DE69423703T2 - Noise reduction device for pre-processing and / or post-processing of speech signals - Google Patents

Description

The invention relates to a noise suppression device for use in suppressing or removing a noise signal from a speech signal.

Typically, a speech signal is subjected to preprocessing before the speech signal is encoded into a sequence of coded signals. For example, such preprocessing has been performed to detect either a speech duration or a non-speech duration, according to a paper published by J. F. Lynch, Jr. et al. in IEEE entitled "SPEECH/SILENCE SEGMENTATION FOR REAL-TIME CODING VIA RULE BASED ADAPTIVE ENDPOINT DETECTION" (Proceedings ICASSP, pages 1348 to 1351, 1987). The paper only describes a distinction between the speech duration and the non-speech duration, but not the removal of a noise signal from the speech signal during preprocessing. That is, Lynch et al. never considered preprocessing that removes the noise signal from the speech signal. Even if the preprocessing described in the article is practically used to remove the noise signal from the speech signal, it is difficult to remove the noise signal, namely a non-speech signal in the speech duration.

On the other hand, in JP-A-2-278 298, spectral subtraction has been proposed to remove a noise component from the speech signal. Thereafter, the speech signal is encoded into a sequence of coded signals. In this method, only a noise spectrum resulting from the noise component is subtracted or removed from a spectrum having the noise spectrum and is obtained as noise-subtracted speech. signal. The noise-subtracted speech signal could be free from the noise component in the spectrum.

Note, however, that speech coding is usually performed not only in conjunction with the spectrum but also with a phase component of the speech signal. This shows that a noise component contained in the phase component cannot be removed by the above-mentioned method.

Therefore, spectral subtraction is disadvantageous in that the noise component cannot be completely removed from the speech signal.

In addition, spectral subtraction cannot be applied to post-processing that is performed after the encoded signal sequence has been decoded into a sequence of decoded signals.

In any case, the suppression of a noise component during post-processing has not been considered at all, although this noise reduction is necessary after decoding.

EP-A-549 364 discloses a noise signal prediction system according to the preamble of claim 1.

It is an object of the invention to provide a noise suppression device capable of completely removing a noise component or a noise signal from a speech signal.

It is a further object of the invention to provide a noise suppression device of the type described which can be used either in pre-processing or in post-processing of the speech signal.

It is a further object of the invention to provide a noise processor of the type described which can suppress the noise signal not only in a speaking period but also in a non-speaking period.

These tasks are solved with the features of the claims.

Fig. 1 is a block diagram of a noise suppressor according to a first embodiment of the invention;

Fig. 2 is a block diagram for describing a part of the noise suppression device shown in Fig. 1;

Fig. 3 is a block diagram of a noise suppressor according to a second embodiment of the invention; and

Fig. 4 is a block diagram for describing a part of the noise suppression device shown in Fig. 3.

A principle of the invention will first be described below in order to facilitate understanding of the invention. Here, it is assumed that a speech signal is given in the form of a sequence of digital speech signals which are to be subjected to preprocessing and postprocessing in order to remove a noise signal from the speech signal. Moreover, the preprocessing is performed in response to an input signal determined by the digital speech signal sequence which is not yet encoded, while the postprocessing is performed in response to an input signal determined by the digital speech signal sequence which is already decoded. Therefore, it should be noted that the terms "digital speech signal sequence" and "input signal" can be used here in two different meanings to include both the preprocessing and the postprocessing.

In any case, the input signal comprises the speech signal (namely the digital speech signal sequence) and the noise signal and can therefore be considered as a combination of the digital speech signal sequence and the noise signal.

According to the invention, feature parameters are extracted from the input signal and may be, for example, a selected or selected spectral parameter representing features of a spectrum in the input signal, fundamental frequency prediction gain factors representing a periodicity of the input signal, etc. The feature parameters are used to determine either a speech duration or a non-speech duration. duration by comparing the feature parameters with a threshold value.

In summary, a preliminary sound source signal that identifies a sound source is determined using the input signal and the feature parameters in preprocessing and postprocessing. In particular, the preliminary sound source signal appears in the form of an error signal that is generated in preprocessing by passing the input signal through an inverse filter controlled by the feature parameters.

On the other hand, the preliminary sound source signal appears in the form of a decoder output signal or a sequence of decoded signals decoded using the feature parameters.

