CN1748250A - System and method for speech processing using independent component analysis under stability restraints - Google Patents

Abstract

A system and method for separating a mixture of audio signal into desired audio signals (430) (e.g., speech) and a noise signal (440) is disclosed. Microphones (310, 320) are positioned to receive the mixed audio signals, and an independent component analysis (ICA) processes (212) the sound mixture using stability constraints. The ICA process (508) uses predefined characteristics of the desired speech signal to identify and isolate a target sound signal (430). Filter coefficients are adapted with a learning rule and filter weight update dynamics are stabilized to assist convergence to a stable separated ICA signal result. The separated signals may be peripherally-processed to further reduce noise effects using post-processing (214) and pre-processing (220, 230) techniques and information. The proposed system is designed and easily adaptable for implementation on DSP units or CPUs in audio communication hardware environments.

Description

Under stable constraint, use the speech processing system and the method for independent component analysis

Invention field

The present invention relates to audio signal processing and method, relate in particular to the system and method that in acoustic environment, improves voice quality.

Technical background

In a lot of zones of periodic traffic, especially exist in the zone of a large amount of noises at those, it is very important that voice signal is handled.Noise in the real world comprises tangible single source noise from multiple source, and they are converted to the multiple sound with echo and reverberation (reverberation) in real world.Unless separate and isolate, otherwise be difficult to from background noise, extract the sound of wanting.Background noise comprises echo, reflection and the reverberation that produces in signal that multiple noise signal that conventional environment produces, other people background conversation produce and each signal from these signals.For the communication that the user talks in the noisy environment of being everlasting, desirable is that the user's voice signal is separated from background noise.Such as cell phone, hands-free phone, headphone, osophone, wireless phone, teleconference, CB radio station, intercom, computer telephony application, computing machine with automobile voice commander uses and voice communication means such as other hand-free application, intercom, microphone system can utilize voice signal to handle, thereby the voice signal that will want is separated from background noise.

Past has worked out the method that a variety of voice signals that will want are separated from background noise.The signal that the prior art noise filter will have predetermined characteristic is identified as white noise signal, deducts these signals then from input signal.Though these methods are enough simple with fast for real-time processing audio signal, are not easy to adapt to different acoustic environments, and can cause waiting to differentiate the obvious deterioration of voice signal.The predetermined hypothesis of feature of noise may contain wide, also may contain deficiency.The result, these methods may be considered as " noise " with some part of individual voice, and therefore it is removed from the output voice signal, simultaneously, these methods may be with some part of background noise, for example music or talk is considered as non-noise, and therefore it is included in the output voice signal.

The nearest additive method of exploitation, independent component analysis (" ICA ") for example provides relatively accurate and method flexibly, is used for background noise is separated from voice signal.For example, thus the open WO 00/41441 of PCT discloses and uses specific ICA technology to handle input audio signal to reduce noise in the output audio signal.ICA separates the technology of mixing source signal (component) at a kind of being used for, and wherein, supposes that these signals are separate.In its reduced form, independent component analysis for example, multiply by mixed signal with this matrix with " removing to mix (un-mixing) " matrix application of weight on mixed signal, thereby produces the signal that separates.Give these weight allocation initial values, adjust them then, make the joint entropy maximum of these signals, thereby information redundance is minimized.Repeating weight adjustment and entropy increases process, is reduced to minimum up to the information redundance with these signals.Because this technology does not need each signal source for information about, so be called as " separation of blind source " method (BSS).Blind separation problem refers to the thought of separating from the mixed signal of a plurality of independent sources.

TonyBell has discussed ICA the earliest in U.S. Patent No. 5706402, this has caused more deep research.A lot of different ICA technology or algorithms are arranged now.Can in relevant books of ICA and reference, find the summary that most popular algorithm and technology are carried out, for example, Te-Won Lee writes, " Independent Component Analysis:Theory and Applications " (in September, 1998) that Bostonian Kluwer Academic Publishers publishes, Hyvarinen etc. write, " IndependentComponentAnalysis " first published (May 18 calendar year 2001) that Wiley-Interscience publishes, " Self-Organizing NeuralNetworks:Independent Component Analysis and Blind SourceSeparation " (Springer Verlag, September 1999) and Mark Girolami (editor) " Advances inIndependent Component Analysis " (the Springer Verlag August 2000) in " Perspectives in Neural Computing " of Mark Girolami in " Perspectives in Neural Computing ".Svd (Singular valuedecomposition) algorithm is disclosed in " Adaptive FilterTheory " third edition (1996) that Simon Haykin Prentice-Hall that write, the New Jersey publishes.

Past has worked out a lot of ICA algorithms commonly used to optimize their performance, comprises the algorithm that the algorithm that only existed was carried out material alteration before 10 years.For example, do not use " Neural Computation " 7:1129-1159 (1995) and the Bell of A.J.Bell and TJ Sejnowski usually, the achievement of describing in the U.S. Patent No. 5706402 of A.J. with its patent form.On the contrary, in order to optimize performance, this algorithm has experienced the several recreation of a plurality of different entities.A kind of such change comprises uses Amari, Cichocki, " natural gradient " that Yang (1996) describes.Other ICA algorithms commonly used comprise method (Cardoso, 1992 of the high-order statistic of calculating such as semi-invariant; Comon, 1994; Hyvaerinenand Oja, 1997).

But, the signal of record during a lot of known ICA algorithms can not be separated in effectively, true environment just comprised echo, for example echo that causes owing to room reflections originally.What should emphasize is that method up to now is only limited to separation and mixes the signal that is caused by the linearity steady (stationary) of source signal.The phenomenon that direct-path signal and echo signal addition thereof are caused is called as reverberation, is the subject matter in enhancing of emulation voice and the recognition system.Now, the ICA algorithm need comprise the long wave filter that can separate those time-delays and echo signal, has got rid of so effectively and has used in real time.

Fig. 1 shows an embodiment of prior art ICA signal separation system 100.In this prior art systems, be used for telling individual signals from any amount of mixed signal of importing this FL-network as the FL-network of neural network.As shown in Figure 1, system 100 comprises two input channels 110 and 120, their receiving inputted signal X ₁And X ₂For signal X ₁, use the direct wave filter W of ICA ₁With ICA cross-filters C ₂For signal X ₂, use the direct wave filter W of ICA ₂With ICA cross-filters C ₁Direct wave filter W ₁And W ₂Communicate by letter for directly adjusting.Cross-filters is a feedback filter, and their filtering signals separately and signal by the filtering of direct wave filter institute are merged.After the convergence of ICA wave filter, the output signal U that is produced ₁And U ₂The signal that expression separates.

The U.S. Patent No. 5675659 of Torkkola etc. has proposed to be used for time-delay is carried out the blind method and apparatus that separates with the signal source of filtering.Torkkola has proposed a kind of ICA system of entropy maximization that will separate output, but what use is compound filter, rather than as the static coefficient in the Bell patent.Yet, when existence has the input signal (as voice signal) of time-varying input energy, being used to described in the Torkkola calculates joint entropy and adjusts the ICA calculating of cross-filters weight, is unsettled on mathematics, and reverberation effect is incorporated in the separate output signals.Therefore, the filters solutions that is proposed can't realize that the stable of real speech signal separates with acceptable blind source in the perception.

