KR100978015B1 - Stationary spectral power dependent audio enhancement system - Google Patents

Abstract

A spectral processor coupled to the signal input for carrying the distorted request signal z, the reference signal input, and both signal inputs and processing the distorted request signal by the reference signal x serving as an estimate for the distortion of the request signal. An audio reinforcement system (1) for speech recognition or voice control comprising (SP) is disclosed. The spectral processor SP determines the factor C 'so that the estimate is a function of the product of C' over the spectral power of the reference signal, and the factor C 'is a component of the signals z, x that are essentially fixed over time. For the processing to be determined according to the spectral ratio between them. The factor determined by the fixed portions of the signals obviates the application of the critical speech detector in the audio enhancement system.

Spectrum Processor, Distorted Request Signal, Spectrum Ratio

Description

Stationary spectral power dependent audio enhancement system

The present invention is coupled to the signal input for carrying a distorted desired signal, a reference signal input for a reference signal, and the two signal inputs to serve as an estimate for distortion of the required signal. The present invention relates to an audio enhancement system including a spectral processor for processing the distorted request signal by a reference signal, and to signals suitable for use therein.

The present invention relates to a system such as a hands-free communication device, in particular a communication system, such as a mobile phone, a voice recognition system, or a voice control system, which system is provided with an audio enhancement system. A method for reinforcing a distorted request signal, wherein the signal is spectrally processed by a reference signal that serves as an estimate for distortion of the request signal.

The audio enhancement system is implemented by a device for suppressing interference components such as distortion noise known from WO 97/45995. Known systems include a number of microphones coupled to audio signal inputs. The microphones include a main microphone for the distorted request signal and one or more reference microphones for receiving an interference signal. The system also includes a spectral processor implemented with a signal processing device coupled to the microphones via audio signal inputs. The interference signal is spectral subtracted from the distortion signal to represent an output signal that includes a reduced interference noise component at the output of the system.

A disadvantage of the known audio enhancement system is that its interfering signal cancellation capabilities depend on the application of the speech detector to be coupled to the speech processing device. The operation of a known audio enhancement system depends critically on proper speech detection by the speech detector.

It is therefore an object of the present invention to provide an improved audio enhancement system and an associated method that is not complicated by the presence and critical operation of the voice detector.

In addition, the audio enhancement system according to the present invention is characterized in that the spectral process of the signals z and x is such that the factor C 'is determined so that the estimate is a C' product of the spectral power of the reference signal, and the factor C 'is essentially fixed with respect to time. Characterized in that it is provided for said processing to be determined in accordance with the spectral ratio between the components.

Similarly, the method according to the invention is characterized in that the estimation is a function of the product of C 'over the spectral power of the reference signal, and C' is a function of It is determined according to.

The inventor has found that the factor C 'as defined is inherently insensitive to the required signal. Only factor C 'accounts for the ratio of fixed components in signals z and x. With the concept of factor C ', a reliable estimate can be provided for the distortion of the distorted request signal actually input to the audio enhancement system without the need for using a speech detector. This leads to improved and less important distortion cancellation characteristics of the simplified audio enhancement system according to the present invention. If one or more reference signals include distortions such as noise, echoes, contention speech, reverberation of the demanded speech signal, etc., distortion cancellation is particularly improved. In addition, the frequency dependent estimate of the distortion may be calculated in any scenario where some reference signal (s) are available.

Another advantage is that no explicit estimation of the individual distortion components, such as the noise floor or the echo tail, is required, and if required a combination technique with these components can be easily achieved. This is particularly desirable when there are no good estimation techniques, such as for microphone beamforming applications.

An embodiment of the audio enhancement system according to the invention has the unique features outlined in claim 2.

In general, spectral powers in the form of average spectral power covering a particular number of time frames are measured. For a period of time, the time period minimums for both spectral powers are determined, without the actual burden on the computational complexity of the audio enhancement system according to the invention.

In another embodiment of the audio enhancement system according to the invention, the time period comprises at least one pause in the distorted request signal. This results in fixed spectral component values of the well-determined minimum value and the distorted required input signal, which minimum accurately represents the fixed distortion in the input signal.

