methods have performance drawbacks. In other words, the operation time is long, annoying hardware is required to perform signal processing, and the signal of interest is distorted.

The present invention relates to speech signals processing.

1 is a block diagram of a NAVSAD system according to one embodiment of the invention.

2 is a block diagram of a PSAD system according to one embodiment of the invention.

3 is a block diagram of a noise cancellation system called a Pathfinder system, in accordance with an embodiment of the invention.

4 is a flowchart of a detection algorithm for use in detecting voiced and unvoiced speech, in accordance with an embodiment of the invention.

FIG. 5A is a graph showing a received GEMS signal 502 for speech, with an average correlation 504 between the GEMS signal and the Mic1 signal, and the limit T1 used for voiced speech detection. FIG.

FIG. 5B is a graph showing the received GEMS signal 502 for speech, along with the standard deviation 506 of the GEMS signal, and the limit T2 used for voiced speech detection.

FIG. 6 is a graph showing voiced speech 602 sensed from an acoustic or audio signal 608 along with a GEMS signal 604 and acoustic noise 606.

7 is a microphone array diagram for use under one embodiment of a PSAD system.

8 is a graph of ΔM versus d 1 for several Δd values under one embodiment.

9 is a graph of gain parameters as the sum of the acoustic data or audio from the microphone 1 and the absolute value of H1 (z).

10 is an alternative graph of the acoustic data shown in FIG. 9;

A system and method for distinguishing voiced and unvoiced speech from background noise are provided below. That is, a non-acoustic sensor voiced speech activity detection (NAVSAD) system and a pathfinder speech activity detection (PSAD) system are provided. The noise reduction and reduction method provided herein overcomes the disadvantages of systems typical of the art by cleaning acoustic signals of interest without distortion, while distinguishing and classifying unvoiced and voiced speech from background noise.

1 is a block diagram of a NAVSAD system 100 in accordance with one embodiment of the invention. The NAVSAD system connects the microphone 10 and sensor 20 to one or more processors 30. Sensor 20 in one embodiment includes a voice activity sensor or an unvoiced sensor. The processor 30 controls the subsystem including the sensing subsystem 50, called the sensing algorithm, and the noise canceling subsystem 40. The operation of the noise canceling subsystem 40 is described in detail in the relevant apparatus section. The NAVSAD system works very well in any background acoustic noise environment.

2 is a block diagram of a PSAD system 200 according to one embodiment of the invention. The PSAD system couples the microphone 10 to one or more processors 30. The processor 30 includes a sensing subsystem 50 called a sensing algorithm and a noise canceling subsystem 40. PSAD systems are very sensitive in low acoustic noise environments and relatively less sensitive in high acoustic noise environments. The PSAD can operate independently or as a backup to NAVSAD. The backup function detects voiced speech when NAVSAD fails.

The sensing subsystem 50 and the noise canceling subsystem 50 of the NAVSAD and PSAD systems according to one embodiment of the invention are algorithms controlled by the processor 30, but are not limited thereto. Alternative embodiments of the NAVSAD and PSAD systems may include the sensing subsystem 50 or the noise canceling subsystem 40, which consists of additional hardware, firmware, and software, and combinations thereof. Moreover, the functions of sensing subsystem 50 and noise canceling subsystem 40 may be distributed among the various components of the NAVSAD and PSAD systems.

