EP0912974B1 - Method of reducing voice signal interference - Google Patents

Abstract

In a method for reducing interferences in a voice signal, a noise reduction method is applied to the voice signal, and spectral psychoacoustic masking is taken into account. A spectral masking curve is determined both for the input signal and the output signal of the noise reduction method. By comparing the signal portions exceeding the respective masking curve, newly-audible portions are detected in the form of interference in the output signal and subsequently damped selectively.

Description

The invention relates to a method for reducing Interference with a speech signal.

Such a method can advantageously be used for Interference-free speech signals for speech communication, especially hands-free kits e.g. in motor vehicles, Find speech recognition systems and the like.

A commonly used method to reduce the amount of noise in speech signals with interference, the so-called spectral subtraction. This procedure has the advantage of simple, low-effort implementation and a significant reduction in noise.

An unpleasant side effect of noise reduction by means of spectral subtraction, the occurrence of is short-term audible tonal noise components, which due to the mediated hearing impression as "musical tones" or "musical noise ".

Measures to suppress "musical tones" at the spectral subtraction are the overestimation of the interference power overcompensation of the disturbance with the disadvantage the increased speech distortion or allowing one relatively high noise level with the disadvantage of only one low noise reduction (e.g. "Enhancement of Speech Corrupted by Acoustic Noise "by Berouti, M .; Schwartz, R .; Makhoul, J .; in Proceedings on ICASSP, pp. 208-211, 1979). Methods for linear or non-linear smoothing and thus to suppress the "musical tones" e.g. in "Suppression of Acoustic Noise in Speech Using Spectral Subtraction "by S.F. Boll in IEEE Vol. ASSP-27, No. 2, pp. 113-120. An effective non-linear Smoothing with median filtering is in the DE 44 05 723 A1 specified.

Methods which are in addition to the spectral are also known Subtraction take into account the psychoacoustic perception (e.g. T. Petersen and S. Boll, "Acoustic Noise Suppression in a Peceptual Model "in Proc. On ICASSP, pp. 1086-1088, 1981). The signals are in the The field of psychoacoustic loudness is transformed all the more to carry out more hearing-appropriate processing. From D.

Tsoukalas, P. Paraskevas and M. Mourjopoulos is published in "Speech Enhancement Using Psychoacoustic Criteria", Proc. on ICASSP, pp. II359-II362, 1993 and by G. Virag in "Speech Enhancement Based on Masking Properties of the Auditory System ", Proc. On ICASSP, pp. 796-799, 1995 uses the calculated masking curve to determine which spectral lines are covered by the useful signal and therefore do not need to be dampened. The quality of the Speech signal is thus improved. The disturbing "musical tones "are not reduced.

The object of the present invention is an improved Method of reducing speech signal interference specify.

The invention is described in claim 1. The subclaims contain advantageous configurations and Developments of the invention.

The invention is essentially based on the fact that signal components, which can only be heard individually through noise reduction appear, recognized as disturbances and afterwards reduced or eliminated by selective damping become. The audibility criterion is in is known to exceed a masking curve (masking threshold).

The determination of masking curves is e.g. from parts of the aforementioned prior art, in more detail general form also e.g. from sound engineering, Cape. 2., psychoacoustics and noise assessment (pp. 10-33), Expert Verlag known in 1994. The determination of the masking curves can both based on the current Speech signals as well as on the basis of a noise signal take place during speech pauses, with different psychoacoustic Effects can be taken into account. The masking curves, which are also called masking curves, listening thresholds, masking threshold and similar in the specialist literature can be referred to as a frequency-dependent level threshold for the perceptibility of a narrow-band tone be considered.

• Kritische Bandanalyse, bei welcher das Spektrum eines Signals in sogenannte kritische Bänder aufgeteilt und aus dem Leistungsspektrum P(i) durch Aufsummierung innerhalb der kritischen Bänder ein kritisches Band-Spektrum B(n) (auch Bark-Spektrum, mit n als Bandindex) gewonnen wird
• Faltung des Bark-Spektrums mit einer Verbreiterungsfunktion (Spreading-Funktion) zur Berücksichtigung der Verdeckungseffekte über mehrere kritische Bänder hinweg; man erhält ein modifiziertes Bark-Spektrum
• evtl. zusätzliche Berücksichtigung der unterschiedlichen Verdeckungseigenschaften von rauschhaften und tonhaften Anteilen durch einen aus der Zusammensetzung des Signals bestimmten Offsetfaktor
• Nach Renormierung im Verhältnis zur jeweiligen Energie in den kritischen Bändern und ggf. Anhebung tieferliegender Werte auf die Werte der Ruhehörschwelle ergibt sich eine barkbezogene Verdeckungskurve T(n) und daraus eine frequenzbezogene Verdeckungskurve V(i) mit V(i) = T(n) für alle Frequenzen i innerhalb des jeweiligen kritischen Bandes n