Since the speech signal has an amplitude larger than the noise signal in the preliminary sound source signal, the noise signal can be suppressed only by comparing an amplitude of the preliminary sound source signal with a predetermined threshold and then generating a noise-suppressed signal. The noise-suppressed signal is converted into a noise-free output signal in preprocessing using the feature parameters and output as a noise-free decoded signal in postprocessing. The noise-free output signal can be encoded by an encoder after preprocessing, while the noise-free decoded signal is converted into a sound signal after postprocessing.

Noise suppression can be performed only in a selected one, namely the speaking period or the non-speaking period, or in both the speaking period and the non-speaking period. The invention can therefore suppress the noise signal on a waveform using the feature parameters and is applicable to both pre-processing and post-processing.

Referring to Fig. 1, a noise suppression device according to a first embodiment of the invention is applicable to preprocessing and is therefore supplied via an input terminal 10 with an input signal IN which is a speech signal and a noise signal superimposed on the speech signal. As mentioned above, the speech signal is given in the form of a sequence of digital speech signals. The input signal IN is passed to a frame division circuit 11 and divided by the frame division circuit 11 into a plurality of frames, each of which has a length of, for example, 40 ms. Each frame is further divided by a sub-frame division circuit 12 into a plurality of sub-frames, each of which has a length of, for example, 8 ms. The input signal IN is divided into the sub-frames as mentioned above and passed in the form of a divided input signal sequence x(n) either in each frame or in each sub-frame to a feature parameter calculator 15 on the one hand and to a noise suppression circuit 20 on the other hand. The divided input signal sequence x(n) can be referred to as an internal input signal.

In the example shown, the feature parameter calculator 15 is supplied with the internal input signal x(n) in each frame. The feature parameter calculator 15 first sets a window to extract a piece of the internal input signal x(n) related to each subframe. The window is longer than each subframe length and may be, for example, 24 ms.

Thereafter, the feature parameter calculation means 15 calculates, as feature parameters, spectral parameters indicating features of a spectrum in the input signal, fundamental frequency prediction gain factors indicating a periodicity of the speech signal, and an average amplitude in each subframe. In this case, an average power may be calculated in the feature parameter calculation means 15. Such calculations of the feature parameters are known and will not be described further. In any case, feature parameters are output from the feature parameter calculation means 15 as feature parameter signals.

Note that the feature parameter calculation means 15 shown in Fig. 1 calculates the spectral parameters of a predetermined order, which may be, for example, a tenth order. In addition, the following calculation is carried out: description assuming that linear prediction coefficients ai are used as spectral parameters. Although such linear prediction coefficients are calculated using a known LPC analysis, Burg analysis or the like, in connection with the example shown, it is assumed that the Burg analysis is used to calculate the linear prediction coefficients. The Burg analysis is described in detail in a book (pages 82 to 87) written by Nakamizo et al. and entitled "Signal Analysis and System Identification", published by Corona Company Ltd, Tokyo 1988. Therefore, the description as far as the Burg analysis is concerned is omitted in the present application.

Alternatively, the linear prediction coefficients can also be calculated using a covariance method and a correlation method.

As mentioned above, the fundamental frequency prediction gains are also calculated in the feature parameter calculator 15. The fundamental frequency prediction gains are represented by Pg and given by:

where: T is a delay time representing one fundamental frequency period; n is a sampling number; N is a maximum sampling number.

Instead of equation (1), the fundamental frequency prediction gain factors Pg can be easily calculated using the following equation:

The mean amplitude is represented by R and given by:

Here, circuits for calculating the above-mentioned linear prediction coefficients, the fundamental frequency prediction gains Pg and the average amplitude R can be easily implemented by a combination of conventional circuit elements. Therefore, no specific circuits for calculating the linear prediction coefficients, the fundamental frequency prediction gains Pg and the average amplitude are described below.

The feature parameter calculator 15 supplies a speech detection circuit 25 and the noise suppression circuit 20 with the feature parameter signals representing the feature parameters as mentioned above. In the example shown, the speech detection circuit 25 detects or determines either the speech duration or the non-speech duration of a speech signal in response to at least one of the feature parameters. For this purpose, a wide variety of methods can be used to determine the speech duration or the non-speech duration. For example, the illustrated speech detection circuit 25 first smoothes the fundamental frequency prediction gains Pg and the average amplitude R to obtain smoothed fundamental frequency prediction gains Pg' and a smoothed average amplitude R', and then compares the smoothed fundamental frequency prediction gains Pg' and the smoothed average amplitude R' with the first and second thresholds TH1 and TH2, respectively.