Typical ICA realizes also facing other obstacles, and the computing power very strong as needs is with the joint entropy of double counting signal and the weight of adjustment wave filter.A lot of ICA also needing to realize the direct correlation of many wheel feedback filters and wave filter.As a result, the ICA filtering that is difficult to finish in real time voice with use a large amount of microphones to separate a large amount of mixing source signals.Under the situation for the position of localization on comfortable space in a steady stream, rational filter tap of quantity available and recording microphone calculate the compound filter coefficient.Yet, if source signal is distributed in the space, as be derived from background noise, sound of the wind or the background conversation of vibration, launch from a plurality of different directions at the signal of microphone position place record so, need very long and the filter construction of complexity or a large amount of microphones.Because the processing power and the hardware complexity of any reality system all have limitation, thus additional disposal route must the ICA filter construction of being discussed be replenished, thereby strengthen a kind of sound method that provides for real-Time Speech Signals.The computation complexity of this system should with the processing power compatibility such as little consumer devices such as cell phone, PDA(Personal Digital Assistant), Voice Surveillance equipment, radio station.

Needed is a kind of simple method of speech processing, and this method is isolating speech signals from background noise in real time, and does not need very strong computing power, but still can produce relatively accurate result and can adapt to different environment neatly.

Summary of the invention

The present invention relates to speech processing system and method, be used in identification and the expectation sound signal of separating in the noisy environment such as at least one voice signal.This speech processes is carried out on such as mobile phone, headphone or cellular equipment with at least two microphones.At least two microphones are positioned on the shell of this equipment, are used to receive the useful signal from target, for example from speaker's voice.Microphone is set makes its receiving target user's voice, but also receive noise, voice, reverberation, echo and other undesirable voice signals from other sources.At least two microphone received audio signals, it comprises that expectation target voice and other do not wish the potpourri of acoustic information.Use improved ICA (independent component analysis) to handle, handle mixed signal from microphone.This speech processes is used predetermined phonetic feature, to assist recognition of speech signals.Like this, this speech processes produces expectation voice signal and the noise signal from the targeted customer.Can use the further filtering of noise signal and handle desired voice signal.

A scheme of the present invention relates to a kind of speech Separation system, and it comprises: at least two input signal channels respectively comprise the combination of a sound signal or sound signal; Two improved independent component analysis cross-filters.By cross-filters the input signal of these two channels is carried out filtering, described cross-filters is preferably the infinite impulse response filter with non-linear limited function.Non-line boundedness function is the nonlinear function with predetermined maximum and minimum value, and these values can be calculated fast, sign function for example, it is according to input value, return on the occasion of or negative value, as output.Repeating after the feedback of signal, produce the output signal of two channels, a channel mainly comprises the sound signal of expectation, and one other channel mainly comprises noise signal.

A scheme of the present invention relates to the system and method that sound signal is separated into expectation voice signal and noise signal.From at least two channel receiving inputted signals, described input signal is the combination of expectation voice signal and noise signal.Use the independent component analysis cross-filters of equal number., and merged by the first cross-filters filtering from the signal of first channel, thereby on second channel, form the signal that strengthens with signal from second channel.Enhancing signal on the second channel is by the second cross-filters filtering, and merges with signal from first channel, thereby forms enhancing signal on first channel.Enhancing signal on first channel can be by the further filtering of first cross-filters.Repeat filtering and merging process, to reduce the information redundancy between these two channel signals.The output signal of two channels that produced represents that one mainly is that the channel of voice signal and one mainly are the channels of non-speech audio.Can carry out other sound enhancement methods, for example the spectrum subtracts, Wiener filtering, denoising and phonetic feature extract, with further raising voice quality.

Another scheme of the present invention relates in the feedback filtering design for scheme and comprises stabilization element.In a stable example, the designing filter weight adapts to rule, makes weight adapt to and dynamically requires consistent with the overall stability of feedback arrangement.Different with previous method is that overall system performance not only points to the expectation entropy maximization that separates output, but also considers stable constraint, thereby reaches more actual purpose.With this purpose be described as under the stable constraint the most general like principle better.These stable constraint in the most general seemingly estimation are carried out modeling corresponding to the temporal characteristics of source signal.In the entropy maximization method, the putative signal source is i.i.d (independent same distribution) stochastic variable.But, not random signal such as the actual signal of sound and voice signal, but relevant in time, level and smooth on frequency.This causes corresponding original I CA filter coefficient learning rules.

In another stable example, because these learning rules directly depend on the input range that is write down, thus reduce input channel in proportion by the self-adaptation scale factor, with restriction filter weight speed-adaptive.Determine scale factor according to recursion equation, and scale factor is the function of channel intake.Therefore, it is irrelevant with the entropy maximization of subsequently ICA filter operations.In addition, the adaptive characteristic of ICA filter construction hint:, then separate output signal and comprise the reverberation interference if filter coefficient is adjusted too soon or shown oscillation behavior.Therefore, must in time domain and frequency domain, the filter weight of acquiring be carried out smoothing processing, to avoid reverberation effect.Because this smoothing processing operation has reduced the speed of wave filter learning process, so the speech intelligibility design aspect of this enhancing has extra stabilizing influence to overall system performance.

For the performance that separate in the blind source of improving the space distribution background noise, can carry out pre-service or aftertreatment respectively to the input and output that ICA calculates, described space distribution background noise may be owing to the restriction to computational resource and number of microphone causes.For example, because voice activity detection is only used institute's processing information relevant temporal information or spectrum information with adaptive wiener filter,, replenish the ICA filter unit thus so an alternative embodiment of the invention imagination comprises these methods.

Last scheme of the present invention relates to the computational accuracy and the computing power problem of this wave filter feedback arrangement.In spacing accuracy computation environment is arranged (common 16 or 32), the filter coefficient quantization error can appear in filtering operation.These cause the constringency performance and the overall system stability that worsen usually.Can be by restriction cross-filters length with by changing original feedback arrangement, the control quantization influence feeds back in the ICA filter construction so the ICA of aftertreatment output changes into.It is emphasized that in limited precision environment and reduce intake in proportion, not only the viewpoint from stability is necessary, and is because the limited range of evaluation.Although the performance in the limited precision environment is reliable and adjustable, the speech processes scheme that is proposed should preferably realize in the floating point precision environment.At last, by selective filter length suitably with adjust the filter coefficient update frequency, finish in the realization of calculating under the constraint.The computation complexity of this ICA filter construction really is the direct function of these variablees of back.

Other schemes shown in the drawings and embodiment are described in " detailed description " part, by the protection domain of claim it are defined.

