Patent Application
20090257609
This application claims priority of German application No. 10 2008 031150.2 filed Jul. 1, 2008, which is incorporated by reference herein in its entirety.
The invention relates to a method for noise reduction for a hearing device and to a hearing device with noise reduction.
Hearing devices are wearable hearing apparatus used to provide assistance those with impaired hearing. To meet the numerous individual requirements different designs of hearing device are provided, such as behind-the-ear hearing devices, with an external earpiece and in-the-ear hearing devices e.g. also Concha or in-canal hearing devices. The typical configurations of hearing device are worn on the outer ear or in the auditory canal. Above and beyond these designs however there are also bone conduction hearing aids, implantable or vibro-tactile hearing aids available on the market. In such hearing aids the damaged hearing is simulated either mechanically or electrically.
Hearing devices principally have as their main components an input converter, an amplifier and an output converter. The input converter is as a rule a sound receiver, e.g. a microphone, and/or an electromagnetic receiver, e.g. an induction coil. The output converter is mostly implemented as an electro acoustic converter, e.g. a miniature loudspeaker or as an electromechanical converter, e.g. bone conduction earpiece. The amplifier is usually integrated into a signal processing unit. This basic structure is shown in
In the processing of digital speech recording, e.g. digital hearing devices, it is often desirable to suppress disruptive background noise without influencing the useful signal (speech). There are known filter methods suitable for this purpose which influence the short-term spectrum of the signal, such as the Wiener filters. However these methods require a precise estimation of the frequency-dependent power of the noise to be suppressed from an input signal. If this estimation is imprecise, either an unsatisfactory noise suppression is achieved, the desired signal is affected or additional artificially-created noise signals, so called “musical tones” occur. There are no methods for noise estimation yet available which solve these problems completely and efficiently.
Previously noise power has been able to be estimated principally using two approaches. Both methods can be undertaken either over a wide bandwidth or preferably in a frequency range split up by means of a filter bank or short-term Fourier transformation:
1. Speech Activity Detection:
Provided no speech activity is detected, the complete (time-variable) input signal power is regarded as noise. If speech activity is detected, the noise estimation is kept constant at the last value before the onset of the speech activity.
2. Noise Power Estimation During Speech Activity (the so Called “Minimum Tracking Method”):
It is known that during speech activity the speech signal power in individual frequency ranges is repeatedly briefly almost zero. If there is now an underlying mixture of speech and noise changing comparatively slowly over time, the minima of the spectral signal power considered over time correspond to the noise power at these times. The noise signal power must lie between the established minima (minimum tracking). Such a minimum tracking can for example be performed with the aid of a smoothing filter, which is described for example in R. Martin, “Noise power spectral density estimation based on optimal smoothing and minimum statistics”, IEEE Trans. Speech Audio Processing, Vol. 5, July 2001, Pages 504-512. The noise power is typically determined separately for different frequency ranges in the input signal. To this end the input signal is first split up by means of a filter bank or a Fourier transformation into individual frequency components. These components are then processed separately from one another.
In the above method 1, on the one hand the reliable detection of speech activity represents a problem, and on the other hand it is not possible to track noise which varies over time during simultaneous speech activity.
In the above method 2 there are fundamental contradictions in the setting of the algorithm to be resolved: If speech is present the noise estimation should only be adapted slowly in order not to classify speech components as noise through fast adaptation and affect the speech quality in this way. If there is no speech present, the noise power estimation should follow the temporal fine structure of the input signal without any delay. This produces conflicting demands for the setting parameters of the method, such as smoothing time constants, window length for a minimum search or weighting factors, which to date have only been able to be resolved averagely optimally. In addition this method is not in a position to track rapid changes in the noise signal.
A further option for enhancing speech and for suppression of “Musical Tones” is promised by “Cepstral smoothing” the weighting of spectral filters. C. Breithaupt et al., “Cepstral Smoothing of Spectral Filter Gains for Speech Enhancement Without Musical Noise”, IEEE Signal Processing Letters, Vol. 14, No. 12, December 2007, pages 1036 through 1039 describes that a recursive, temporary smoothing is essentially applied to higher cepstral coefficients, with each coefficient representing sound level information being removed. This method is also effective with non-stationary noise.
The object of the present invention is now to specify a further method and a hearing device for an enhanced noise reduction, with speech in particular being less adversely affected and disruptive artifacts being avoided more effectively.
In accordance with the invention the given object is achieved with the method and with the hearing device of the independent claims.
Inventively the method for noise reduction of an input signal comprises a modification of the coefficients of the cepstrum of the input signal, of the changed input signal and/or of at least one parameter derived from the input signal, with cepstral coefficients replacement signal or of a parameter derived from the replacement signal being accepted depending on a specific point in time (this correspondence to an acceptance varying from point in time to point in time), as well as a use of the modified cepstral coefficients for forming an output signal from the input signal with the noise in the output signal being reduced in relation to the input signal.
In a further development the input signal can be obtained from an acoustic signal picked up by a hearing device.
