EP0669606A2 - Method for noise reduction in disturbed voice channels - Google Patents

Abstract

Filtering is applied to the band spectrum of the noisy signal before and optionally also after spectral subtraction of the noise is applied. The required transfer function for the spectral subtraction process is obtained during speech pauses and may itself be subjected to median filtering. The filter may operate in the time domain over a window of 3 successive time segments, in total less than 50 ms. Alternatively, the filter operates in the frequency domain with a window size of less than 5 frequency units.

Description

The invention relates to a method for noise reduction of a disturbed speech signal with the aid of spectral subtraction.

The noise reduction using the spectral subtraction method is used in automatic speech recognition or in hands-free systems to improve speech quality, e.g. when calling from the motor vehicle.

The noise reduction by spectral subtraction is characterized in that relatively stationary interference can typically be reduced by approximately 10 dB without additional information about the interference being required. Only the disturbed voice channel is required. The speech signal is divided into short overlapping time segments and processed segment by segment. In the case of spectral subtraction, an estimate of the interference is determined in the speech pauses, and this estimate is subtracted in terms of amount in the spectral range. The spectral subtraction can be implemented in different ways, but is usually implemented as a multiplicative filter in the frequency domain. This spectral subtraction shows the undesirable side effect of a residual musical noise, the "musical tones" and a speech distortion.

"Musical tones" are usually suppressed by excessive damping. The excessive damping can be done by overestimating the interference with an overestimating factor or by choosing a special transmission characteristic. The values of the current transfer function are determined for each frequency from the transfer characteristic. It is common to implement a magnitude characteristic in the spectral subtraction filter that has a higher attenuation than e.g. a characteristic curve based on the quadratic error criterion. Specially designed characteristic curves are also possible. Depending on the characteristic used, an overestimation of the disturbance by a factor of 1 to 3 is common. The excessive damping due to the characteristic curve and the overestimation factor gives the desired effect of suppressing "musical tones", but it also has the side effect of a partly. considerable language distortion.

Another common method of suppressing "musical tones" is masking by allowing a certain proportion (e.g. 20%) of the original noise as background noise ("spectral floor"). This makes "musical tones" less audible, but the sound is no longer completely suppressed.

mit

und für das Beispiel einer sogenannten Betragskennlinie als Übertragungskennlinie

sowie beispielsweise die Auswahl eines minimalen Übertragungswertes für den spectral floor

Mit den Größen:

• S: geschätztes Ausgangssignal
• K: Übertragungsfunktion
• Y: gestörtes Sprachsignal
• S: Sprachsignal
• N: Störgeräusch
• b: Hintergrundrestgeräusch (spectral floor)
• a: Überschätzfaktor (overestimate)
• N ²: in Sprachpausen geschätzte Störung
• i: Frequenzindex
• I: Zeitindex des Segments

Methoden zur Unterdrückung der "musical tones", durch Kennlinie, "overestimation" und "spectral floor", sind in vielfältiger Variation durch zahlreiche Veröffentlichungen bekannt, z.B.:

• Boll, S.: Suppression of Noise in Speech Using the SABER Method, Proc. IEEE Int. Conf. on ASSP, 1978, pp. 600-609.
• Boll, S.: Suppression of Acoustic Noise in Speech Using Spectral Substraction, IEEE Trans. on ASSP, Vol. ASSP-27, No. 2, April 79, pp. 113-120.
• Berouti, M.; Schwartz, R.; Makhoul, J.: Enhancement of Speech Corrupted by Acoustic Noise, Proc. Int. Conf. on ASSP, 1979, pp. 208-211.
• Vary, P.: Noise Suppression by Spectral Magnitude Estimation - Mechanism and Theoretical Limits-, Signal Processing, Vol. 8, No. 4, 1986, pp. 387-400.
• Xie, F.; Compernolle.: Speech Enhancement by Nonlinear Spectral Estimation - A Unifying Approach, Int. Conf. Eurospeech, 1993, pp. 617-620.