The accompanying drawing summary

Fig. 1 illustrates the block diagram of prior art ICA signal separation system;

Fig. 2 is the block diagram according to an embodiment of speech Separation of the present invention system;

Fig. 3 is the block diagram according to an embodiment of improvement ICA processing sub of the present invention;

Fig. 4 is the block diagram according to an embodiment of improved ICA speech Separation process of the present invention;

Fig. 5 is the process flow diagram according to method of speech processing of the present invention;

Fig. 6 is the process flow diagram according to speech de-noising process of the present invention;

Fig. 7 is the process flow diagram according to phonetic feature leaching process of the present invention;

Fig. 8 is the table that illustrates according to the combination examples of speech processes process of the present invention;

Fig. 9 is the block diagram that has according to a cellular embodiment of speech Separation of the present invention system;

Figure 10 is the block diagram with cellular another embodiment of speech Separation system.

Detailed description of preferred embodiments

Below in conjunction with accompanying drawing, the preferred embodiment of speech Separation system is described.In order to realize handling in real time by limited computational power, this system uses the improvement ICA processing sub of the cross-filters with limited function simple and that easily calculate.Compare with classic method, the ICA method of this simplification has reduced the requirement of computing power and isolating speech signals from non-speech audio successfully.

The speech Separation systematic review

Fig. 2 illustrates an embodiment of speech Separation system 200.System 200 comprises: voice strengthen module 210; Optional speech de-noising module 220; Optional pronunciation extracting module 230.Voice strengthen module 210 and comprise: improve ICA processing sub 212 and optional aftertreatment submodule 214.Improve ICA processing sub 212 and use simplification and improved ICA to handle, so that realize real-time speech Separation with low relatively computing power.In the application that does not need real-time voice to separate, improved ICA handles the requirement that can further reduce computing power.As what here use, term " ICA " and " BSS " can exchange, and be meant and be used for directly or minimize or maximized method by approximate mathematical notation indirectly interactive information, comprise decorrelation method, for example delay time decorrelation or any other decorrelation method based on second order or higher order statistical based on time domain and frequency domain.

" module " used herein or " submodule " can refer to any method, device, equipment, unit or computer-readable data storage medium that comprises the computer instruction of software, hardware or form of firmware.Should be understood that, a plurality of modules or system can be merged into a module or system, also a module or system can be separated into a plurality of modules or system, to carry out identical functions.In preferred embodiment about cellular phone application, combine with improved ICA processing sub 212 self or with other modules, be embedded into the microprocessor chip that is arranged in cell phone.When realizing with software or the executable instruction of other computing machines, in fact unit of the present invention is exactly the code segment of carrying out necessary task, for example routine, program, object, parts, data structure etc.Program or code segment can be stored in the processor readable medium, or are transmitted by transmission medium or communication link by the computer data signal that is included in the carrier wave." processor readable medium " can comprise any medium that can store or transmit information, comprises volatility, non-volatile, removable and removable medium not.The example of processor readable medium comprises and can be used for storing expectation information and circuit, semiconductor storage unit, ROM, flash memory that can be accessed, can wipe ROM (EROM), floppy disk or other magnetic storage mediums, CD-ROM/DVD or other optical storage medias, hard disk, fiber medium, radio frequency (RF) link or any other medium.Computer data signal can comprise any signal that can pass through such as the propagation of transmission mediums such as electronic network channels, optical fiber, air, electromagnetic wave, RF link.Code segment can be by downloading such as computer networks such as internet, Intranets.Under any circumstance, the present invention should not be considered limited to these embodiment.

Speech Separation system 200 comprises that one or more voice strengthen the various combinations of module 210, speech de-noising module 220 and pronunciation extracting module 230.Speech Separation system 200 also comprises the one or more sound identification module (not shown)s that the following describes.Each module can self be used as an autonomous system, perhaps uses as the part of a bigger system.As described below, this speech Separation optimum system choosing is integrated in the electronic equipment, and the latter accepts phonetic entry, thereby controls some function, or the phase cross-demand separates the noise of expectation from background noise.A lot of application needs strengthen clean desired audio or clean desired audio are separated from the background sound from a plurality of directions.Such application comprises such as the man-machine interface in electronics or computing equipment, they are integrated such as speech recognition with detection, voice enhancing and separate, ability such as voice-activated control.Because the lower computing power of speech Separation system's needs of the present invention is so it is applicable to the equipment that limited processing power only is provided.

Improved ICA handles

Fig. 3 illustrates an embodiment 300 of improved ICA or BSS processing sub 212.Respectively from channel 310 and 320 receiving inputted signal X ₁And X ₂Generally, each signal is from least one microphone, but what will be appreciated that is also can use other sources.With cross-filters W ₁And W ₂Be applied on each input signal, thus the separation signal U of generation channel 330 ₁Separation signal U with channel 340 ₂Channel 330 (voice channel) mainly comprises wanted signal, and channel 340 (noise channel) mainly comprises noise signal.Should be understood that, although used term " voice channel " and " noise channel ", but term " voice " and " noise " can exchange according to expection, for example, a certain voice and/or noise compare other voice/or noise may be desirable.In addition, can use this method, separate mixing noise signal from two above sources.

The preferred infinite impulse response filter that uses in improved ICA processing procedure.The output signal of infinite impulse response filter as at least a portion of input signal, feeds back in this wave filter.The output signal of finite impulse response filter is not fed back as input.Cross-filters W ₂₁And W ₁₂Can have the coefficient of sparse distribution in time, to catch macrocyclic time delay.In the form of simplifying most, cross-filters W ₂₁And W ₁₂Be the gain coefficient that each wave filter has only a filter coefficient, for example, time delay gain coefficient between output signal and the feedback input signal and the amplitude gain coefficient that is used for amplification input signal.In other forms, cross-filters can respectively have dozens of, hundreds of or thousands of filter coefficients.As described below, aftertreatment submodule, denoising module or pronunciation extracting module can further be handled output signal U ₁And U ₂

Separate to realize blind source although can obtain the ICA learning rules clearly, the specific implementation of its speech processes in acoustic environment can cause the nonsteady behavior of wave filter scheme.In order to ensure the stability of this system, W ₁₂Adaptation W dynamically and similarly ₂₁At first must be stable.The gain margin of such system is very low usually, this means: the increase of input gain, for example run into the non-stationary voice signal, and can cause the unstable exponential increase that also therefore causes weight coefficient.Because voice signal is usually expressed as the sparse distribution with zero-mean, so the sign function can vibrate in time continually, and causes nonsteady behavior.At last, because convergence needs bigger learning parameter fast, thus the compromise of an inherence arranged between stability and the performance, because the input gain conference makes that this system is more unstable.Known learning rules not only cause instability, but also easily owing to nonlinear s ign function vibrates, and especially near limit of stability the time, thereby cause filtering output signal Y ₁[t] and Y ₂The reverberation of [t].In order to address these problems, need to stablize W ₂₁And W ₁₂The adaptation rule.If the learning rules of these filter coefficients are stable, analysis and empirical studies show widely, and system is stable in BIBO (output of bounded input bounded).Like this, the final respective objects of entire process scheme is under stable constraint the noise voice signal to be carried out blind source to separate.