3 is a block diagram of a noise cancellation subsystem 300 called a pathfinder in accordance with one embodiment of the invention. The pathfinder system is briefly described below and described in detail in the relevant apparatus section. Two microphones Mic1 and Mic2 are used in the Pathfinder system, and Mic1 is considered a "signal" microphone. Referring to FIG. 1, when the voice activity detector (VAD) 320 is a non-acoustic voiced sensor 20 and the noise canceling subsystem 340 includes a sensing subsystem 50 and a noise canceling subsystem 40. The pathfinder system 300 is equivalent to the NAVSAD system 100. Referring to FIG. 2, when the VAD 320 is absent and the noise canceling subsystem 340 includes the sensing subsystem 50 and the noise canceling subsystem 40, the pathfinder system 300 is the PSAD system 200. Equivalent to

NAVSAD and PSAD systems support two commercial approaches. That is, relatively inexpensive PSAD systems support acoustic approaches that function from the lowest to the middle noise environments, while the NAVSAD system adds non-acoustic sensors to detect voiced speech in any environment. In general, unvoiced speech is not detected using the sensor. This is because unvoiced speech does not sufficiently vibrate a person's tissue. However, it is not important to detect unvoiced speech in high noise environments. This is because the energy of unvoiced speech is very small and easily washed away by noise. Thus, in high noise environments, unvoiced speech has little effect on voiced speech noise cancellation. Therefore, unvoiced speech information is most important in the absence of almost no noise, so unvoiced speech detection should be very sensitive in low noise environments and insensitive in high noise situations. This is not easily achieved and comparable acoustic unvoiced detectors known in the art cannot operate under such environmental constraints.

NAVSAD and PSAD systems include an array algorithm for speech sensing that uses the frequency content difference between two microphones to compute the relationship between the signals of the two microphones. This contrasts with existing arrays that attempt to use the time / phase difference of each microphone to remove noise outside the “sensitivity area”. The methods introduced here offer significant advantages as they do not require a specific orientation of the array for the signal.

Moreover, while existing arrays rely only on specific noise directions, the systems introduced here are sensitive to noise of all kinds and in all directions. As a result, the frequency-based array presented here is unique. This is because it depends only on the relative direction of the two microphones, without depending on the direction of the signal and the noise with respect to the microphone. This results in a robust signal processing system with respect to noise type, microphone, and direction between the noise source / signal source and the microphone.

The systems introduced here use information from non-acoustic sensors or pathfinder noise suppression systems described in the relevant device paragraphs to determine the vocalization of the input signal. Voicing states include silent, voiced and unvoiced states. For example, the NAVSAD system includes a non-acoustic sensor to sense vibrations of human tissue related to speech. The non-acoustic sensor according to an embodiment is a general electromagnetic movement sensor (GEMS), but is not limited thereto. However, alternative embodiments may use any sensor that can detect human tissue movement related to speech while being immune to environmental acoustic noise.

GEMS is an RF device (2.4GHz) that can detect the movement of the dielectric interface of human tissue. GEMS is an RF interferometer that uses homodyne mixing to detect small phase excursions related to target motion. In essence, the sensor emits weak electromagnetic waves (less than 1 milliwatt) that reflect anything around the sensor. The reflected wave is mixed with the original transmission wave and the result is analyzed for the target position change. Anything moving around the sensor will change the phase of the reflected wave, which will be amplified and displayed by the voltage output from the sensor. A similar sensor is introduced in Gregory C. Burnett's 1999 Thesis at the University of California, Davis, "The Physiological Basis of Glottal Electromagnetic Micropower Sensors (GEMS) and Their Use in Defining an Excitation Function for the Human Vocal Tract". .

4 is a flow diagram of a detection algorithm 50 for use in detecting voiced and unvoiced speech in accordance with one embodiment of the invention. Referring to FIGS. 1 and 2, the NAVSAD and PSAD systems include a sensing algorithm 50 as the sensing subsystem 50. This sensing algorithm 50 operates in real time and, in one embodiment, operates with a 20 millisecond window and steps, but is not limited to, 10 milliseconds at a time. Voice activity decisions are recorded for the first 10 milliseconds, and the next 10 milliseconds serve as a "look-ahead" buffer. While one embodiment uses 20/10 windows, alternative embodiments may use many other combinations of window values.