Preferably, the time period lasts from 4 seconds to 5 seconds so that speech pauses in the distorted request signal input to the audio enhancement system are normally included.

Another embodiment of the audio enhancement system has the unique features outlined in claim 5.

In general, the distortion estimate of the required signal can be preferably expressed as a positive function, for example by signal power or signal energy, and is defined by one of the spectral units.

The substantially preferred embodiment has the unique features of claim 6.

In that case, the audio enhancement system is included cost-effectively and easily to implement shift registers for storing values of spectral powers and / or smoothed spectral powers.

The audio enhancement system and method according to the invention will now be described in more detail with additional advantages, with reference to the accompanying drawings, wherein like components are referred to by the same reference numerals.

1 is a basic diagram of an audio enhancement system according to the present invention.

2 is a basic diagram implemented in another embodiment of an audio enhancement system according to the present invention having a filter and a sum beamformer.

3 shows a detailed embodiment of an audio enhancement system according to the invention.

1 shows a basic diagram of an audio enhancement system 1 implemented by a spectrum processor SP, in which the frequency domain input signals z, x and output signal q are shown. These frequency domain signals are spectrally calculated block-wise at the processor SP by a Discrete Fourier Transform, such as a short time DFT, simply referred to as STFT. This STFT is a function of both time and frequency, which can be represented only by the arguments kB, lw ₀ or sometimes the argument w _k . Where k denotes a discrete time frame index, B denotes a frame shift, l denotes a (discrete) frequency index, w ₀ denotes a fundamental frequency spacing, and w _k denotes the spectral component of the index k. The input signal z represents the distorted request signal. It is usually in the form of speech, which includes the sum of the distortion of the demand signal, such as noise, echo, contention speech or reverberation, and the demand signal. Signal x represents a reference signal from which distortion estimates within the distorted request signal z are derived. Signals z and x may originate from one or more microphones 2, as shown in FIGS. 1 and 2. In the multi-microphone audio enhancement system 1, there are two or more separate microphones 2 for deriving a reference signal from one or more microphones.

The audio enhancement system 1 may comprise adaptive filter means (not shown) to derive the reference signal x from it. In this case, the reference signal is generated at the shortest end of the communication system.

In the embodiment of Figure 1, only signal x includes a reference or noise signal, while signal z includes both a request signal and a noise signal. FIG. 2 shows an embodiment of an audio enhancement system 1 for the case where both of the microphones ₂ detect voice and noise via the microphone array signals u ₁ , u ₂ . The filter and sum beamformer 3 are now coupled between the microphones 2 and the spectral processor SP. In addition, the spectral processor receives the signals z and x described above, where only signal x includes a reference or noise, and signal z includes a request signal and a noise signal. Through the respective transfer functions f ₁ (w), f ₂ (w), the beam former ( ₁ ) is adapted such that the distorted request signal z is obtained by a linear combination of each of the microphone array signals u ₁ , u ₂ . 3) is designed. In order to project these signals into a subspace orthogonal to the request signal, the reference signal x is derived from the individual microphone array signals by a blocking matrix B (w). Ideally, the output signal x of the matrix B (w) does not contain the desired speech but contains distortions. The signals z and x are then supplied to the spectral processor SP for spectral processing the distorted request signal z by the reference signal x. The signal q from the processor SP is in fact an output signal without distortion. q = G × z, where G is the gain function to be described below.

The audio reinforcement system 1 may be included in a system, in particular a communication system such as a handsfree communication device such as a mobile phone, a voice recognition system, or a voice control system.

The spectral processor SP operates to serve as a controllable gain function for subsequent frequency bins generated by the Discrete Fourier Transform (DFT) described above. This gain function is applied to the distorted desired speech signal z, but the phase of the signal z is not changed. For good performance audio enhancement, gain function, i.e., estimation of distortion present in the input signal is important. However, various gain functions may be used depending on the optimization criteria being processed. Examples include spectral subtraction, Wiener filtering, or Minimum Mean-Square Error estimation or log-MMSE estimation based on spectral amplitude or magnitude, or squared spectrum. Squared spectral magnitude, power spectral density, or Mel-scale smoothed spectral density of the included signals. These techniques may be combined with the applications described above for audio enhancement systems 1 having one or more microphones and / or loudspeakers.