Therefore, guarantee that the main way of stability is suitably to adjust input, as shown in Figure 3.In this framework, scale factor sc_fact adapts to according to the input signal feature that enters.For example, if input is too high, this will cause the increase of sc_fact, thereby reduce input range.Between performance and stability, there is one to trade off.To import scaled sc_fact, and reduce SNR, this causes the reduction of separating property.Like this, should only input be adjusted to and guarantee stable needed degree.Can realize the additional stability of cross-filters by filter construction of operation, this filter construction has been considered the short-term fluctuation in the weight coefficient when each is sampled, and avoids the reverberation of being correlated with thus.It is level and smooth the regular wave filter of this adaptation can be considered as time domain.It is level and smooth to carry out further filtering in frequency domain, thereby strengthens the coherence of convergent separation filter on the side frequency section.By length L is arrived in zero tap (zero tapping) of K tap filter, support this wave filter is carried out Fourier transform with the enhancing time then, carry out inverse fourier transform subsequently, can finish this point easily.Owing to the rectangle time-domain window this wave filter has been carried out window treatments effectively, it has been undertaken correspondingly level and smooth by the sinc function in the frequency domain.This frequency domain filtering smoothly can be finished with Fixed Time Interval, thereby periodically the filter coefficient that adapts to is reinitialized to relevant separating.

Below equation be to can be used for the example of non-linear limited function that each size is the time-sampling window of k for t, time variable.

U ₁(t)=X ₁(t)+W ₁₂(t)  X ₂(t) (equation 1)

U ₂(t)=X ₂(t)+W ₂₁(t)  X ₁(t) (equation 2)

Y1=sign (U1) (equation 3)

Y2=sign (U2) (equation 4)

Δ W _12k=-f (Y ₁) * U ₂[t-k] (equation 5)

Δ W _21k=-f (Y ₂) * U ₁[t-k] (equation 6)

Function f (x) is non-linear limited function, promptly has the nonlinear function of predetermined maximum and predetermined minimum value.F (x) preferably one according to the symbol of variable x and reach the non-linear limited function of maximal value or minimum value fast.For example, top equation 3 and equation 4 use the sign function as simple limited function.Sign function f (x) is a plus or minus according to x, and value is 1 or-1.Exemplary non-linear limited function including, but not limited to:

$f\left(x\right)=\mathrm{sign}\left(x\right)=\left\{\begin{array}{cc}1& x>0\\ -1& x\≤0\end{array}\right\}$ (equation 7)

$f\left(x\right)=\tanh \left(x\right)=\frac{e^x-e^{-x}}{e^x+e^{-x}}$ (equation 8)

$f\left(x\right)=\mathrm{simple}\left(x\right)=\left\{\begin{array}{cc}1& x\≥\mathrm{\ϵ}\\ x/\mathrm{\ϵ}& -\mathrm{\ϵ}>x>\mathrm{\ϵ}\\ -1& x\≤\mathrm{\ϵ}\end{array}\right\}$ (equation 9)

These regular hypothesis can obtain floating point precision and be used to carry out necessary calculating.Although preferably adopt floating point precision, also can adopt fixed-point arithmetic, especially when it is applied to have the equipment of minimum of computation processing power.Though adopt the ability of fixed-point arithmetic, converge to best ICA and separate more difficult.The ICA algorithm is really based on the principle that must cancel interference source.Because fixed-point arithmetic is in some cases some mistake, when deducting almost equal numeral (perhaps adding very different numerals), the ICA algorithm may show less than best convergence property.

Another factor that may influence separating property is a filter coefficient quantization error effect.Because limited filter coefficient resolution, the adaptation meeting of filter coefficient from improvement, produces the consideration of determining convergence attribute in specified point generation line bonus gradually thus.The quantization error effect depends on several factors, but mainly is the function of filter length and employed bit resolution (bit resolution).Aforesaid adjustment convergent-divergent problem also is necessary in preventing the limited accuracy computation that numerical value overflows.Because the convolution that relates in this filtering may be increased to potentially and be higher than the numerical value that can obtain resolving range, so scale factor must guarantee that the wave filter input is sufficiently little, to prevent the generation of this situation.

Multichannel improves ICA and handles

Improved ICA processing sub 212 receives from least two input signals such as the audio frequency input channel of microphone.Can increase the quantity of audio frequency input channel, surpass the minimum value of two channels.When the quantity of input channel increased, the speech Separation quality can improve, and reaches the point that input channel quantity equals audio signal source quantity usually.For example; if the input audio signal source comprises: the generic background noise that speaker, background speaker, background music source and distant place road noise and wind noise produce, four channel speech Separation systems usually can be better than the performance of two channel systems so.Certainly, the input channel of use is many more, just needs many more wave filter and strong more computing power.

Improved ICA processing sub can be used for separating the input signal that surpasses two channels with process.For example, in cellular phone application, a channel can mainly comprise the voice signal of expectation, and one other channel mainly comprises the noise signal from a noise source, and another channel mainly comprises the noise signal from another noise source.For example, in multi-user environment, channel can mainly comprise the voice from a targeted customer, and one other channel can mainly comprise from another different target user's voice, the 3rd channel can comprise noise, is used for further handling this two voice channels.Will be appreciated that, also can use additional voice or destination channel.

Only relate to an expectation source speech signal although some is used, in other are used, may have a plurality of expectation source speech signals.For example, conference call application or Voice Surveillance application requirements are isolated a plurality of speakers' voice signal and they are separated from each other and come from background noise.Improved ICA process not only can be used for from background noise a source speech signal being separated, but also a speaker's voice signal can be separated from another speaker's voice signal.

Auxiliary process

In order to improve the inventive method and system in the performance aspect effect and the robustness, various auxiliary process technology can be applied on the input and output signal, and with different degree.Preconditioning technique and post-processing technology that method and system described here is replenished obviously can improve the performance that is applied to the blind source separate technology on the audio mix.For example, utilize undesirable output or unsegregated input, post-processing technology can be used to improve the quality of wanted signal.Equally, the finishing by improving mixed scenario is to replenish method and system described here, and preconditioning technique or information can strengthen the performance that is applied in the blind source separate technology on the audio mix.

Improved ICA handles voice signal is separated at least two channels, for example, and a channel (noise channel) and a channel (voice channel) that is used to expect voice signal that is used for noise signal.As shown in Figure 4, channel 430 is voice channels, and channel 440 is noise channels.Situation likely is, voice channel comprises the not noise signal of level of hope, and the noise channel still comprises some voice signals.For example, if plural important sound source is arranged and have only two microphones, perhaps, far away if two microphone close proximity and sound source are isolated, so, improved ICA handles the voice that may always not be enough to separate individually expectation from noise.Therefore, the signal demand after the processing carries out aftertreatment, with the background noise of removing the residue level and/or the quality that further improves voice signal.For example, by through list or multichannel voice enhancement algorithm, feed back the ICA output that separates, thereby realize this point.Can use S filter, realize SNR preferably for those background noise of being supported for a long time institute degraded signals, wherein said S filter has the noise spectrum that estimates by with the detected non-voice of the voice activity detector time interval.In addition, limited function is simple being similar to that joint entropy is calculated, and always can not reduce the information redundance of signal fully.Therefore, after using improved ICA processing separation signal, can carry out aftertreatment, with the quality of further raising voice signal.