It applies to spectral subtraction

With

and for the example of a so-called amount characteristic as a transfer characteristic

as well as, for example, the selection of a minimum transmission value for the spectral floor

With the sizes:

• S: estimated output signal
• K: transfer function
• Y: disturbed speech signal
• S: voice signal
• N: noise
• b: background noise (spectral floor)
• a: overestimate factor
• N ² : Estimated disturbance during pauses in speech
• i: frequency index
• I: Time index of the segment

Methods for suppressing the "musical tones" by means of characteristic, "overestimation" and "spectral floor" are known in numerous variations through numerous publications, for example:

• Boll, S .: Suppression of Noise in Speech Using the SABER Method, Proc. IEEE Int. Conf. on ASSP, 1978, pp. 600-609.
• Boll, S .: Suppression of Acoustic Noise in Speech Using Spectral Substraction, IEEE Trans. On ASSP, Vol. ASSP-27, No. April 2, 79, pp. 113-120.
• Berouti, M .; Schwartz, R .; Makhoul, J .: Enhancement of Speech Corrupted by Acoustic Noise, Proc. Int. Conf. on ASSP, 1979, pp. 208-211.
• Vary, P .: Noise Suppression by Spectral Magnitude Estimation - Mechanism and Theoretical Limits-, Signal Processing, Vol. 8, No. 4, 1986, pp. 387-400.
• Xie, F .; Compernolle .: Speech Enhancement by Nonlinear Spectral Estimation - A Unifying Approach, Int. Conf. Eurospeech, 1993, pp. 617-620.

In addition to the methods mentioned above, other special methods are known which are also used to reduce the "musical tones": the amplitude values of temporally successive disturbed speech spectra are averaged (e.g. Boll "magnitude averaging"). This dampens noise components, but since speech is highly unsteady, the speech signal is smeared over time even with short averaging lengths (echo-like effect). Boll continues to describe a "magnitude plus bandwith measurement test" after spectral ranges with a bandwidth below 300 Hz and an amplitude smaller than a predetermined threshold are recognized as "residual noise". These areas are then additionally damped. It is proposed by Boll to reduce the "residual noise" by using the minimum value as the output signal from three temporally successive spectra of the filtered signal. The output of the minimum spectral line from three adjacent lines leads to a significant reduction in the residual noise and thus the "musical tones", but occasionally sudden, short "noise bundles" occur at irregular intervals.

Another method uses so-called nonlinear spectral subtraction. The overestimation factor is calculated here depending on the pause noise and the current signal. However, the optimal setting of this regulation is difficult. (Lockwood, P .; Boudy, J .: Experiments with a Nonlinear Spectral Subtraction (NSS), Hidden Markov Models and the projection, for robust speech recognition in cars, Speech Communication, No. 11, 1992, p. 215-228) .

It is an object of the present invention to provide a method for reducing the noise of a disturbed speech signal which, with a high speech quality of the output signal, enables a strong reduction in the noise, in particular also the "musical tones".

Solutions to this task according to the invention as well as advantageous designs and further developments are described in the patent claims.

Median filtering has proven to be an advantageous method for further substantially improving the spectral subtraction method for reducing the noise of a disturbed speech signal. The median filtering can be applied both to the magnitude spectrum of the disturbed input signal or the noise-reduced output signal after spectral subtraction and to the transfer function determined from the application of a transfer characteristic and can be carried out in the time direction or in the frequency direction. The magnitude spectrum of the speech signal is composed of a sequence of segment spectra in accordance with the segmentation of the speech time signal. The transfer function is represented by the time and frequency discrete values K _i , ₁ (eg equation (3)). A combination of different of these procedures can also be advantageous. Thus, a preferred method provides, in speech pauses by using median filtering, preferably in the temporal direction, to preserve the natural impression of a weak background noise on the transmission function, and during speech activity by applying median filtering to the spectrum of magnitude of the speech signal to strongly suppress the "musical tones" to reach. The separate detection of speech pauses and speech activity is provided and known in any case for determining an average noise signal during speech pauses, so that no special effort is required for this. The methods according to the invention are easy to implement.

The principle of median filtering is generally known (e.g. Mitra, S.K .: Handbook for Digital Signal Processing, John Wiley & Sons, 1993).

Median filtering is also already known for processing speech signals. In DE 32 43 231 A1 and DE 32 43 232 A1, for example, a median filter for successive short-term mean values, which represent a measure of the average power of speech signal sections, is used as a smoothing filter. Speech pauses are recognized by comparing the smoothed sequence of values with a threshold value. A This does not result in interference-free speech signals.