The above-mentioned smoothing process of the fundamental frequency prediction gains Pg and the average amplitude R are carried out according to the following equation:

P'j = (1- δ) P'j-1 + δ·P, (4)

where: P represents the fundamental frequency prediction gains or the mean amplitude to be smoothed; δ represents the time constant for smoothing and takes a value between 0 and 1 inclusive;

and P% and P'j-1 represent smoothed values at times j and j-1.

If, as a result of comparison, the smoothed fundamental frequency prediction gains Pg' and the smoothed average amplitude R' are smaller than the first and second threshold values TH1 and TH2, respectively, the speech detection circuit 25 decides that the non-speech duration in the internal input signal x(n) continues. Otherwise, the speech detection circuit 25 decides that the speech duration in the internal input signal x(n) continues. Thus, the non-speech duration and the speech duration are determined by the speech detection circuit 25. In the example, the first and second threshold values TH1 and TH2 may be invariable or time-variable.

As mentioned above, the speech detection circuit 25 comprises a calculation circuit for calculating the smoothed values (namely, the smoothed fundamental frequency prediction gains Pg' and the smoothed average amplitude R') according to the equation 4 and a comparator unit for comparing the smoothed values with the first and second threshold values TH1 and TH2, respectively. As a result, the illustrated speech detection circuit 25 can generate the smoothed average amplitude R' in each frame or in each subframe and a detection signal DT representing either the speech or non-speech duration in each frame and in each subframe.

The smoothed average amplitude R' is supplied to a storage circuit 30, while the detection signal DT is supplied to the noise suppression circuit 20.

According to Fig. 2 in addition to Fig. 1, the noise suppression circuit 20 can be operated to suppress the noise signal in at least one of the speaking and/or non-speaking periods. In Fig. 2, the noise suppression circuit 20 comprises an inverse filter 201 which is supplied with the internal input signal x(n) from the input terminal 10 via the frame and sub-frame division circuits 11 and 12. The feature parameters ai are also supplied from the feature parameter calculation means 15 to the inverse filter 201. The inverse filter 201 performs an inverse filter process to generate an inverse filtered signal e(n), which can be called a preliminary sound source signal because the inverse filtered signal e(n) determines a sound source. Here, the inverse filtered signal e(n) is given by:

e(n) = x(n) - ai x(n - i), (5)

where P represents an order of the inverse filter 201. The inverse filtered signal e(n) is therefore dependent on the feature parameters and determines the sound source.

The inverse filtered signal e(n) has a speech signal component and a noise signal component superimposed on the speech signal component, and appears in the form of a continuous signal. The inverse filter 201 may be simply referred to as a filter circuit.

Now note that the inverse filtered signal e(n) is determined by a pulse with a comparatively large amplitude in a portion of the speech signal component that occurs in the speaking period, since the speech signal has a fundamental frequency. On the other hand, the inverse filtered signal e(n) shows a comparatively small amplitude in a portion of the noise signal.

Consequently, the noise signal can be suppressed by comparing the inverse filtered signal e(n) with the threshold TH₁.

In particular, the noise reduction circuit 20 shown in Fig. 2 comprises a threshold calculation circuit 202 which is supplied with the smoothed average amplitude R' calculated by the feature parameter calculation means 15 according to the equation 4 and stored in the storage circuit 30. The threshold calculation circuit 202 calculates the threshold TH1 which is given by:

TH₁ = K2·R' (6)

to generate a threshold signal representing the threshold TH1, where K2 is greater than zero. The threshold TH1 is thus determined by the average amplitude R stored in the memory circuit 30.

The inverse filtered signal e(n) and the threshold signal are passed to a suppression unit 203, to which the detection signal DT from the speech detection circuit 25 is also passed. The suppression circuit 203 is set to an active state or an inactive state in response to the detection signal DT. In this case, the suppression unit 203 can suppress the noise signal in at least one of the speaking period and the non-speaking period. In the example shown, it is assumed that the suppression unit 203 is set to the active state in response to the detection signal DT in the non-speaking period, although the suppression unit 203 can be set to the active state in the speaking period.

Furthermore, the suppression unit 203 compares the inverse filtered signal e(n) with the threshold signal. The suppression unit 203 attenuates the inverse filtered signal e(n) by a predetermined amount or sets the inverse filtered signal e(n) to zero if the inverse filtered signal e(n) is smaller than the threshold TH₁. As a result, the suppression unit 203 generates a noise-suppressed signal e' which is determined by:

where K is greater than zero and less than one.