Can abandon the noise signal channel of separation, but also can be with it as other purposes.If reasonably suppose remanent noise signal in the voice channel have to the noise channel in the similar signal characteristic of noise signal, so in post-processing unit, should filter the signal of its feature and noise channel signal feature similarity in the expectation voice channel.For example, can use spectrum to subtract technology and carry out aftertreatment.Signal characteristic in the identification noise channel.Compare with the prior art noise filter of being scheduled to hypothesis based on feature of noise, aftertreatment is more flexible, because its analyzes the feature of noise of specific environment, and removes the noise signal of expression specific environment.Therefore unlikelyly when noise eliminating contained wide or contained deficiency.Also can use such as other filtering techniques such as Wiener filtering and Kalman filtering and carry out aftertreatment.Because this ICA wave filter is separated the limit cycle of truly separating only converging on,, and can not cause better separating property so filter coefficient can continue adaptation.Observed the resolution limit that some coefficients float to them.Therefore, by IIR feedback arrangement as shown in Figure 4, feedback comprises the aftertreatment version of the ICA output of expectation speaker signal, thus overcome limit of convergence ring, and stablized the ICA algorithm.The useful spinoff of this process is that convergence is obviously quickened.

Can strengthen with voice and use such as other processing such as denoising, phonetic feature extraction, with the quality of further raising voice signal.Speech recognition application can be utilized the voice signal that separates by voice enhancing process.When voice signal separates substantially with noise, based on can be with higher accuracy work such as the speech recognition engine of methods such as hidden markov models chain, neural network learning and support vector machine.

With reference now to Fig. 5,, it illustrates the process flow diagram of speech processes.For example, method 500 can be used for speech ciphering equipment, for example portable radiotelephone mobile phone, head telephone or hand-free automobile external member.Will be appreciated that method 500 can be used for other speech ciphering equipments, and available dsp processor, general-purpose computations processor, microprocessor, gate array or other computing equipments are realized.In use, method 500 reception forms are the audible signal of voice signal 502.Voice signal 502 can be from a lot of sources, and can comprise voice from the targeted customer, near other people voice, noise, reverberation, echo, reflected sound and other undesirable sound.Although shown method 500 is used for identification and separates the simple target voice signal, should be understood that, can amending method 500, discern and separate other target sound signal.

In addition, the preconditioning technique of change or information can be used for raising or promotion morbid sound Signal Processing and separate, and for example utilize priori, feature in runout information or the input signal and condition are maximized, improve the adjustment of mixing situation etc.For example, because the output order of the ICA acoustic channel that separates normally can't learn in advance, so the mode of additional channel choice phase 510 to repeat based on the priori 501 about the expectation speaker, handled the content of segregated channel.Be used to discern expectation speaker phonetic feature standard 504 can based on but be not limited to room and time feature, energy, volume, frequency content, zero-crossing rate or with the relevant and incoherent speech recognition value with loudspeaker of detachment process parallel computation.For example, standard 504 can be configured to: response for example " wakes up " such as vocabulary that is tied such as particular commands.In another example, speech ciphering equipment can be to making response from the voice signal of ad-hoc location or direction emission, for example the position of front driver in the automobile.Like this, hand-free automobile external member can be configured to: only response made in the voice from this driver, and ignore the sound from passenger or wireless device.Perhaps, by modulation or handle the feature of input signal, for example, wait by space, time, energy, frequency spectrum and to modulate and to handle, can improve the condition of mixing sight.

On some speech ciphering equipments, always according to placing microphone, or make it self to have some feature to speech source, background noise or with respect to the preset distance of other microphones, adjust input signal, for example, the direction microphone.Shown in module 506, two microphones can be isolated separately, and be placed on the shell of speech ciphering equipment.For example, adjust head telephone usually, make microphone in about 1 inch of speaker's mouth, speaker's sound is normally from the nearest sound source of microphone.Equally, the microphone of handheld wireless phone, headphone or lapel microphone are rational known distance from target speaker's mouth usually.Because the distance from the microphone to the target source is known, so can be with the feature of this distance as the recognition objective voice signal.In addition, will be appreciated that, can use a plurality of features.For example, process 510 can only select within two inches, frequency component is shown as the voice signal of male sex's sound.Under the situation of using two microphones to be provided with, with the mouth of microphone near the expection speaker.The voice signal that this set allow to be isolated the expectation speaker enters in the ICA channel of separation, thus the phase residue that only comprises noise can be separated delivery channel as the noise reference, being used for subsequently, isologue speaker channel carries out aftertreatment.

In using the recording occasion of two above microphones, with to the similar mode of two channel sights explained earlier, with the individual ICA cross-filters of N* (N-1), two channel ICA algorithms are expanded to N channel (microphone) algorithm.The latter is used from the source location purpose with the channel option program one shown in [ad2], to select two best channel combinations from N recording channel, with two channel ICA algorithms it is handled then, to isolate the speaker of expectation., use from separating obtained various information of N channel ICA and the ICA cross-filters coefficient of learning, described information changes as be input to the relative energy that separates output source from record, but is not limited thereto for this reason.

Microphone is separately respectively received a signal, and it is an expectation target sound and the mixing of multiple noise and reverberation source.Handle in 508 at ICA, receive the voice signal 507 and 509 that mixes, separate.After identifying the target voice signal with identification processing 510, ICA processing 508 is separated into expectation voice signal and noise signal with morbid sound.This ICA handles can use noise signal, for example, further improves and be provided with weighting factor by using noise signal, with further processing 512 voice signals.In addition, additional filtering or handle 514 and also can use noise signal is further to remove the noise content, as will be described as further below from voice signal.

Denoising

Fig. 6 is the process flow diagram of an embodiment of denoising process.In cellular phone application, preferably use denoising to separate spatially unfixed noise source, for example from the wind noise of all directions.Noise-removed technology also can be used for removing the noise signal with fixed frequency.This process proceeds to module 610 from beginning module 600.In module 610, this process receives comments tone signal x.This process proceeds to module 620, wherein system-computed source coefficient s, the preferably formula below the employing:

$s_i=\underset{j}{\mathrm{\Σ}}{w}_{\mathrm{ij}}*x_j$ (formula 10)

In the equation above, w _IjExpression ICA weight matrix.Can in the denoising process, use in the United States Patent (USP) 5706402 the ICA method of explanation in the ICA method of explanation or the United States Patent (USP) 6424960.Then, this process proceeds to module 630, module 640 or module 650.Module 630,640 and 650 other embodiment of expression.In module 630, this process is according to signal s _iPower, select a plurality of important source coefficients.In module 640, this process will the most probably be applied on the source coefficient that calculates like contracting function, thereby eliminates unessential coefficient.In module 650, this process is with the basis function of each time-sampling t, filtering voice signal x.