In developing the sensing algorithm 50, a number of multidimensional elements may be considered. The most painstaking must be to maintain the effectiveness of the pathfinder noise canceling technique described in the related technical section. Pathfinder performance can be compromised if adaptive filter training is performed on speech rather than noise. Therefore, it is important not to exclude a significant amount of speech from the VAD (uttering activity determination) while keeping this confusion to a minimum.

Care should also be taken in the accuracy of the characterization between voiced and unvoiced speech signals, and in distinguishing each of these speech signals from the noise signal. This kind of characterization can be useful for devices such as speech recognition and speaker identification.

Moreover, systems using detection algorithms in accordance with one embodiment of the invention function in an environment that includes varying amounts of background acoustic noise. If a non-acoustic sensor is available, external noise is not a problem for voiced speech. However, in the case of unvoiced speech (and in the case of voiced speech when a non-acoustic sensor is not available or malfunctioning), a dependency is placed on the acoustic data itself to separate the noise from the unvoiced speech. In one embodiment of the pathfinder noise suppression system, there are advantages when using two microphones, and the spatial relationship between the microphones is used to help detect unvoiced speech. However, it can often be that there is a level of noise that is large enough for speech to be almost undetectable and the acoustic-only method fails. In this situation, a non-acoustic sensor (hereafter just a "sensor") will be required to ensure good performance.

In a two-microphone system, the speech source should be relatively loud in one microphone than in one microphone. As any noise had to cause H1 with unit-level gain, the test showed that the existing microphone could easily meet this requirement when the microphone was in the head.

Referring to the NAVSAD system and FIGS. 1 and 3, NAVSAD relies on two parameters for voiced speech detection. These two parameters are related to the sensor energy of the window of interest determined in one embodiment by the standard deviation (SD), and in addition, the cross-correlation (XCORR) between the acoustic signal and the sensor data from the microphone 1 and the sensor data. Include. Sensor energy can be determined in one of a number of ways, and SD is the only convenient way to determine energy.

In the case of a sensor, the standard deviation (SD) is close to the signal energy corresponding to some degree of accuracy in the vocal state, but is susceptible to motion noise (relative movement of the sensor to the user) and electromagnetic noise. To further separate sensor noise from tissue movement, XCORR can be used. XCORR is calculated with 15 delays, which corresponds to an operation under 2 milliseconds at 8000 Hz.

XCORR may be useful when the sensor signal is distorted or modulated in some way. For example, there is a sensor location (eg, behind the jaw or neck) where speech generation may be detected but the signal may have inaccurate or distorted time-based information. That is, they may not have features and shapes that match the acoustic waveform (well formed in time). However, XCORR is more susceptible to errors from acoustic noise, and XCORR is of little use in high acoustic noise environments. Therefore, XCORR should not be the only source of speech information.

The sensor senses human tissue movement associated with the closure of the vocal fold, so that the acoustic signal produced by the vocal closure is highly correlated to the closure. Therefore, sensor data highly correlated to the acoustic signal is classified as speech, and sensor data not highly correlated is classified as noise. As a result of the delay time caused by the relatively slow sound speed (330m / s), acoustic data is expected to lag behind sensor data by 0.1-0.8 milliseconds (or about 1-7 samples). However, one embodiment uses a 15 sample correlation because the acoustic wave waveform varies greatly depending on the sound produced, and a larger correlation width is needed to ensure detection.

Although standard deviation and XCORR signals are involved, they differ enough from each other so that voiced speech detection is reliable. Nevertheless, for simplicity, either parameter can be used. The values for standard deviation and XCORR are compared with the experimental limit, and if the two values are greater than the limit, they are classified as voiced speech. An example data is presented and introduced below.

