CZ2012831A3 - Method of suppressng noise and accentuation of speech signal for cellular phone with two or more microphones - Google Patents

Abstract

The principle of noise suppression and speech enhancement in a mobile phone with two or more microphones is that a speech suppression filter bank (20) is created for the given mobile phone during manufacture or calibration. Each bank filter is designed for a particular speaker position relative to the phone so that its output signal has a maximum noise-to-speech ratio. The speaker's position on the phone for which the bank's filters are derived is selected to cover the most likely speaker positions. To suppress noise and enhance speech during call, microphone signals are parallel filtered by all filters from the bank. At that moment, its noise reference signal is selected by the output signal from that filter, whose output variance is minimal. The speech signal is highlighted by the focuser (40) and the noise is suppressed by the adaptive filter. For a given speaker position with respect to the phone, the speech-suppressing filter is derived from a silent speaker record. The filter is designed so that the variance of its output, if input is a record, is minimal.

Description

A method of suppressing noise and enhancing speech for a mobile phone with two or more microphones

Technical field

The invention relates to a method of suppressing noise and interfering sounds (hereinafter referred to as "noise") and enhancing a speech signal in a mobile phone with two or more microphones. This method of noise reduction is based on the use of a pre-measured speech-suppression filter system.

BACKGROUND OF THE INVENTION

The standard method of eliminating unwanted noise using microphone fields is to focus, in English literature referred to as beamforming, to search for a linear combination of microphone field output that maximizes the ratio of useful signal energy and noise energy. This focusing is effective when the geometry of the microphone field - signal source - noise source system does not change over time or does not change rapidly over time, and at the same time if the interference sources are rather point-like. (An alternative to point source noise is diffuse noise, which seems to come from all directions simultaneously.)

An example of unwanted noise suppression by focusing is the method described in Flaks et al., US2012245933, January 2010, which includes a number of techniques and options. A certain similarity to the solution proposed herein is in claim 17, where the possibility of using predetermined focusing parameters for a number of relative positions of the useful signal source (speaker mouth) and the microphone field is mentioned.

Another standard method of suppressing unwanted noise is to use a Wiener filter in the time-frequency domain. This method is suitable for diffuse type noise removal. It needs to estimate the instantaneous frequency spectrum of unwanted noise. This method exists in many different variants, for example in S.F. Pain, Suppression of Acoustic Noise in Speech Using Spectral Subtraction, IEEE Trans. Acoustics, Speech and Signal Processing, vol. 113-120, 1979. In a variant called double spectral subtraction, it appears in the patent by H. Gustafsson, I Claesson and S. Nordholm under US6549 ^ 86, April 1999. Another variant of spectral subtraction is the Power Level Differences, designed by M. Jeub et al., " Noise Reduction for Dual Microphone Mobile Phones Exploiting Power Level Differences, Proc. IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 1693-1696, Kyoto, Japan, March 2012.

In the latter two works it is assumed that the acoustic signal is sensed by two microphones, one microphone oriented towards the mouth of the speaker and the other directed towards the space. The PLD method is based on the assumption that the first microphone captures mainly the speech signal and the second microphone captures unwanted noise.

1 '

A limitation of this method is that the speaker's speech signal is partially present on the distant microphone, and although it is typically 10 dB less, it negatively affects the accuracy of the noise spectrum estimation and hence the efficiency of the separation of the useful signal.

A similar concept of a mobile, dual-microphone phone solution appears in the more recent patent by H.J.Belt and US2007 / 0230712A1 of 11.8.2005. Here, it is recognized that the efficiency of spectral readings (Wiener filter) depends on the position of the mobile phone relative to the mouth of the speaker, and therefore it is proposed to add a position detector to the device.

There are other methods of noise suppression and speech enhancement in a dual microphone system in the patent and other technical literature, but the above patents and work are conceptually closest to the solution proposed here. Among those other works, for example, Hao Deng, et al., US 2007/0165879 of January 13, 2007 (adaptive filtering method) or Kai Li et al., Two-Microphone Noise Reduction Using Spatial Information-Based Spectral Amplitude Estimation , IEICE Trans, on Information and Systems, vol. 1454-1464, 2012.

It is an object of the present invention to provide a method of suppressing unwanted noise, which is particularly suitable when the noise is diffuse in nature or is a point source but rapidly changing its position, does not need a mobile phone position detector and has better quality than known solutions.

SUMMARY OF THE INVENTION

The principle of the method of suppressing the noise and enhancing the speech signal in a mobile phone with two or more microphones is that for a given mobile phone (hereinafter referred to as the "telephone") a bank of 10 to 256 speech filter filters of 100 to 1000 length referred to as "the bank"). The input of each filter from the bank is signals from microphones and the output is one signal. Each bank filter is designed for a specific or group of speaker-to-telephone positions so that its output signal contains only noise and suppresses the speaker's voice as much as possible. The speaker-to-telephone positions for which the filters in the bank are derived are selected to cover together the closest neighborhood of the telephone and the most likely speaker-to-telephone positions in normal use. The position and shape of the speaker's head, the position and shape of the hand held by the speaker, and the typical noise in the environment where the telephone is used may also be taken into account.