This process proceeds to module 660 from module 630,640 or 650, wherein this process reconstructed speech signal, the preferably formula below the use:

$x_{\mathrm{new}}=\underset{j}{\mathrm{\Σ}}{a}_{\mathrm{ij}}*s_{j,\mathrm{shrinked}}$ (formula 11)

In the equation above, a _IjThe training signal of expression by producing with weighting factor filtering input signal.Like this, the denoising process abates the noise, and produces the voice signal x that rebuilds _NewWhen the information that can obtain about noise source, can obtain the good denoising result.As top illustrated in conjunction with improved ICA processing, the denoising process can be utilized the signal characteristic in the noise channel, removing in the signal of noise from voice channel.This processing proceeds to from module 660 and finishes module 670.

Phonetic feature extracts

Fig. 7 illustrates an embodiment of the phonetic feature leaching process that uses ICA.This process proceeds to module 710 from beginning module 700, there, and this process received speech signal x.As figure below in conjunction with Fig. 9 illustrated, voice signal x can be an input speech signal, the signal by the voice enhancement process, the signal by denoising or the signal by voice enhancing and denoising.

With reference to figure 7, this process proceeds to module 720 from module 710, and wherein this process is used the formula s shown in the formula 10 _{Ij, new}=W*x _Ij, calculate the source coefficient.Then, this process proceeds to module 730, wherein the voice signal of receiving is decomposed into basis function.This process proceeds to module 740 from module 730, wherein the source coefficient that calculates is used as proper vector.For example, the coefficient S that is calculated _{Ij, new}Or 2logs _{Ij, new}Be used for the calculated characteristics vector.Then, this process proceeds to and finishes module 750.

The phonetic feature that extracts can be used for recognizing voice or distinguish discernible voice from other sound signal.Can use the phonetic feature that extracts separately or it is used with cepstrum feature (MFCC).The phonetic feature that extracts can be used for discerning the speaker, for example, identifies individual speaker person from a plurality of speakers' voice signal, or voice signal is identified as the class that belongs to specific, for example from sex speaker's voice.Sorting algorithm also can be used the feature that is extracted, and detects voice signal.For example, can use the most general seemingly calculating, determine the possibility of described signal for true man's voice signal.

The phonetic feature that extracts also can be used in the application of text-to-speech, and this uses the computing machine reading matter that produces literal.The text-to-speech system uses a huge voice signal data storehouse.A challenge is to obtain good phonemic representation database.Prior art systems is used cepstrum feature, so that speech data is divided into phoneme database.By voice signal is decomposed into basis function, improved phonetic feature extracting method can be divided into voice the phoneme section better and therefore produce better database, thereby realizes the better voice quality of text-to-speech system.

In an embodiment of phonetic feature leaching process, use a basis function set for all voice signals, thus recognizing voice.In another embodiment, use a basis function set for each speaker, thereby discern each speaker.For the application such as a plurality of speakers of teleconference, this may be especially favourable.In another embodiment, use a basis function set for a class speaker, to discern each class.For example, use a basis function set, use another basis function set for women speaker for male speaker.United States Patent (USP) 6424960 has illustrated with the ICA mixture model discerns inhomogeneous sound.Such model can be used for discerning the speaker's of different speakers or different sexes voice signal.

Speech recognition

Speech recognition application can be utilized the voice signal that is separated by improved ICA process.By voice signal is separated substantially with noise, speech recognition application can be with higher accuracy work.Can be used for speech recognition application such as methods such as hidden markov models, neural network learning and support vector machine.As mentioned above, in being provided with of two microphones, improved ICA handles the voice channel that input signal is separated into expectation voice signal and some noise signals, and the noise channel of noise signal and some voice signals.

In order in noisy environment, to improve the accuracy of speech recognition, preferably have noise reference signal accurately, with according to this noise reference signal, noise is separated with voice signal.For example, use speech manual to subtract, from mainly being to remove signal the channel of voice signal with noise reference signal characteristics.Therefore, in preferred speech recognition system at unusual noisy environment, the voice channel of this system's received signal and noise channel, and identification noise reference signal.

Treatment combination

, some embodiment of phonetic feature extraction, denoising and speech recognition process have been described with the voice enhancement process.It should be noted that not all process all needs to use together.Fig. 8 is a table 800, has listed some typical combination of voice enhancings, denoising and phonetic feature leaching process.The left column of table 800 has been listed the type of signal, the right preferred process of handling the respective type signal of having listed.

Be expert in the configuration shown in 810, at first use voice to strengthen input signal is handled, use speech de-noising to handle then, and then use phonetic feature to extract and handle.When input signal comprised heavy noise and competition source, the combination of these three processes ran well.Heavy noise is meant the relative low amplitude noise signal from a plurality of information sources, and for example in the street, various types of noises are loud especially from different directions but there is not the noise of a certain type.The competition source is meant that for example, when the driver spoke by automobile telephone, auto radio was transferred to louder volume from one or several high amplitude signals with the source of expectation voice signal competition.Be expert in the another kind configuration shown in 820, at first use voice enhancement process input signal, use phonetic feature to extract then and handle.Omitted the speech de-noising process.When original signal comprises the competition source and when not comprising heavy noise, voice strengthen and the work in combination of phonetic feature leaching process good.

Be expert in another configuration shown in 830, at first use voice denoising input signal, handle with the voice feature extraction then.Here omitted the voice enhancing.When input signal comprises heavy noise and when not comprising the competition source, the work in combination of speech de-noising and phonetic feature leaching process is good.Be expert in another configuration shown in 840,, only carry out phonetic feature and extract for input signal.For the voice of the relative clean that does not comprise heavy noise or competition source, this process just is enough to reach good result.Certainly, table 800 has just been listed some examples, also can use other embodiment.For example, can voice enhancing, speech de-noising and phonetic feature extraction all be used for processing signals regardless of signal type.

Cellular phone application

Fig. 9 illustrates an embodiment of cellular telephone apparatus.Cellular telephone apparatus 900 comprises: two microphones 910 and 920 are used for the recording voice signal; And a speech Separation system 200, be used to handle the signal that is write down, from background noise, to separate the voice signal of expectation.Speech Separation system 200 comprises an improved ICA processing sub at least, and it is applied in cross-filters on the signal that is write down, to produce the separation signal on the channel 930 and 940.Then, the transmitter 950 expectation voice signal that will separate is transmitted to such as sound signal receiving equipments such as wire telephony or other cell phones.

Can abandon isolated noise signal, but also can be used as other purposes.Isolated noise signal can be used for correspondingly determining environmental characteristic and adjusting the cell phone parameter.For example, noise signal can be used for determining the noise level of speaker's environment.Then, if the speaker is in the high environment of noise level, then cell phone increases the volume of microphone.As mentioned above, also can be with noise signal as reference signal, further from the voice signal that separates, to remove remaining noise.

For ease of explanation, other cell phone part, for example battery, display panels etc. in Fig. 9, have been omitted.For ease of explanation, also omitted the cellular phone signal treatment step that relates to analog to digital conversion, modulation or realize FDMA (frequency division multiplexing), TDMA (time division multiplex) or CDMA (CDMA) etc.