5A, 5B, and 6 are graphs of examples in which the subject sounds twice the phrase “pop pan” according to one embodiment. FIG. 5A is a graph showing the received GEMS signal 502 for this utterance along with the average correlation 504 between the GEMS signal and the Mic1 signal and the limit T1 used for voiced speech detection. 5B is a graph showing the received GEMS signal 502 for this utterance, along with the standard deviation 506 of the GEMS signal, and the limit T2 used for voiced speech detection. FIG. 6 is a graph showing voiced speech 602 sensed from acoustic or audio signal 608 along with GEMS signal 604 and acoustic noise 606. At this time, no unvoiced speech is detected in this example due to the enormous background crosstalk noise 606. The limits are set such that there is no error negative value at all and only an error positive value exists intermittently. Over 99% voiced speech activity detection accuracy is achieved under any acoustic background noise condition.

NAVSAD can determine when voiced speech is occurring with a high degree of accuracy due to non-acoustic sensor data. However, sensors rarely help in separating unvoiced speech from noise. This is because unvoiced speech does not cause a detectable signal in most non-acoustic sensors. NAVSAD can be used if there is a detectable signal, but the use of the standard deviation method is indicated because the degree of correlation of unvoiced speech is poor. In the absence of a detectable signal, the system and method of the pathfinder noise cancellation algorithm is used to determine when unvoiced speech is occurring. The pathfinder algorithm is briefly introduced below, and a detailed description is provided in the relevant device section.

Referring to FIG. 3, the acoustic information flowing into the microphone 1 is represented by m1 (n), the information flowing into the microphone 2 is represented by m2 (n), and the GEMS sensor determines the voiced speech area. It is assumed to be available. In the z (digital frequency) domain, these signals are represented by M1 (z) and M2 (z). next,

M1 (z) = S (z) + N2 (Z)

M2 (z) = N (z) + S2 (Z)

Where N2 (z) = N (z) H1 (z)

S2 (z) = S (z) H2 (z)

Thus M1 (z) = S (z) + N (z) H1 (z)

N2 (z) = N (z) + S (z) H2 (z) (1)

This is a common case for a two-microphone system. There is some noise leakage with Mic1 and some signal leakage with Mic2. Equation 1 has only four unknowns and two relationships, so no solution can be obtained.

However, there are also ways to obtain solutions to some unknowns in equation 1. Investigate the case where no signal is being generated. In other words, consider when the GEMS signal indicates that no speech is occurring. In this case, s (n) = S (z) = 0 and equation 1 is

M1n (z) = N (z) H1 (z)

N2n (z) = N (z)

Is reduced to. At this time, the n subscript in the M variable indicates that only noise is being received. Because of this,

M1n (z) = M2n (z) H1 (z)

H1 (z) = M1n (z) / M2n (z) (2)

H1 (z) can be computed using the microphone output and available system identification algorithm when only noise is being received. This operation can be made adaptively so that H1 (z) can be quickly recomputed if the noise changes significantly.

As a solution to one of the unknowns in Equation 1, the amplitude of GEMS or similar device, along with the amplitudes of the two microphones, can be solved for another value H2 (z). If GEMS shows utterance but the recent (less than 1 second) history of the microphone shows a low level of noise, it is assumed that n (s) = N (z) to 0. Equation 1 can be reduced to

M1s (z) = S (z)

M2s (z) = S (z) H2 (z)

As a result,

M2s (z) = M1s (z) H2 (z)

H2 (z) = M2s (z) / M1s (z)

This is the inverse of the H1 (z) operation value. However, it should be noted that several different inputs are used.

After the above calculation of H1 (z) and H2 (z), they are used for noise removal from the signal. Equation 1 can be rewritten as

S (z) = M1 (z)-N (z) H1 (z)

N (z) = M2 (z)-S (z) H2 (z)

S (z) = M1 (z)-[M2 (z)-S (z) H2 (z)] H1 (z)

S (z) [1-H2 (z) H1 (z)] = M1 (z)-M2 (z) H1 (z)

Solving for S (z) gives the following results:

S (z) = (M1 (z)-M2 (z) H1 (z)) / (1-H2 (z) H1 (z))

In fact, since H2 (z) is very small, H2 (z) H1 (z) << 0, thus,

S (z) = M1 (z)-M2 (z) H1 (z)

The PSAD system is described with reference to FIGS. 2 and 3. As sound waves propagate, they lose energy due to diffraction and scattering. Assuming that sound waves originate from the point source and radiate isotropically, its amplitude will decrease as a function of 1 / r, where r is the distance from the point source. Since the reduction will be small if limited to a small area, this function proportional to 1 / r of amplitude is the worst case. However, this is a suitable model for the structure of noise and speech propagation into a microphone somewhere above the user's head.