In any case, a combination of the threshold calculation circuit 202 and the suppression unit 203 serves to suppress the noise signal contained in the inverse filtered signal e(n) and to generate the noise-suppressed signal e'(n), and can be collectively referred to as a noise suppression part.

The noise-reduced signal e'(n) is passed to a playback circuit 204 together with the feature parameters ai. The playback circuit 204 converts the noise-reduced signal e'(n) into a noise-reduced speech signal x'(n) taking into account the feature parameters a;. In this case, the noise-reduced speech signal x' is given by:

x'(n) = e'(n) + ai x'(n - i). (8th)

The noise-reduced speech signal x'(n) is supplied to an encoder (not shown) via an output terminal 35 of the noise reduction circuit 20 to be encoded. Thus, the noise-reduced speech signal x'(n) is generated during processing before encoding. Since the noise reduction is performed taking into account the feature parameters of the input signal IN, a phase component of the noise signal can be suppressed even in the above-mentioned example.

Referring to Fig. 3, a noise suppressor (denoted 40) according to a second embodiment of the invention can be operated to perform post-processing after coding. For this purpose, the illustrated noise processor 40 is connected to a decoder 45 which is supplied with a decoder input signal or an input signal DIN with feature parameters of a speech signal and an index signal relating to the sound source. The encoder 45 itself can be similar to the known decoder and generates a sequence of decoded sound source signals v(n) representing a sound source, together with feature parameters and the index signal, as known. The decoded sound source signal sequence v(n) and the feature parameters and the index signal are passed to the noise suppressor 40.

In the noise suppressor 40, the decoded sound source signal sequence v(n) is passed to a noise suppression circuit, designated 50, which may be operated as described in detail later. Furthermore, the illustrated noise suppressor 40 comprises a speech detection circuit 25' and a memory circuit 30', each of which may be similar to those shown in Fig. 1. From this fact, it is readily understood that the speech detection circuit is responsive to the feature parameters, e.g., the spectral parameters, the fundamental frequency prediction gains Pg and the average amplitude R, to either determine the speech duration or the non-speech duration. The speech detection circuit 25' supplies the noise suppression circuit 50 with a detection signal DT' indicative of either the speech duration or the non-speech duration. As in Fig. 1, the speech detection circuit 25' calculates the smoothed average amplitude R' which is stored in the memory circuit 30'.

According to Fig. 4 together with Fig. 3, the noise suppression circuit 50 comprises a threshold calculator 501 which is supplied with the smoothed average amplitude R' to calculate the threshold signal representing a threshold TH2 as in the threshold calculation circuit 202. The threshold signal is passed to the suppression unit 502 together with the detection signal DT'.

The suppression unit 502 is set to an active state in at least one of the speaking and/or non-speaking periods. Here, it is assumed that the illustrated suppression unit 502 only becomes active in the non-speaking period, as in the case of the suppression unit 203. In any case, the suppression unit 502 generates a sequence of noise-suppressed sound source signals v'(n) given by:

where K is identical to K in equation 7. The threshold TH2 can be equal to that in equation 7.

Referring again to Fig. 3, noise-reduced sound source signals v'(n) are supplied to a speech reproduction circuit 52 which is supplied with the feature parameters from the decoder 45. The speech reproduction circuit 52 converts the noise-reduced sound signals into a converted speech signal in a known manner, taking into account the feature parameters. The converted speech signal is supplied to a loudspeaker or the like.

The noise suppression device according to the invention can be used in post-processing the decoded sound source signals DIN in the manner described above.

Although the invention has been described in connection with a few embodiments, it will be readily possible for those skilled in the art to put it into practice in various other ways. For example, the feature parameters need not always be limited to linear prediction coefficients, but may be any other known parameters. In addition, any parameters other than the average amplitude and the fundamental frequency prediction gains may be used. The speech detection circuit 25 or 25' may be operated in a manner different from that shown in Figs. 1 and 3.

In addition, post-processing may be performed to suppress the noise signal even if the feature parameters are not transmitted by a transmitter and are not received by a decoder 45 (Fig. 3). In this case, the speech signal is once reconverted by a receiver to form a reconverted speech waveform and then to calculate feature parameters from the reconverted speech waveform in a manner described in connection with Fig. 1. The feature parameters may be used to suppress the noise signal in the manner described above.

With this structure, noise suppression is possible both during preprocessing and postprocessing of the speech signal. In addition, not only the noise signal that occurs in the non-speech period can be suppressed, but also a non-speech signal that is superimposed on the speech signal that occurs in the speech period. Such suppression can be performed on the waveform.