For example, for the Wiener filter type described below, the gain function implemented in the spectral processor SP has the following form:

는 왜곡에 적용된 억제량을 조정하게 하는 소위 초과 감산 인자를 나타낸다. 이처럼 트레이드-오프(trade-off)는 왜곡 억제량과 프로세서 출력 신호의 지각 품질 사이에 형성될 수 있다.Where P _{zz, n} (kB, lw ₀ ) and P _zz (kB, lw ₀ ) are estimates of the power distribution of the distortion in the input signal z and the power distribution of the input signal z itself.

Denotes a so-called excess subtraction factor that allows to adjust the amount of suppression applied to the distortion. As such, a trade-off may be formed between the amount of distortion suppression and the perceived quality of the processor output signal.

이 이하와 같이 제안된다:In equation (1), P _{zz, n} (kB, lw ₀ ) is not generally known and should therefore be estimated. calculation

This is suggested as follows:

Where the ratio term is:

은 음성과 같은 요구 신호의 부재 동안 측정된, 왜곡된 요구 신호(z)에 대한 왜곡의 시평균 스펙트럼 전력(time averaged spectral power)이고,

은 기준 신호(x)의 시평균 스펙트럼 전력이다. 예를 들면, 스펙트럼 진폭 또는 크기와 같이 스펙트럼 전력에 대한 양(陽, positive)의 측정값으로서, 포함된 신호들의 제곱 스펙트럼 크기, 전력 스펙트럼 밀도, 또는 멜-스케일 평활화된 스펙트럼 밀도가 취해질 수 있다. 프로세서(SP)에서 식(3)의 구현은 음성 검출기를 요구한다. 상기 음성 검출기가 정확히 수행하지 않는다면, 요구 음성이 영향받을 수 있고, 이것은 방지되어야 하는 가청 아티팩트들(audible artifacts)을 초래한다. 그러나, 차나 공장과 같은 잡음 조건들에서 신뢰성 있는 음성 검출이 태스크를 수행하는 것은 어렵다.here

Is the time averaged spectral power of the distortion for the distorted demand signal z, measured during the absence of a demand signal such as speech,

Is the time average spectral power of the reference signal x. For example, as a positive measure of spectral power, such as spectral amplitude or magnitude, the squared spectral magnitude, power spectral density, or mel-scale smoothed spectral density of the included signals can be taken. The implementation of equation (3) in the processor SP requires a speech detector. If the voice detector does not perform correctly, the required voice may be affected, resulting in audible artifacts that should be avoided. However, it is difficult for reliable voice detection to perform this task in noisy conditions such as cars or factories.

In general, a robust algorithm that does not require a speech detector by suggesting a new factor C ', which is virtually insensitive to the demanded speech, actually done by concentrating on the fixed proportions of the ratio of equation (3) as an estimate for factor C. Is generated. In a practical implementation of the idea, the factor C 'is defined by the ratio of the minimum spectral power of the distorted request signal z and the minimum spectral power of the reference signal x, wherein the two minimum values are determined for one time period. . It is expressed by the following formula:

The time period between l-L time frames and l time frames covers L time frames including at least one pause present in the distorted request signal. If the request signal is a voice signal, this is typically a voice pause. The minimum value thus determined concentrates the ratio of equation (4) on the fixed components of each of the signals z, x, where the minimum represents the fixed components of the distortion or noise. Generally, the time period lasts at least 4 to 5 seconds. The factor C 'given in equation (4) is determined based on the fixed components of the signals z, x. It is assumed that even if non-fixed components such as voice are present in these signals, the operation performed by the spectral processor SP is based on this assumption.