7 is a diagram of a microphone array according to one embodiment of a PSAD system. By placing the microphones Mic1, Mic2 on the array midline with the mouth and linear array, the signal strength difference between Mic1 and Mic2 will be proportional to d1 and Δd. Assuming a 1 / r relationship (or 1 / d in this case), the following results can be obtained.

ΔM = | Mic1 | / | Mic2 | = ΔH1 (z) ∝ (d1 + Δd) / d1

DELTA M is a gain difference between Mic1 and Mic2 and, therefore, H1 (z) in equation (2). The variable d1 is the distance from Mic1 to the speech or noise source.

8 is a graph 800 of d1 versus ΔM for various Δd values according to one embodiment of the invention. The larger Δ and the closer the noise source, the larger ΔM. The variable Δd will vary with the azimuth angle for the speech / noise source from the maximum on the array midline to zero perpendicular to the array midline. It can be seen from the graph 800 that ΔM is close to the unit value for a distance of Δd being small and approximately 30 cm or more. Since most noise sources are located farther than 30 cm and do not lie on the midline of the array, when calculating H1 (z) as in Equation 2, ΔM (or gain of H1 (z)) will be close to unit values. Conversely, for a nearby noise source (within a few centimeters), the gain difference may appear depending on whether the microphone is closer to noise.

If "noise" is the speaker's speech and Mic1 is closer to the mouth than Mic2, the gain increases. Since environmental noise typically originates farther from the user's head than speech, noise will be found when the gain of H1 (z) is near unit or some fixed value, and speech can be found after a sharp increase in gain. . Speech can be unvoiced or voiced as long as it has sufficient volume relative to ambient noise. The gain will remain somewhat high during the speech portion, and then decreases rapidly after the speech is stopped. The rapid increase and decrease of the H1 (z) gain should be sufficient to detect speech under almost any situation. The gain of this example is calculated by the sum of the absolute values of the filter coefficients. This sum is not equal to gain, but the two values have a relationship in which the increase in the sum of the absolute values reflects the increase in gain.

As an example of this behavior, FIG. 9 shows the gain parameter 902 and the acoustic data 904 from the microphone 1 or the graph 900 of audio as the sum of the absolute values of H1 (z). The speech signal is utterance when the phrase "pop pan" is repeated twice. The bandwidth covers a frequency range of 2500 to 3500 Hz, with an additional bandwidth of 1500 to 2500 Hz. The gain increased rapidly when unvoiced speech appeared first, and then quickly returned to normal value when speech terminated. Significant gain changes resulting from transitions between noise and speech can be detected by standard signal processing techniques. The standard deviations of the last few gain operations are used and the limits are defined by the standard deviation noise floor and the operating average of the standard deviations. Later gain variations for voiced speech are suppressed in this graph 900 for simplicity.

FIG. 10 is an alternative graph 1000 of acoustic data shown in FIG. 9. The data used to form the graph 900 is presented again in this graph 1000 and presented together with the acoustic data 1004 and GEMS data 1006 without noise so that unvoiced speech appears. The speech signal 1002 may have three values. Noise is 0, voiceless is 1, voiced sound is 2. Noise rejection is only achieved when V = 0. Apart from the two single dropouts of unvoiced detection near the end of each "pop", it is clear that unvoiced speech is easily captured. However, this single-window dropout is not common and does not significantly affect the noise cancellation algorithm. These can be easily removed using standard smoothing techniques.