Claims (7)

1. Noise suppression device supplied with an internal input signal IN comprising both a speech signal and a noise signal to produce an output signal substantially free of the noise signal, the speech signal being determined from a sound source, the noise suppression device comprising: feature parameter calculation means (15) supplied with the internal input signal for calculating a feature parameter determining a feature of the speech signal to produce a feature parameter signal representing the feature parameter, and noise suppression means (20) connected to the feature parameter calculation means (15) for removing the noise signal from the internal input signal to produce the output signal, wherein the noise suppression device comprises: a suppression unit (203) for removing the noise signal from a residual signal (e(n)) by estimating the noise signal to generate a noise-suppressed signal (e'(n)); and an output device (204) for generating the noise suppressed signal as the output signal; and characterized by a filter device (201) supplied with the feature parameter signal (ai) and the internal input signal for filtering the internal input signal (x (n)) to produce a filtered signal which is dependent on the feature parameter (ai) and which calculates the sound source in the residual signal (e(n)) which represents the difference between the feature parameter signal representation and the internal input signal wherein the suppression unit is coupled to the filter device (201). 2. Noise suppression device according to claim 1, wherein the speech signal is divisible into a speaking period and a non-speaking period, the noise suppression device (20) further comprising: a speech detection device (25) coupled to the feature parameter calculation device (15) for calculating the speaking and non-speaking durations in response to the feature parameter signal to produce a detection signal representing either the speaking or non-speaking duration; an average calculation device (30) coupled to the speech detection device for calculating an average of either the power or an amplitude in the non-speech period to produce an average signal representing the average; wherein the noise suppression device (20) further comprises: a threshold value calculator (202) for calculating the threshold value from the mean value signal to supply the suppression unit (203) with a threshold value signal (TH₁) representing the threshold value to cause the suppression unit (203) to compare the filtered signal with the threshold value signal and the suppression unit to suppress the noise signal. 3. Noise suppression device according to claim 2, wherein the suppression unit (203) is further supplied with the detection signal (DT) in order to be put into an active state in at least one of the speaking and/or non-speaking periods. 4. A noise suppression device according to claim 1, 2 or 3, wherein the feature parameter calculation means (15) calculates as the feature parameter (ai) spectral parameters representing a spectrum of the internal input signal, a fundamental frequency period of the internal input signal and an average amplitude of the internal input signal. 5. Noise suppression device according to claim 1, 2, 3 or 4, wherein the internal input signal is divided into a sequence of frames each lasting for a predetermined time interval, the speech signal is generated by a sound source and has a spectrum determined by at least one feature parameter and is divisible into a speaking duration and a non-speaking duration, the suppression device comprises: feature parameter calculation means for calculating the at least one feature parameter to generate a feature parameter signal representative of the at least one feature parameter, and speech detection means coupled to the feature parameter calculation means (15) for determining the speaking and non-speaking durations in response to the feature parameter signal to generate a detection signal representative of either the speaking or non-speaking duration, an average storage device is coupled to the speech detection device for storing an average of either the power or an amplitude of the internal input signal in the non-speech period to generate an average signal corresponding to the average; and the noise suppression device (20) is coupled to the feature parameter calculation device (15), the speech detection device and the mean value calculation device for suppressing the noise signal taking into account the feature parameter signal, the detection signal, the mean value signal and the internal input signal in order to generate the output signal. 6. Noise suppression device operable in response to a feature parameter signal defining a speech signal and to a sound source signal (v(n)) representing a sound source of the speech signal, to remove a noise signal from the sound source signal and to produce an output signal (v'(n)) substantially free of the noise signal, the speech signal being divisible into a speech duration and a non-speech duration, the sound source signal appearing in the form of an error signal generated during preprocessing by an input signal is passed through an inverse filter controlled by the feature parameter signal, the noise suppression device being characterized by: a noise suppression circuit (50) for removing the noise signal from the sound source signal taking into account the feature parameter signal to generate a noise-suppressed signal (v'(n)); means (52) for generating the noise suppressed signal as the output signal. 7. Noise suppression device according to claim 6, characterized by: a speech detection device (25') supplied with the feature parameter signals for determining the speech and non-speech durations to generate a detection signal representing either the speech or non-speech duration; and an average storage device (30') coupled to the speech detection device for storing an average of either the power or an amplitude of the speech signal in the non-speech period to generate an average signal representing the average; wherein the noise suppression circuit (50) suppresses the noise signal also taking into account the mean value signal.