In IEEE Transactions on Acoustics, Speech and Signal Processing, Vol. ASSP-23, No. 6, Dec. 1975, pp. 552-557 describes the application of a median filter in combination with a linear smoothing filter to samples of the intensity of a speech signal. Signal processing in the spectral range is not provided and only a slight interference superposition can be managed.

FIG. 1 shows an example of an input signal E and an output signal A filtered with a median filter of length 3. The median filter first sorts the values within the data window F and then outputs the mean value med. The median filter hides short signal peaks, but receives the remaining signal edges.

oder bei Filterung in Frequenzrichtung

Der Filterung am Betrag ist die Filterung am Betragsquadrat im Prinzip gleichwertig.For the example of the application of a median filter of length 3 to a noise-reduced magnitude spectrum of a speech signal, time filtering applies

or with filtering in the frequency direction

Filtering on the amount is basically equivalent to filtering on the amount square.

The effect of median filtering on the reduction of the "musical tones" is illustrated by means of representations of a typical temporal and spectral course of such "musical tones". The spectrum of magnitudes of an output signal obtained in a speech pause and estimated with the aid of the spectral subtraction is shown. Since there are no parts of the language during the language break, especially the "musical tones" appear clearly.

The following was used as an example of spectral subtraction: standard method with magnitude characteristic, 20% background noise (b = 0.2), without overestimation factor (a = 1.0).

The following noise example was used: vehicle interior noise at 140 km / h, 12 kHz sampling frequency, segment length 512 values, the last 256 values of each segment are set to zero, the first 256 values of each segment are multiplied by the Hanning window, segments are half overlapped, i.e. a new segment every 10.67ms.

FIG. 2 shows first the frequency (linear 0 to 6 kHz) the spectrum for 4 successive segments (time interval 10.67 ms, index I) and then the time (0 to 2.5 seconds) the signal curve for 4 successive discrete frequencies (index i ), representative of all 256 frequencies. It is shown as a typical property of the "musical tones" that the course over the frequency has relatively extensive disturbances (broad impulses), whereas the course over time has a strong impulsive character (narrow impulses). It is precisely the pulse-like character in the temporal direction that makes median filtering particularly effective. A pulse-like fault is deleted. A longer window length of the median filter is required for pulse-like disturbances with wider pulses. In contrast to linear filtering methods (smoothing filters, "linear smoother") there is no smearing of the signal curve. The representation of the signals filtered in time with the 3-fold median in FIG. 3 illustrates this property. The filtered signal clearly shows a smoother course over time. In the frequency response, some of the (broader) pulses are also deleted by filtering in the time direction.

In the case of speech activity, the use of the median filter in the temporal direction of the individual spectral lines leads to an improvement in the speech quality, since impulsive disturbances in the speech spectrum are "repaired". The voice signal itself is changed very little. Increasing the window length from 3 to 5 (in the time direction) results in an even better extinction of the "musical tones", but a weak echo-like character of the language is already audible.

Instead of the output signal, the median filter can also be carried out on the input signal before the spectral subtraction. Ideally, this does not result in the creation of "musical tones" that would otherwise be solved by post-filtering with the median filter. The median filtering on the input signal can then be advantageous if "musical tones" influence the various processing steps implemented in the spectral subtraction filter (apart from the characteristic function). It shouldn't go on possible advantages or disadvantages of median filtering on the input or output signal are discussed. In principle, both options are available and, apart from special cases of implementation, are equivalent.

The median filter can also be carried out on the transfer function K instead of on the magnitude spectrum of a speech signal.

oder

FIG. 4 zeigt die Übertragungsfunktion K über der Zeit und über der Frequenz. Dargestellt ist der gleiche Ausschnitt wie in FIG. 2. Die Übertragungsfunktion zeigt ein ähnliches Verhalten wie das Ausgangssignal in FIG. 2.The following applies to the median of 3:

or

FIG. 4 shows the transfer function K over time and over frequency. The same section is shown as in FIG. 2. The transfer function shows a similar behavior as the output signal in FIG. 2nd

FIG. 5 shows the transfer function filtered in time with the 3-fold median. The same section is shown as in FIG. 3. Here too, the median filtering in the temporal direction is extremely effective for the same reasons as for the output signal.

