DE4405723A1 - Method for noise reduction of a disturbed speech signal - 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 a disturbed speech signal using the spectral Subtraction.

The noise reduction fits the spectral sub method traction is used in automatic speech identifier or in hands-free systems to improve the Voice quality, e.g. B. when calling from the motor vehicle stuff.

The noise reduction by spectral subtraction draws is characterized by the fact that relatively stationary  Interference can typically be reduced by approx. 10dB NEN without additional information about the fault is needed. Only the disturbed voice channel is used does. The speech signal is in short overlapping time segment elements divided and processed segment by segment. In the spectral subtraction is used in the speech pauses Estimate of the fault, and this estimate is subtracted in the spectral range. The spectral subtraction is real in different ways can be used, but is usually used as a multiplicative filter implemented in the frequency domain. This spectral subtrak tion shows the undesirable side effect of a musical residual noise, the "musical tones" and a speech distortion.

Usually "musical tones" are exaggerated Damping suppressed. The excessive damping can be caused by an overestimation of the disturbance with an overestimation factor done or by choosing a special transfer characteristic curve. The transmission characteristic curve becomes each frequency the values of the current transfer function certainly. It is common in the spectral subtraction filter implement an amount characteristic that is higher Damping has z. B. a characteristic curve after the quadra table error criterion. Specially designed characteristic curves are also possible. Depending on the identifier used line is an overestimation of the disturbance by a factor of 1 up to 3 usual. The excessive damping due to the characteristic and the overestimation factor gives the desired Ef effect of the suppression of "musical tones", but also has the side effect of a z. T. considerable distortion of the Language.  

Another common way to suppress "musical tones" is masking by allowing a specific one Share (e.g. 20%) of the original noise as a hint background noise ("spectral floor"). "musical tones" who which makes it less audible, but the noise also becomes no longer completely suppressed.

It applies to spectral subtraction

_{i, l} = K _{i, lYI , l} (1)

With

Y _{i, l} = S _{i, l} + N _{i, l} (2)

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

and, for example, the selection of a minimal transfer values for the spectral floor

Min (K _{i, l} ) = b (4)

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

Methods to suppress the "musical tones", by Characteristic curve, "overestimation" and "spectral floor", are in diverse variation through numerous publications known, e.g. 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. 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 Estima tion - Mechanism and Theoretical Limits-, Signal Proces sing, vol. 8, no. 4, 1986, pp. 387-400.

Xie, F .; Compernolle .: Speech Enhancement by Nonlinear Spectral Estimation - A Unifying Approach, Int. Conf. Eu rospeech, 1993, pp. 617-620.  

Beyond the methods mentioned above, there are others known special methods that are also used for reduction the "musical tones" can be used:

The amplitude values are interfered with one another in time ter speech spectra are averaged (e.g. Boll "magni tude averaging "). This does in fact generate noise components dampens but since language is highly transient, it already occurs with short averaging lengths a time smear of the speech signal on (echo-like effect). At Boll still a "magnitude plus bandwith measurement test" described after spectral ranges with a band width below 300 Hz and an amplitude less than one predetermined threshold, can be recognized as "residual noise".

These areas are then additionally damped. It will von Boll proposed to reduce the "residual noise" reduce that from three consecutive times Spectra of the filtered signal the minimum value is used as the output signal. The output of the minima len spectral line of three temporally adjacent lines leads to a significant reduction in the residual ge noise and thus the "musical tones", occasionally kicking however, sudden short "Ge bundle of noise ".

Another method uses a so-called non-li linear spectral subtraction. The overestimation factor is depending on the pause noise and the current concern calculated the signal. The optimal setting of this re However, success is difficult. (Lockwood, P .; Boudy, J .: Ex periments 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).

The object of the present invention is a method to reduce the noise of a disturbed speech signal admit which with high speech quality of the output si gnals a strong reduction in noise, in particular the "musical tones" also makes it possible.

Solutions according to the invention of this task and of advantage sturdy designs and further training are in Pa claims described.

The median filtering proves to be an advantageous method for further substantially improving the method of spectral subtraction for the noise reduction 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 transmission function determined from the application of a transmission characteristic and can be carried out in the time direction or in the frequency direction. The transfer function is represented by the time and frequency discrete values K _{i, l} (e.g. equation (3)). A combination of various of these procedures can also be advantageous.

This is a preferred procedure during language breaks by using median filtering, preferably in time direction, on the transfer function the natural to give the impression of a weak background noise true and during speech activity by using the Median filtering on the magnitude spectrum of the speech signal  a strong suppression of the "musical tones" too to reach. The separate recognition of speech pauses and Speech activity is used to determine a mean Ge noise signal provided anyway during speech pauses and known, so that this requires no special effort is. The methods of the invention are simple can be implemented.

The principle of median filtering per se is general knows (e.g. Nitra, S.K .: Handbook for Digital Signal Pro cessing, John Wiley & Sons, 1993).

Fig. 1 shows an example for an input signal E and a filtered with a median filter of length 3 shows Ausgangssi gnal A. The median filter sorted first the values within half of the data window F, and then outputs the average value of med. The median filter hides short signal peaks, but receives the remaining signal edges.

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

| _{i, l, m} | = med (| _{i, l-1} |, | _{i, l} |, | _{i, l + 1} |) (5)

or with filtering in the frequency direction

| _{i, l, m} | = med (| _{i-1, l} |, | _{i, l} |, | _{i + 1, l} |) (6)

The amount filtering is the amount filtering in principle equivalent.  