Although Fig. 9 shows two microphones, also can use plural microphone.Existing manufacturing technology can be produced coin, syringe needle size or littler microphone, a plurality of microphones can be arranged on the equipment 900.

In one embodiment, traditional echo cancellation process of carrying out in the cell phone is replaced by the ICA process, the process that for example improved ICA submodule is carried out.

Because audio signal source is separated from each other usually, thus preferably microphone is arranged on the cell phone, and on sound, separate.For example, a microphone can be arranged on cellular front, and another microphone is arranged on the cellular back side.A microphone can be arranged near cellular top or the left side, and another microphone is arranged near cellular bottom or the right side.Two microphones can be arranged on the diverse location of cell phone headphone.In one embodiment, two microphones are arranged on the headphone, and two microphones are arranged on the cell phone handheld unit in addition.Therefore, the speech that two microphones can recording user, and no matter the user is to use handheld unit also to be to use headphone.

Although illustrated to have the cell phone that improved ICA handles as example, but other voice communication medium, for example the voice control of household electrical appliances, wire telephony, hand-free phone, wireless phone, teleconference, CB radio station, intercom, computer telephony application, computing machine and automobile speech recognition application, surveillance equipment, intercom etc., also can utilize improved ICA to handle, so that the voice signal of expectation is separated from other signals.

Figure 10 illustrates another embodiment of cellular telephone apparatus.Cellular telephone apparatus 1000 comprises two channels 1010 and 1020, is used to receive the voice signal from such as other cellular other communication facilitiess.Channel 1010 and 1020 receives the voice signal with the same dialog of two microphone records.Can use plural receiving element to receive two input signals with upper signal channel.Equipment 1000 also comprises speech Separation system 200, is used to handle received signal, thereby will expect that voice signal separates from background noise.Then, amplifier 1030 amplifies isolated expectation voice signal, thereby arrives the ear of cellular telephone subscribers.Receive on the cell phone by speech Separation system 200 is arranged on,, receive cellular user and also can hear high-quality voice even the emission cell phone does not have speech Separation system 200.Yet this need receive the session signal by two channels of two microphone records on the emission cell phone.

For convenience of explanation, Figure 10 has omitted other cell phone component, for example battery, display panel etc.For convenience of explanation, also omitted the cellular phone signal treatment step that relates to digital-to-analog conversion, demodulation or realize FDMA (frequency division multiplexing), TDMA (time division multiplex), CDMA (CDMA) etc.

Some aspect of the present invention, advantage and novel feature have been described here.Certainly, should be understood that all these aspects, advantage or feature needn't be included in any one specific embodiment of the present invention.Embodiment discussed here is an example of the present invention, can also increase, revises and adjust.For example, although formula 7,8 and 9 has provided the example of non-linear limited function, non-linear limited function is not limited to these examples, but can comprise any nonlinear function with predetermined maximum and minimum value.Therefore, protection scope of the present invention should be defined by the following claim book.

List of references

Hyvaerinen，A.，Karhunen，J，Oja，E.Independent?componentanalysis.John?Wiley?&?Sons，Inc.2001

Te-Won?Lee，Independent?Component?Analysis：Theory?andApplications，Kluwer?Academic?Publishers，Boston，September?1998

Mark?Girolami，Self-Organizing?Neural?Networks：IndependentComponent?Analysis?and?Blind?Source?Separation.In?Perspectives?inNeural?Computing，Springer?Verlag，September?1999

Mark?Girolami(Editor)，Advances?in?Independent?ComponentAnalysis.In?Perspectives?in?Neural?Computing，，Springer?Verlag，August2000

Simon?Haykin，Adaptive?Filter?Theory，Third?Edition，Prentice-Hall(NJ)，1996.

Bell，A.，Sejnowski，T.，Neural?Computation?7：1129-1159，1995

Amari，S.，Cichocki，A.，Yang，H.，A?New?Learning?Algorithm?forBlind?Signal?Separation，In：Advances?in?Neural?Information?ProcessingSystems?8，Editors?D.

Touretzky，M.Mozer，and?M.Hasselmo，pp.757-763，MIT?Press，Cambridge?MA，1996.

Cardoso，J.-F.，Iterative?techniques?for?blind?source?separation?usingonly?fourth?order?cumulants?In?Proc.EUSIPCO，pages739-742，1992.

Comon，P.，Independent?component?analysis，a?new?concept？SignalProcessing，36(3)：287-314，April?1994.

Hyvaerinen，A.and?Oja，E，A?fast?fixed-point?algorithm?forindependent?component?analysis.Neural?Computation，9，pp.1483-1492，1997

Claims (54)

1, a kind of method of separating the expectation voice signal in acoustic environment comprises:

Receive a plurality of input signals, described input signal is in response to described expectation voice signal and other voice signals produce;

Under stable constraint, use independent component analysis (ICA) or blind source to separate (BSS), handle received input signal; And

Received input signal is separated into one or more expectation sound signals and one or more noise signal.

2, method according to claim 1, wherein, one in the described expectation sound signal is described expectation voice signal.

3, method according to claim 1, wherein, described ICA or BSS handle and comprise by approximate and directly or indirectly the mathematical notation of interactive information is minimized or maximizes.

4, method according to claim 1 also comprises by coordinating the ICA weight adapting to the step of dynamically stablizing described ICA processing.

5, method according to claim 1 also comprises by using the self-adaptation scale factor to adjust the ICA input and stablizes the step that described ICA handles with constraint weight speed-adaptive.

6, method according to claim 1 also comprises by filter the filter weight of acquiring in time domain and frequency domain and stablizes the step that described ICA handles to avoid reverberation effect.

7, method according to claim 1 wherein, with different degree, is applied in the auxiliary process technology on described input and the separation signal.

8, method according to claim 1 also comprises and utilizes preconditioning technique or information to improve the performance of separation.

9, method according to claim 8 also comprises and improves the adjustment that is applied in the mixing sight on the described input signal.

10, method according to claim 2 comprises that also the characteristic information that utilizes described expectation voice signal discerns the channel that comprises described separation expectation voice signal.

11, method according to claim 10, wherein, described characteristic information is space, frequency spectrum or temporal information.

12, method according to claim 1, wherein, post-processing technology is used for utilizing at least one or the described input signal in the described noise signal at least one to improve the quality of described wanted signal.

13, method according to claim 12 comprises that also the noise signal that uses described separation is with further separation with strengthen the step of described expectation voice signal.

14, method according to claim 13, wherein, described use step comprises the noise spectrum that uses described noise signal to come the estimated noise wave filter.

15, method according to claim 1 also comprises:

At least two microphones are separated; And

Produce one of described input signal at each corresponding microphone place.

16, method according to claim 15, wherein, described division step comprises separates distance between about 1 millimeter and about 1 meter with described microphone.

17, method according to claim 15, wherein, described division step comprises separates described microphone on telephone receiver, headphone or hands free kits.