What is not clear from the graph 1000 is that the PSAD system functions as an automatic backup for the NAVSAD system. This is because if the sensor or NAVSAD system fails for some reason, voiced speech (since it has the same relationship as unvoiced to the microphone) will be detected as unvoiced. Voiced speech will be misclassified as unvoiced, but noise cancellation will still not occur, preserving the quality of the speech signal.

However, for NAVSAD systems, this automatic backup works best in low noise environments (approximately 10 + dB SNR). This is because high amounts of acoustic noise (less than 10 dB SNR) directly overwhelm any acoustic-only unvoiced detector, including the PSAD. This is evident in the difference between the vocal signal data 602 and 1002 shown in the graphs 600 and 1000 of FIGS. 6 and 10, where the same utterance occurs but the graph 600 data shows no unvoiced speech. Do not give. Because you can't detect unvoiced speech. This is the desired behavior when running noise reduction. If unvoiced speech cannot be detected, it will not significantly affect the noise rejection process. Detecting unvoiced speech using a Pathfinder system ensures the detection of unvoiced speech large enough to distort noise rejection.

Looking at the hardware considerations with reference to Figure 7, the structure of the microphone can have an effect on the gain required for speech detection and the limits related to speech. In general, each structure will require testing to determine the appropriate limits, but tests using two very different microphone structures have shown that work is normally done with the same limits and different parameters. The first set of microphones has a noise microphone a few centimeters away from the ear and a nearby signal microphone. The second structure also places the noise and signal microphones back within a few centimeters at the mouth. The results presented here were derived using the first microphone structure, but the results using the remaining sets are also substantially the same, so that the detection algorithm is relatively robust for microphone placement.

Multiple structures are possible using the NAVSAD and PSAD systems to detect voiced and unvoiced speech. One architecture uses the NAVSAD system for voiced speech detection along with the PSAD system for unvoiced speech detection. PSAD also acts as a backup to the NAVSAD system for voiced speech detection. An alternative architecture uses a NAVSAD system for voiced speech detection in conjunction with a PSAD system for unvoiced speech detection. PSAD also acts as a backup to the NAVSAD system for voiced speech detection. Another alternative structure can detect voiced and unvoiced speech using a PSAD system.

Although the above systems have been described listening to distinguish voiced and unvoiced speech from background acoustic noise, there is no reason why a more complex classification cannot be made. When you characterize speech more deeply, the system can bandpass information from Mic1 and Mic2 so that you can see which band of Mic1 data consists of more noise and which band consists of more speech. . Using this knowledge, it is possible to group utterances by spectral characteristics similar to existing acoustic methods. This method works better in noisy environments.

As an example, "k" of "kick" has a frequency content between 500 and 4000 Hz, while "sh" of "she" has only energy from 1700-4000 Hz. Voiced speech can also be classified in a similar manner. For example, / i / ("ee") has an energy of 300-2500 Hz and / a / ("ah") has an energy of 900-1200 Hz. Thus, this ability to distinguish between voiced and unvoiced speech in the presence of noise is very useful.

Each of the steps depicted in the flowchart presented herein may itself include a series of actions that need not be introduced herein. Those skilled in the art will be able to create routines, algorithms, source code, microcode, program logic arrays, and implement the invention based on the flowcharts and detailed description presented herein. Routines introduced herein may be provided with one or more of the following, or a combination of one or more of the following: That is, stored in non-volatile memory that forms part of the associated processor, or implemented using existing programmable logic arrays or circuit elements, or stored on removable media, such as disk, or downloaded from a server and stored locally on the client. One or more of these, or a combination of one or more of these, may be provided, wired or preprogrammed to a chip such as an EEPROM semiconductor chip, an ASIC, or a DSP integrated circuit.