Such concealment curves are used in addition to interference-free applications, for example also for data reduction when encoding audio signals. A detailed procedure for determining a masking curve is, in addition to the publications already mentioned, for example from "Transform Coding of Audio Signals Using Perceptual Noise Criteria" by J. Johnston in IEEE Journal on Select Areas Commun., Vol. 6, pp. 314-323, Feb. 1988. The essential steps of a typical method for determining a masking curve from the short-term spectrum of a disturbed speech signal are in particular

• Critical band analysis, in which the spectrum of a signal is divided into so-called critical bands and a critical band spectrum B (n) (also Bark spectrum, with n as the band index) is obtained from the power spectrum P (i) by adding up within the critical bands
• Folding the Bark spectrum with a spreading function to take into account the masking effects across several critical bands; a modified Bark spectrum is obtained
• Possibly additional consideration of the different masking properties of noisy and tonal parts by an offset factor determined from the composition of the signal
• After renormalization in relation to the respective energy in the critical bands and, if necessary, lowering values to the values of the resting hearing threshold, a bark-related masking curve T (n) results and from this a frequency-related masking curve V (i) with V (i) = T (n) for all frequencies i within the respective critical band n

With the determined masking curve V (i) the spectral components of the signal by comparing the range of services P (i) with the masking curve V (i) in audible, (P (i)> V (i)) and hidden (P (i) <V (i)) parts become.

FIG. 1

ein Blockschaltbild eines Standardverfahrens zur spektralen Subtraktion

FIG. 2

ein Blockschaltbild zu einem Verfahren nach der Erfindung

FIG. 3

ein Sprachsignal in verschiedenen Stufen des erfindungsgemäßen Signalverarbeitungsverfahrens.

The invention is illustrated below by means of examples with reference to the figures. It shows

FIG. 1

a block diagram of a standard method for spectral subtraction

FIG. 2nd

a block diagram of a method according to the invention

FIG. 3rd

a speech signal in different stages of the signal processing method according to the invention.

The methods for spectral subtraction are based on processing the short-term magnitude spectrum of the disturbed input signal. During speech pauses, the interference power spectrum is estimated and then subtracted in phase from the disturbed input signal. This subtraction is usually carried out as filtering. The filtering results in a weighting of the disturbed spectral components with a real factor, depending on the estimated signal-to-noise ratio of the respective spectral band. The noise reduction therefore results from the fact that disturbed spectral regions of the usage signal are damped in the ratio of their interference component. A simplified block diagram in FIG. 1 shows a typical implementation of the spectral subtraction algorithm. In an analysis stage, the disturbed speech signal is broken down, for example by a discrete Fourier transformation (DFT), into a series of short-term spectra Y (i). The unit KM forms a short-term mean value from the Fourier coefficients, which represents an estimated value for the mean power Y ² (1) with i as the discrete frequency index of the disturbed input signal. In a unit LM, controlled by the speech pause detector SP, an average interference power spectrum N ² (1) is estimated in the speech-free sections. Each spectral line Y (i) of the input signal is then multiplied by a real filter coefficient H (i), which is calculated from the short-term mean value Y ² (1) and the interference power mean value N ² (1) in the unit FK. The process step of noise reduction is shown as the multiplication level GR. An inverse discrete Fourier transformation (IDFT) results in the noise-reduced speech signal at the output of the synthesis stage.

The filter coefficients H (i) can be calculated according to different weighting rules known per se. The coefficients are typically estimated according to H (i) = max {(1- N 2nd (i) / Y 2nd (i) ), fl}

With fl as the predeterminable basic value (also spectral floor), which represents a lower bound for the filter coefficients and is usually 0.1 <fl <0.25. He decides one in the output signal of the spectral subtraction remaining amount of residual noise that the lowering of the Listening threshold limited and so narrowband portions in noise-reduced output signal of the spectral subtraction partially covered. Compliance with a core value fl improves the subjective hearing impression.

To cover up any residual disturbances of the kind of "musical tones "a base value of approx. 0.5 would have to be chosen, which means the maximum achievable noise reduction to about 6dB would be limited.

A characteristic used in the method according to the invention The characteristic of musical tones is that they in the output signal of the noise reduction process for that human ear perceptible as a disturbance in appearance to step. The second concealment curve makes it noticeable quantified for this output signal become. Compared to the also the level threshold of second concealment curve exceeding language useful portions in the output signal, which is already in the input signal as exceeding the level of the first masking curve The musical tones can be perceived by comparison the perceptible signal components in the output signal and input signal noise reduction as new audible components distinguished and in a subsequent processing step be selectively damped selectively.

The detection and suppression method according to the invention of narrowband interference like musical tones is based on the block diagram in FIG. 2 explained. It represents an extension of the in FIG. 1 standard procedure shown for spectral subtraction. As far as that outlined method in FIG. 2 with the in FIG. 1 outlined known methods matches, are the same reference numerals related. From the input signals Y (i) of Noise reduction GR becomes the first in a VE unit Masking curve V1 (i) determined. From the output signals Y '(i) of noise reduction becomes a second masking curve in VA V2 (i) determined.