• Ein Eingangssignal mit einer impulsartigen Störung verursacht die entsprechende impulsartige Änderung der Übertragungsfunktion. Im ursprünglichen Geräusch gehört dieser lokale Impuls zum natürlichen Geräusch und wird deshalb nicht als besonders störend empfunden. Das Spektrum des Eingangssignals wird mit der Übertragungsfunktion multipliziert. Die impulsartige Störung wird dadurch zusätzlich verstärkt ist jetzt als "musical tone" hörbar.

The effective suppression of the "musical tones" by median filtering can be explained as follows:

• An input signal with a pulse-like disturbance causes the corresponding pulse-like change in the transfer function. In the original sound, this local impulse is part of the natural sound and is therefore not perceived as particularly disturbing. The spectrum of the input signal is multiplied by the transfer function. The pulse-like disturbance is additionally amplified and can now be heard as a "musical tone".

The pulse-suppressing property of median filtering has a particularly significant effect on the increased impulse disturbance and thus on the "musical tones". Median filtering has a repairing effect on the pulse-like disturbance.

Compared to the median filtering on the transmission values, the median filtering on the magnitude spectrum of the input or output signal results in the higher gain in the suppression of impulse-like disturbances, but can also lead to changes which are particularly noticeable during speech pauses, while the median filtering of the transmission values in speech pauses essentially increases a pure attenuation of the signal, which makes it sound quieter but natural. Ideally, there are no "musical tones". A preferred embodiment of the invention takes advantage of this by performing the median filtering in the case of speech activity on the magnitude spectrum and in speech pauses on the transmission values. The required speech-pause decision is available anyway with the spectral subtraction, since the formation of the noise estimate is only carried out during the speech pauses.

Instead of the median filtering in the time direction as described, a median filtering in the frequency direction can also be carried out in accordance with equation (6). The detailed explanations given apply analogously to filtering in the frequency direction. It can be seen that as the number of samples within a time segment decreases, the median filtering in the frequency direction gains advantages over the filtering in the time direction and vice versa.

In the case of the customary values for the segment length according to time and sample values, the use of median filtering in the time direction is particularly advantageous.

In the described application of median filtering in the temporal direction with the values for sample rate and window length given by way of example, the window length, as indicated in the example, is equal to the minimum median window length 3. In this case, larger window lengths lead to a further suppression of the "musical tones", but under certain circumstances also to a leveling of the speech signal which is perceived as unnatural. The preferred window length is therefore 3 as indicated by way of example. For segments that are shorter in time, a larger window length can be appropriate for median filtering. However, the time interval covered by the median temporal filtering window should not exceed 50 ms ten.

For filtering in the frequency direction, the window length of the median filter is based on the data segment length. In the numerically described example, the data segment length should be less than 64, the median filter not greater than 5.

Claims (11)

1. A method for noise reduction of a disturbed speech signal with the aid of spectral subtraction, characterized in that the magnitude spectrum of the speech signal is subjected to median filtering. 2. The method according to claim 1, characterized in that the median filtering is applied to the magnitude spectrum of the disturbed input signal. 3. The method according to any one of claims 1 and 2, characterized in that the median filtering is applied to the magnitude spectrum of the output signal of the spectral subtraction. 4. A method for noise reduction of a disturbed speech signal with the aid of spectral subtraction, a transfer function for the spectral subtraction being determined from a predefinable transfer characteristic, characterized in that the transfer function is subjected to median filtering. 5. A method for noise reduction of a disturbed speech signal with a combination of the preceding claims. 6. The method according to claim 5, characterized in that the median filtering in speech pauses is applied to the transmission values and in speech activity to the magnitude spectrum of the speech signal. 7. The method according to any one of claims 1 to 6, characterized in that the median filtering is applied in the temporal direction. 8. The method according to claim 7, characterized in that the window length of the median filter comprises three successive time segments. 9. The method according to claim 7 or claim 8, characterized in that the window length of the median filter is less than 50ms. 10. The method according to any one of claims 1 to 6, characterized in that the median filtering is applied in the frequency direction. 11. The method according to claim 10, characterized in that the window length of the median filter comprises no more than 5 frequency values.

Images