The information introduced here is known or described in detail in the relevant apparatus paragraph. In addition, much of the description provided herein is explicitly disclosed in the relevant apparatus paragraphs. Many or all of the additional aspects of the invention are to be understood by one of ordinary skill in the art as would be inherent in the detailed description provided in this relevant apparatus paragraph, or will be recognized as known in the art. Those skilled in the art can implement aspects of the invention based on the detailed description provided herein and the matter presented herein.

The teachings of the invention provided herein are not limited to the speech signal processing described above, but may be applied to signal processing systems. Moreover, the elements and steps of the various embodiments described above can be combined to provide further embodiments.

Claims (7)

A system for detecting voiced and unvoiced speech in an acoustic signal with varying levels of background noise. At least two microphones for receiving acoustic signals, and One or more processors coupled to the microphones Including, the one or more processors, Generate a difference parameter between the acoustic signals received at each of the two microphones, wherein the difference parameter represents the relative difference in signal gain for a portion of the received acoustic signal, -Classify the acoustic signal information into unvoiced speech when the difference parameter exceeds the first limit, and To classify the acoustic signal information into voiced speech when the difference parameter exceeds the second limit, System for detecting voiced and unvoiced speech, characterized in that. As a method for detecting voiced and unvoiced speech in an acoustic signal having various levels of background noise, Receiving acoustic signals at two receivers, Generating a difference parameter between acoustic signals received at each of the two receivers, where the difference parameter represents the relative difference in signal gain for a portion of the received acoustic signal, Classify the acoustic signal information into unvoiced speech when the difference parameter exceeds the first limit, and Classifying acoustic signal information into voiced speech when the difference parameter exceeds a second limit, Method for detecting voiced and unvoiced speech comprising the above steps. The method of claim 2, Generating the first and second limits using a standard deviation corresponding to the occurrence of the difference parameter. And further comprising a step. The method of claim 2, Classify the acoustic signal information as noise when the difference parameter is less than the first limit, and To remove noise for the classified noise, And further comprising a step. The method of claim 2, wherein the method is To receive physiological information related to human speech activity Wherein the physiological information comprises one or more detectors selected from RF devices, electroglottographs, ultrasound devices, acoustic throat micorphones, and airflow detectors. And receiving physiological data related to human speech. A system for detecting voiced and unvoiced speech in an acoustic signal having several levels of background noise, the system comprising: At least two microphones for receiving acoustic signals, One or more speech sensors that receive physiological information related to human speech activity, and At least one processor coupled to the microphone and the voice sensor Includes, wherein the one or more processors, Generate cross-correlation data between physiological information and acoustic signals received at one of the two microphones, -When the cross-correlation data corresponding to a part of the acoustic signal received from one receiver exceeds the correlation limit, the acoustic signal information is classified into voiced speech, Generate a difference parameter between the acoustic signals received at each of the two receivers, where the difference parameter represents the relative difference in signal gain for a portion of the received acoustic signal, -Classify the acoustic signal information into unvoiced speech when the difference parameter exceeds the gain limit, and Classifying acoustic signal information as noise when the difference parameter is less than the gain limit, System for detecting voiced and unvoiced speech, characterized in that. As a noise reduction method of an acoustic signal, this method Receiving acoustic signals at both receivers, receiving physiological information related to human speech activity at the speech sensor, Generate cross-correlation data between the acoustic signal received at one of the two receivers and the physiological information, -When the cross-correlation data corresponding to a part of the acoustic signal received from one receiver exceeds the correlation limit, the acoustic signal information is classified into voiced speech, Generate a difference parameter between the acoustic signals received at each of the two receivers, where the difference parameter represents the relative difference in signal gain for a portion of the received acoustic signal, -Classify the acoustic signal information into unvoiced speech when the difference parameter exceeds the gain limit, and Classifying acoustic signal information as noise when the difference parameter is less than the gain limit, Noise canceling method of an acoustic signal comprising the above steps.