A bank of speech-suppressing filters is created by the learning process after the phone prototype has been constructed. Furthermore, the bank can be customized by the user prior to normal use to calibrate the method to the user's voice, head shape and manner of holding the phone, and possibly to the most common environmental noise. For a given position or group of speaker positions relative to the telephone, the position and shape of his or her head and the position and shape of his hand (hereinafter referred to as "situation"), the filter is derived by recording or using multiple microphone recordings in the situation. The records shall be such that they contain only a negligible amount of noise. The filter is designed so that the variance of its output, if its input is these records, is minimal and, at the same time, so that the transmitted noise has the least altered spectrum.

2,

To suppress the noise and enhance the speech signal during a call, the microphone signals are filtered in parallel by all bank filters and the signal variation at their output is measured. At the moment, the output signal from the filter whose output variance is minimal is selected as the noise reference signal. The noise suppression and enhancement of the speech signal in the microphone signal or a signal that is a mixture of microphone signals (the "noise signal") is performed by subtracting noise from the focusing output signal by an adaptive Wiener filter. Another variant of the subtraction is the double spectral subtraction proposed in the aforementioned U.S. Pat. A third subtraction method is the PLD algorithm proposed by Jeub et al., ICASSP 20, also cited above. Finally, for all of the aforementioned spectral subtraction methods, the listening quality can be improved by smoothing (averaging) the spectrogram at higher frequencies, as described in P. Echt and P. Vary, Effective musical noise suppression for speech enhancement systems, Proceedings of IEEE Int. Conference on Acoustics, Speech and Signal Processing, ICASSP, Taipei, Taiwan, 2009 4409-4412.

Clarification of drawings /

Giant. 1 is a block diagram of a method of suppressing noise and enhancing a speech signal in a mobile phone that may be used in any embodiment of the invention.

Giant. 2 is a representation of typical positions from which pure speaker recordings are taken to calculate and derive a bank of speech-suppressing filters.

EXAMPLES OF EMBODIMENTS OF THE INVENTION

Example 1: Creating a Speech-Suppressing Filter Bank

A bank of speech-suppressing filters is created by the learning process after the phone prototype has been constructed. First, the phone records the speaker from different positions in relation to the phone. The number of recordings is 50, the length of each record is 5 seconds, and each record contains the spoken words of the person to be recorded "one, two, three, four. Speaker positions relative to the phone are selected to cover together the closest neighborhood of the phone use (see Fig. 2). Recordings are made in a quiet environment where the noise level is below 40 dB.

In addition, a 5-second noise recording is performed, typical of the phone's environment.

For each speaker recording made, a speech-suppressing filter is calculated, which for the fifth recording is determined by vectors g _{P / L and} g _{P /} R which contain filter coefficients and each have a length of 300 according to the formula [gp.L 8 _P , r1 ^ ^{ar] min} {£ | {GT * ^X p, ⁿ H) + gR * ^{X p, _R ⁿ M) | ² +

Sl »8r π e | _L {g *} y _L (n) + g _{{R} R} * y (n) -y _L (nd) | ² }, where x _{P / L} (n) and χ _ρΚ (η) denote samples of the _pth record of pure speech, y _L (n) and y _R (n) denote samples above

3 '' * í? ! The noise recordings, * denotes a convolution operation, ε is a regularization constant equal to 0.1 and D is an integer noise delay constant equal to 20. The purpose of the second term in the formula is not to change the speech suppression filter too much the spectrum of noise it transmits. The task of minimization is transformed into a system of linear equations with a block toeplitz matrix, which is quickly and economically solved by the block Levinson-Durbin algorithm, derived by H. Akaike, Block Toeplitz Matrix Inversion, SIAM J. Appl. Math. 24 (2): 234-241, 1973.

Calculated speech-suppressing filters, more precisely the FFT transformation of zero-extended filters to block lengths, form a filter bank that is stored in the phone memory.

Example 2: A simpler variant of creating a bank of speech-suppressing filters

The procedure is the same as in Example 1 except that the noise recording y _L (n) and Υκ (η) is not needed and the speech-suppression filter is calculated for the given speaker recording using the formula gp, L = argmin X | {g _L * x _p , _L } (n) - x _{p R} (n - D) | ² . In this case, the coefficients g _{P (R} all Si n are zero except D-th, which is equal to -1 and g _{) R} is therefore the same for all speech-suppressing filters (for all p). The task of minimization is transformed into a system of linear equations and solved by the Levinson-Durbin algorithm.

This simpler variant of calculation of speech-suppressing filters is less computationally demanding and has less memory requirements (it is sufficient to store coefficients _{gP; L} in memory). However, it does not allow adaptation to the type of noise removed and the speaker signal suppression is weaker.