The spectrum of the numerator and denominator of factor C 'in equation (4) is obtained by smoothing the power spectrum in the first order recursions implemented in blocks LPF1, LPF2 having smoothing constants β, respectively. The recursive implementation in these blocks uses combined multipliers (X), adders (+), and delay lines (z ⁻¹ ) as shown to obtain smoothed power spectral density versions of the input x, z signals. Include. For example, the z signal then follows the smoothing rule:

Here, the smoothing constant β is assumed to be a value between 0 and 1. The same rule can apply for the x signal spectrum. The value of β can be controlled in any desired manner. Its value generally corresponds to a time constant of 50-200 ms. For each time frame index m, each of these smoothed amounts is stored in a buffer in the form of shift registers SR1, SR2, respectively. In addition to the L smoothed values stored in each of the register locations, individual minimum values are supplied to divisor D to represent the calculated value of C 'according to equation (4). Of course, appropriate measures are taken to prevent division by small values in the denominator.

If the average level of the demanded speech signal is too high relative to the average level of distortion, a problem may arise in that the average output by LPF1 and LPF2 is dominated by the demanded speech. This is because these averages take a long time to return to the low distortion level after the occurrence of the high speech level. When the estimate of C 'can be affected by the demanded speech, it results in unwanted suppression of the demanded speech signal. This effect is reduced, for example, by applying a multivariable compression function f _c in the recursions as follows:

The same rule can apply for the x signal. If the new input power value is relatively large compared to the values in the filters LPF1, LPF2, the compression function is selected to reduce the updater phase of recursion. Therefore, the compression function reduces the impact of high demanded speech levels on average signal power. Examples of suitable compression functions are given below:

Where δ is a positive constant. The smaller the value of δ, the slower rise in signal value follows the recursive filters LPF1, LPF2. An embodiment comprising a compression block f _c is shown in FIG. 3. If no compression is needed, it can simply be omitted.

Although the foregoing has been described with reference to the preferred embodiments and the best possible modes, the embodiments of the invention are related to the implementation of the system as the various changes, features, and combinations of the features within the scope of the appended claims are within the understanding of those skilled in the art. It will be understood that it is not to be construed as limiting the examples of examples and methods.

Claims (12)

Signal input for distorted desired signal z, Reference signal input for reference signal x, and A spectrum coupled to both of the signal inputs and processing the distorted request signal z by the reference signal x to yield a distortion estimate of the distorted request signal z, yielding a substantially distortion free output signal q An audio enhancement system comprising a processor, The spectral processor is provided for processing the distorted request signal z using factor C ', The distortion estimate of the distorted request signal z is a function of the product of the factor C 'and the spectral power of the reference signal x, Wherein the factor C 'is determined by the spectral ratio of the distorted demand signal z and the essentially fixed components of the reference signal x. The method of claim 1, The factor C 'is defined by the ratio of the minimum of the spectral power of the distorted request signal and the minimum of the spectral power of the reference signal, Wherein both minimum values are determined for a specific time period. The method of claim 2, Wherein said time period comprises at least one pause in said distorted request signal. The method of claim 3, wherein Wherein said time period lasts at least 4 to 5 seconds. The method according to any one of claims 1 to 4, The spectral power is defined by a function of any quantity related to the spectral power, such as spectral amplitude, squared spectral magnitude, power spectral density, or Mel-scale smoothed spectral density. An audio enhancement system, characterized in that. The method according to any one of claims 1 to 4, And the spectral processor comprises shift registers for storing values of spectral power. The method according to any one of claims 1 to 4, And the spectral power is smoothed spectral power. A communication system provided with an audio enhancement system according to claim 1. As a method of enhancing the distorted request signal z, Processing the distorted request signal z by a reference signal x that causes a distortion estimate of the distorted request signal z to be derived, thereby reinforcing the distorted request signal z comprising a substantially distorted output signal q. In the way, Said processing step comprises processing said distorted request signal z using factor C ', The distortion estimate of the distorted request signal z is a function of the product of the factor C 'and the spectral power of the reference signal x, Wherein said factor C 'is determined by the spectral ratio of the distorted demand signal z and the essentially fixed components of said reference signal x. delete A speech recognition system provided with an audio enhancement system according to claim 1. A voice control system provided with an audio enhancement system according to claim 1.

Images