Alternatively, the first masking curve V1 (i) can also from the mean interference power spectrum at the entrance of the noise reduction can be determined during breaks in speech. The second masking curve can also be from the first masking curve derived, e.g. by multiplication with the basic value fl, V2 (i) = fl ° Vl (i).

The advantage of determining the masking curves from the current input and output signals of noise reduction consists in particular that also transient Noise components and the concealing effect of the speech components be taken into account. In contrast, becomes the first Masking curve from the medium interference power spectrum determined and the second masking curve approximately determined according to V2 (i) = fl · V1 (1), the result is considerable Reduction of the computing effort. The computing effort can be further reduced in that the masking curve need to be updated much less often, since the average interference power spectrum is usually only is slowly changing over time. The better quality synthesized Speech signal is however with the determination of the Masking curves from the current signals Y (i), Y '(i) achieved.

An advantageous development of the invention provides one further improvement by detection of stationary signal components, exempt from selective damping even if they only meet the criterion in the output signal Y '(i) to be perceivable. In FIG. 2 is for this a stationarity detector STAT is drawn.

It can be implemented in various ways, for example by tracking individual spectral lines over time or the filter coefficients. A simple form of implementation results from the requirement that a plurality of successive filter coefficients each have to exceed a certain threshold value thr _stat , so that: H kn (i), ..., H small (i), H k (i)> thr stat , with e.g. n = 2 and thr _stat = 0.35.

In the decision maker ENT, audible tonal components in the output signal of the noise reduction system are first determined with the aid of the second masking curve V ₂ (i). If this is not a stationary component, it is examined whether the spectral component was audible before filtering (noise reduction). This is done using the first masking curve V _l (i). If the frequency component in the input signal Y (i) is found to be hidden, the spectral component in the output signal is assumed to be a musical tone and is attenuated in a post-processing stage NV. In the other case, ie if the input signal is not masked, a decision is made for speech and no additional attenuation is carried out.

The additional damping in post-processing can done in different ways. For example, for one as Fault detected newly audible spectral component of the level value set to the value of the second masking curve become. The detected level value is preferably the interfering spectral component to a corrected Set value resulting from the filtering of the spectrally corresponding Input signal component with the basic value fl results as a filter coefficient.

In FIG. 3 are different stages of signal processing for a disturbed speech signal according to the invention Procedure outlined.

FIG. 3A shows a power spectrum P (i) of a disturbed signal at the input of the noise reduction and a first masking curve V1 (i) with the signal components s exceeding the masking curve. After performing the spectral subtraction, there is a noise-reduced power spectrum P '(i) = Y' ² (i) with a second masking curve V2 (i) determined therefrom, in which, in addition to the also shown in FIG. 3A, the signal components exceeding the masking curve V1 (i) s further signal components m exceeding the second masking threshold, which appear as non-masked and thus newly audible signal components in the manner of the musical tones. These newly audible signal components can be detected and suppressed by selective attenuation without impairing the speech components s. The power spectrum P "(i) resulting from the selective damping is sketched in FIG. 3C. Only the signal components s evaluated as speech signals exceed the masking curve, these signals now being a much larger extent than masking curve V2 (i) corresponding portions in the input signal above the masking curve V1 (i) (FIG. 3A) and are therefore more clearly audible The musical tones m from FIG. 3B are pressed below the masking curve V2 (i) and are therefore no more than individual tones noticeable.

The invention is not for spectral subtraction Noise reduction limited. The process, the masking curves at the entrance and at the exit of a noise reduction to be determined and based on newly audible shares in To detect and suppress output disturbances, lets other signal processing systems, e.g. to Transfer signal coding.

Claims (9)

1. Method of reducing disturbances of a speech signal, in which a noise-reducing process is used and the spectral psychoacoustic masking is taken into consideration, characterised in that a first spectral masking curve for the input signal and a second spectral masking curve for the output signal of the noise-reducing process are determined and that newly audible portions, which exceed the second masking curve and opposite which there are no spectrally corresponding input signal portions exceeding the first masking curve, of the output signal of the noise-reducing process are in addition selectively damped.
2. Method according to claim 1, characterised by a spectral subtraction process as noise-reducing process.
3. Method according to claim 2, characterised in that the newly audible portions are reduced to their fundamental value of the spectral subtraction.
4. Method according to claim 1 or claim 2, characterised in that the newly audible portions are reduced to their value of the spectral masking curve.
5. Method according to one of claims 1 to 4, characterised in that static newly audible portions of the output signal are excluded from the additional selective damping by way of a presettable time interval.
6. Method according to one of claims 1 to 5, characterised in that the second masking curve is determined from the output signal of the noise-reducing method.
7. Method according to one of claims 1 to 5, characterised in that the second masking curve is derived from the first masking curve.
8. Method according to one of claims 1 to 7, characterised in that the first masking curve is determined from the input signal of the noise-reducing process.
9. Method according to one of claims 1 to 8, characterised in that the first masking curve is determined from noise signals in pauses in speech.

Images