The effect of median filtering on reducing the "musical tones" is illustrated by means of representation against a typical temporal and spectral course such "musical tones". The amount spectra is shown dream of a language break obtained with help the spectral subtraction of the estimated output signal. There there are no speech parts during the language break especially the "musical tones" clearly appear.

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

The following was used as a sound example: vehicle interior noise at 140 km / h, 12 kHz sampling frequency, segment length 512 Values, the last 256 values of each segment become zero are set, the first 256 values of each segment are marked with Hanning window multiplied, segments are half over laps, d. H. a new segment every 10.67 ms.

Fig. 2 shows first the frequency (linear 0 to 6 kHz), the spectrum for 4 successive segments (time interval 10.67 ms, index 1) and then over time (0 to 2.5 sec) the signal curve for 4 successive discrete frequencies (index i), representative of all 256 frequencies. It shows up 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 pulse-like character (narrow impulses). It is precisely the pulse-like character in the temporal direction that makes median filtering particularly effective here. A pulse-like fault is deleted. A longer window length of the media filter is required for pulse-like interference with wider pulses. In contrast to linear filtering methods (smoothing filter, "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.

When there is voice activity, the median filter is used in the temporal direction of the individual spectral lines ner improvement of speech quality, since impulsive disturbances of the language spectrum are "repaired". The voice signal itself is changed very little. An increase hung the window length from 3 to 5 (in the time direction) although an even better erasure of the "musical tones", it but is already a weak echo-like character of the Speech audible.

The median filter can also be used instead of the output signal Input signal, prior to spectral subtraction leads. Ideally, this will not result in any "musical tones ", which otherwise alternate through the post-film can be solved with the median filter. The median file The input signal can be advantageous if "musical tones" the various implementations implemented processing steps in the spectral subtraction filter (except the characteristic function). It should go on not on possible advantages or disadvantages of a median filter  tion on the input or output signal. in the In principle, both possibilities are given and of special len cases of implementation apart from equivalent.

The median filter can be used instead of the magnitude spectrum Speech signal also on the transfer function K out leads.

The following applies to the median of 3:

| K _{i, l, m} | = med (| K _{i, l-1} |, | K _{i, l} |, | K _{i, l + 1 | )} (7)

or

| K _{i, l, m} | = med (| K _{i-1, l} |, | K _{i, l} |, | K _{i + 1, l | )} (8)

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 the temporal direction with the 3-series media. The same section is shown as in Fig. 3. Here, too, the media filtering in the temporal direction is extremely effective for the same reasons as for the output signal.

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

An input signal with a pulse-like disturbance gently the corresponding pulse-like change in the transfer  supply function. In the original sound it belongs local impulse to natural sound and therefore becomes not perceived as particularly disturbing. The spectrum of the Input signal is with the transfer function multiplied. The impulse-like disturbance thereby becomes additionally amplified is now audible as a "musical tone".

The pulse-suppressing property of median filtering affects particularly clearly on the amplified impulse malfunction and thus on the "musical tones". The median filtering has a repairing effect on the pulse-like disturbance.

The median filtering on the magnitude spectrum of the input or Output signal indicates compared to the median filtering the transfer values the higher profit on the sub pressure from impulsive disturbances, but can also particularly noticeable in speech pauses as unnatural perceived changes while the median filters Essentially, the transmission values during pauses in speech leads to a pure attenuation of the signal sounds quieter but natural. Ideally, none arise "musical tones". A preferred embodiment of the Er This takes advantage of the invention by median filtering in the case of language activity on the range of amounts and during language breaks is carried out on the transmission values. The required The linguistic pause decision is the spectral one Subtraction is available anyway since the formation of the Ge Noise estimate only carried out during the pauses in speech becomes.

Instead of median filtering in the time direction as described A median filtering in the frequency direction can also be used  according to equation (6). The given detailed explanations apply to the filtering in Frequency direction analog. It turns out that you lose weight the number of samples within a time segment Median filtering in the frequency direction gains advantages versus filtering in the time direction and vice versa.

When using median filtering in temporal direction with the exemplified who The window length is the sampling rate and window length as shown in the example, equal to the minimum median star length 3. Larger window lengths lead in this case to further suppress the "musical tones", u. However, it may also lead to a level that is perceived as unnatural voice signal. The preferred window length is therefore 3 as indicated by way of example. For shorter ones Segments can have a larger window length at the median fil be appropriate. The one from the window of temporal Median filtering covered time interval should however be 50 ms do not exceed.

For filtering in the frequency direction, the Window length of the median filter at the data segment length. The data segment length should be numerically described Example is less than 64, the median filter is 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 on the magnitude spectrum of the disturbed Input signal is applied. 3. The method according to any one of claims 1 and 2, characterized characterized that the median filtering on the amount spectrum of the output signal of the spectral subtraction is applied.   4. Procedure for noise reduction of a disturbed Speech signal using spectral subtraction, where from a predefinable transmission characteristic Transfer function for the spectral subtraction be is true, characterized in that the transmission median filtering function. 5. Method for noise reduction of a disturbed Speech signal with a combination of previous claims che. 6. The method according to claim 5, characterized in that the median filtering in speech pauses on the transmission values and in the case of voice activity on the range of amounts of the Speech signal is applied. 7. The method according to any one of claims 1 to 6, characterized characterized that the median filtering in temporal Direction is applied. 8. The method according to claim 7, characterized in that the window length of the median filter three consecutive included time segments. 9. The method according to claim 7 or claim 8, characterized ge indicates that the window length of the median filter is small is less than 50 ms. 10. The method according to any one of claims 1 to 6, characterized characterized that the median filtering in frequency tion is applied.   11. The method according to claim 10, characterized in that the window length of the median filter does not exceed 5 Frequency values included.