Example 3: Performing a method of suppressing noise and enhancing a speech signal

Fig. 1 is a block diagram of a method of suppressing noise and enhancing a speech signal in a mobile phone that may be used in any embodiment of the invention.

The signals from the microphones sampled at 16 kHz x _L (n) and x _R (n), where n is the sample index, are first transformed by a fast window Discrete Fourier Transform (window FFT) 10 where window length is 1024 samples and window overlap is 50% . The block (window) of the transformed signals is denoted X _L (k) and X _R (k), where k is the frequency band index.

X _L (k) and X _R (k) are input to the bank of 20 speech-suppressing filters, where they are parallel filtered by all filters from the bank. The filter coefficients are read from the phone memory. Ptého filter output is computed according to the formula Z _p (k) = _pL G (k) x _L (k) + G _{p, r} (k) -X _r (k) where G _{p, L} (k) and G _p , _R (k) are the FFT transform coefficients of the p-th filter.

The filter selector 30 evaluates the output variations of the speech suppressing filters. The variation of the p-th filter is calculated according to the formula I T | from _p (k) | ^Z. The output of the filter selector 30 labeled Z (k) is the output of that k filter, the variance of which is the least. If it is the p-th filter, then Z (k) = Z _p (k).

In parallel to the calculation of the signal Z (k), the focuser 40, which is inputted by the signals X _L (k) and X _R (k), calculates the signal X (k), which is the input to the noise subtractor 50. The object is to increase the speech-to-noise ratio by means of a focusing device 40 which can be selected according to the manner in which the microphones are deployed on the telephone. If the microphones are positioned both in front, one of the known focusers 40 may be used, such as a delay-and-sum beamformer or set X (k) equal to the signal from the microphone whose variance is higher. If the microphones are placed one at the front (signal X _L (k)) and the other at the rear (signal X _R (k)), then X (k) is equal to X _L (k).

The noise reader 50 reads the Z (k) signal from the X (k) signal and outputs the S (k) signal. The reading method used is an adaptive Wiener filter, realized by the formula, | X (k) | ²

S (k) = ---- 5 - (----— X (k), where τ is an optional parameter equal to 10 to control the rate | x (k) | + r | z (k) | noise suppression To maintain the listening quality of the signal, the formula is used only for index values k from 0 to K, where K corresponds to a frequency of 3 kHz, for k> K then S (k) = X (k).

The inverse FFT transformation 60 and the OLA method convert the S (k) signal into the time domain s (k) using the inverse FFT and the overlap-add (OLA) method, which is described, for example, in B. Porat, "A Course in Digital Signal Processing" , John Wiley & Sons, Inc., 1997.

Industrial applicability

The invention is designed to be implemented in mobile phones having two or more microphones to capture sound. The invention is intended to facilitate telephony in noisy environments by suppressing environmental noise and amplifying the caller's speech signal in a transmitted telephone call.

List of reference marks

10-window FFT

- Bank of speech-suppressing filters

- filter selector

- focuser

- Noise reader

- inverse FFT transformation and OLA method

Claims (3)

PATENT CLAIMS A method for suppressing noise and enhancing a speech signal for a mobile phone with two or more microphones, characterized in that the noise component is estimated by means of a speech suppression filter bank (20) and a filter selector (30), wherein the filter selector (30) always selects as the noise reference signal an output signal from the filter whose output variance is minimal, and the actual speech signal estimation is made by the noise guide (50) by subtracting the estimated noise component from the focusing signal (40) by one of the known spectral subtraction methods such as Wiener filter, or derived methods such as Power level difference. The method of claim 1, wherein the speech suppression filter bank (20) is formed for an existing mobile phone prototype based on a set of speaker recordings holding the phone at various positions as expected in normal use of the phone, minimizing expression. g _pL = argmin ^ | {g _L * x _pL } (n) - x _pR (nD) | ² where x _p , _L (n) and x _{P / R} (n) denote samples of the p-th Sl n recording of pure speech from both microphones, n is the time index, * is the convolution operator and D is the delay parameter. The method of claim 1, wherein the speech suppression filter bank (20) is formed for an existing mobile phone prototype based on a set of speaker recordings holding the phone prototype at various positions as expected in normal use of the phone and audio recording. noise (noise) of the environment, which is expected to be typical for the use of the phone, by minimizing the expression [gp.L gp, R1 = argmin X {| {g _L * x _{p> L} } (n) + {g _R * x _{p R} } (n) | ² + Bl »8r not | {g _L * y _L } (n) + {g _R * y _R } (n) -y _L (nD) | ² }, where χ _ρΛ (η) and x _p , _R (n) denote samples of the _pth th record of pure speech from both microphones, y _L (n) and y _R (n) denote samples of the aforementioned noise record, n is a time index, * is the convolution operator, D is the delay parameter and ε is the regularization constant.