18, method according to claim 15, wherein, described ICA handles and comprises:

Be connected to first self-adaptation independent component analysis (ICA) wave filter of first delivery channel and second input channel,, make described first filter adaptation by relating to the recursive learning rule that non-linear bounded sign function is applied to described noise signal channel;

Be connected to the second self-adaptation independent component analysis wave filter of first input channel and second delivery channel,, make described second filter adaptation by relating to the recursive learning rule that non-linear bounded sign function is applied to described expectation voice signal channel;

Wherein, described first wave filter of repeated application and described second wave filter are to produce the voice signal of described expectation.

19, method according to claim 18, wherein, (a) feedback is by the described expectation voice channel of the described first self-adaptation independent component analysis wave filter recursive filtering, and adds it to from described second microphone described input channel, thereby produces described noise signal channel; And (b) feedback and is added it to from described first microphone described input channel by the described noise signal channel of the described second self-adaptation independent component analysis wave filter recursive filtering, thereby produces described expectation voice signal channel.

20, method according to claim 19, wherein, by the self-adaptation scale factor of the described input signal energy function of the conduct that calculates according to recursion equation, scaled described input channel signal.

21, method according to claim 18, wherein, by level and smooth described filter coefficient in time, stablize the described filter weight learning rules of the described first self-adaptation ICA cross-filters, and wherein, by level and smooth described filter coefficient in time, stablize the described filter weight learning rules of the described second self-adaptation ICA cross-filters.

22, method according to claim 18 wherein, is filtered the described first self-adaptation ICA cross-filters weight in frequency domain, and wherein, filters the described second self-adaptation ICA cross-filters weight in frequency domain.

23, method according to claim 18, also comprise the post-processing module that is connected to described expectation voice signal, its application comprises that the list of voice activity detection or multichannel voice strengthen module, and wherein, the output with described aftertreatment does not feed back to described input channel.

24, method according to claim 18, wherein, realize that in a station accuracy environment described ICA handles, wherein, use described self-adaptation ICA cross-filters at each sampled point, but, upgrade filter coefficient at the multiple place of described sampled point, and variable-sized filter length is used for being fit to available computing power.

25, method according to claim 18 comprises also and uses described noise signal that described expectation voice signal is carried out aftertreatment, and described post-processing module subtracts spectrum the voice signal that is applied to described expectation according to described noise signal.

26, method according to claim 18 comprises also and uses described noise signal that described expectation voice signal is carried out aftertreatment, and described post-processing module is applied to described expectation voice signal according to described noise signal with Wiener filtering.

27, method according to claim 18 also comprises from the 3rd channel receiving the 3rd group of audio input signal, and uses the 3rd wave filter, and non-linear limited function is applied to input signal.

28, a kind of speech ciphering equipment comprises:

At least two microphones that separate are configured to receive voice signal, and described microphone and speech source are at a distance of the distance of expectation; And

Be connected to the ICA or the BSS processor of described microphone,

Described processor is carried out following steps, comprising:

Reception is from the voice signal of described two microphones,

Under stable constraint, described voice signal is separated at least one expectation voice signal line and at least one noise signal line.

29, speech ciphering equipment according to claim 28 also comprises the post-processing filter that is connected to described noise line and described expectation voice signal line.

30, speech ciphering equipment according to claim 28, wherein, described microphone separates about 1 millimeter to about 1 meter.

31, method according to claim 30 comprises that also the voice signal to receiving at each microphone place carries out pre-service.

32, speech ciphering equipment according to claim 28, wherein, one of described microphone is positioned at the one side of described device housings, and another microphone is positioned at the another side of described device housings.

33, speech ciphering equipment according to claim 28 wherein, constitutes wireless telephone with described speech ciphering equipment.

34, speech ciphering equipment according to claim 28 wherein, constitutes wireless telephone with described speech ciphering equipment.

35, speech ciphering equipment according to claim 28 wherein, constitutes hand-free automobile external member with described speech ciphering equipment.

36, speech ciphering equipment according to claim 28 wherein, constitutes headphone with described speech ciphering equipment.

37, speech ciphering equipment according to claim 28 wherein, constitutes personal digital assistant with described speech ciphering equipment.

38, speech ciphering equipment according to claim 28 wherein, constitutes hand-held bar code scanning equipment with described speech ciphering equipment.

39, a kind of system that is used for separating at acoustic environment the expectation voice signal comprises:

A plurality of input channels, each receives one or more voice signals;

At least one ICA or BSS wave filter, wherein, described wave filter is one or more expectation sound signals and one or more noise signal with the Signal Separation of receiving under stable constraint; And

A plurality of delivery channels, the signal that transmission is separated.

40, according to the described system of claim 39, wherein, described expectation sound signal is the voice signal that receives in described a plurality of voice signal.

41, according to the described system of claim 39, wherein, described wave filter is by the approximate mathematical notation of directly or indirectly adjusting interactive information.

42, according to the described system of claim 39, wherein, described wave filter adapts to dynamically by coordinating the ICA weight, stablizes described ICA and handles.

43, according to the described system of claim 39, wherein, described wave filter is adjusted the ICA input with constraint weight speed-adaptive by using the self-adaptation scale factor, stablizes described ICA and handles.

44, according to the described system of claim 39, wherein, described wave filter, is stablized described ICA and is handled to avoid reverberation effect by filter the filter weight of acquiring in time domain and frequency domain.

45,, also comprise one or more auxiliary process wave filters that are applied to described input and/or output signal according to the described system of claim 39.

46, according to the described system of claim 45, also comprise one or more pre-processing filters.

47, according to the described system of claim 45, also comprise one or more post-processing filters.

48,, also comprise one or more microphones that are connected to described input channel according to the described system of claim 39.

49, according to the described system of claim 48, comprise two or more microphones, respectively separate the distance between about 1 millimeter and about 1 meter.

50, according to the described system of claim 39, wherein, the described system of structure on handheld device.

51, according to the described system of claim 39, wherein, described wave filter comprises:

Be connected to first self-adaptation independent component analysis (ICA) wave filter of first delivery channel and second input channel,, make described first filter adaptation by relating to the recursive learning rule that non-linear bounded sign function is applied to described noise signal channel;

Be connected to the second self-adaptation independent component analysis wave filter of first delivery channel and second input channel,, make described second filter adaptation by relating to the recursive learning rule that non-linear bounded sign function is applied to described expectation voice signal channel;

Wherein, described first wave filter of repeated application and described second wave filter are to produce the voice signal of described expectation.

52, be used for the system of isolating speech signals, comprise:

One group of signal generator is provided with each signal generator and makes it produce the mixed signal of mixing of the described voice signal of expression and other voice signals;

Processor is configured to receive each described mixed signal;

Described processor is carried out a kind of method, also comprises:

Under stable constraint, use independent component analysis (ICA) or blind source to separate (BSS), handle described mixed signal group; And

Described mixed signal is separated into described voice signal and at least one noise signal;

And

The voice activation unit is used to receive described voice signal.

53,, wherein, described signal generator is constituted acoustic transducer according to the described system of claim 52.

54, according to the described system of claim 53, wherein, described acoustic transducer is to constitute the microphone that receives the voice signal in the human voice frequency range.

Images