In a further embodiment the method can comprise the following steps:
Furthermore the method can comprise the following steps:
In a further development the method can comprise the following additional steps:
The advantage of processing in the cepstral domain lies in the fact that coefficients can be determined robustly, which are predominantly dominated by speech. This allows the other coefficients to be assigned to the noise/interference. Speech can be broken down in the cepstral domain into the transmission function of the vocal tract and the excitation function. The information about the transmission function of the vocal tract is mapped onto the lower cepstral coefficient. With voiced sounds the information about the excitation function will essentially be reflected in a cepstral maximum in the upper cepstral range. The knowledge of the cepstral coefficients which are dominated by speech can be used as a-priori knowledge for a robust noise reduction or for reconstruction of a naturally sounding residual noise. In particular for the case of instationary noises an enhanced estimation and thus an enhanced auditive quality is possible.
Inventively a hearing device with noise reduction according to an inventive method is also specified. It comprises a signal processing unit with a noise power estimator, a speech power estimator and a first and/or second replacement unit for modification of cepstral coefficients.
The invention also claims a computer program product with a computer program featuring software means for executing the inventive method when the computer program is executed in an inventive hearing device.
Further special features and advantages of the invention are evident from the subsequent explanations of a number of exemplary embodiments which refer to schematic drawings.
The drawings show:
A general overview of the inventive method for noise reduction is first given below, before specific embodiments are examined with reference to
The cepstrum of an input speech signal s(t) overlaid with noises can be determined as follows. Assuming that a discrete time signal s(t) sampled with the sampling rate fs is given. This time signal is subdivided into segments of length M. The segments are offset from each other with an advance of R and are weighted with an analysis window. The discrete Fourier-transform of the segment, Sk(1), is indexed by the frequency index k and the segment index 1. The cepstrum is calculated from the inverse Fourier transformation of the logarithmized magnitude spectrum
s
q(1)=IDFT{log(|Sk(1)|)},
with q being the cepstral coefficient index, the so-called Quefrency index, and IDFT { } being the inverse discrete Fourier transformation.
Cepstral coefficient zero (q=0) gives information about the even proportion of the logarithmized magnitude spectrum. The lower cepstral coefficients contain the information about the envelope of the speech signal, and thus also about the formants important for the comprehensibility. Formants are identified a maxima of the spectral envelopes which result from the resonances of the vocal tract. With voiced sounds maxima at multiples of the basic voice frequency are to be found in the spectrum. These maxima are essentially mapped in the cepstrum onto one strong maximum. Thereafter the maxima contain the lower cepstral coefficients a maximum in the upper cepstral domain the information about speech, while the remaining cepstral coefficients very probably do not to originate from speech.
Some of the output signals of spectral noise reduction algorithms contain unnatural artifacts, for example peaks in the spectral domain which lead to so-called “Musical Noise”. These local spectral maxima change the fine structure of the spectrum, which is reflected in the upper cepstral bins. Since it is known in the cepstral domain which coefficients very probably do not originate from speech, this information can be used to avoid spectral outliers in the output signal. To this end the cepstral coefficients of certain parameters of the noise reduction algorithm are modified. The modification can be undertaken for example by a replacement of the cepstral coefficients which very probably do not originate from speech by the corresponding coefficients of the noise-affected signal.
The following three parameters are suitable for a cepstral modification:
The flowchart of the inventive method for noise reduction shown in
From the spectral coefficients RLS, SLS thus obtained, by means of inverse Fourier transformation SCR, SCS of the logarithmized magnitude spectrum, the cepstra of the estimated noise power and speech power are formed. In this way the cepstral coefficients RLC, SLC are determined. From the spectrum of the input signal LS the cepstrum with the cepstral coefficients LSC is determined.
All three cepstra RLC, SLC, LSC are evaluated within the framework of a first replacement strategy ES1 and used for a modification of the cepstral coefficients RLC, SLC of the noise power or the speech power such that an optimum possible noise reduction of the input signal S and high naturalness of the output signal SR or aSR can be achieved. As the result of the first replacement strategy ES1 the modified cepstral coefficients mRLS, mSLS of the noise power and the speech are determined.
Modified spectral coefficients mRLS, mSLS of the noise power or the speech power are subsequently generated from the modified cepstral coefficients mRLC, mSLC by back transformations CSR, CSS. By means of a weighting method the weighting factors GF for the weighting of the spectral coefficients LS of the input signal are determined from the modified spectra mRLS, mSLS of the noise power and the speech power taking into account the spectrum LS of the input signal. With a subsequent multiplication MP the spectrum LS of the input signal is multiplied by the weighting factors. The modified spectral coefficients mLS thus produced are subsequently transformed by an inverse discrete Fourier transformation into a noise-reduced output signal SR.
Shown in
Before a back transformation in the time domain however the cepstrum with the cepstral coefficients ALS is formed from the noise-reduced spectrum mLS by means of inverse Fourier-Transformation SCA of the logarithmized magnitude spectrum. With the aid of a second replacement strategy ES2, which is intended to suppress artifacts, as well as taking into account the cepstrum LSC of the input signal S modified cepstral coefficients mALC of the noise-reduced output signal mLS are generated. Through a spectrum formation CSA modified spectral coefficients mALS are determined from them, which are subsequently transformed by an inverse discrete Fourier-transformation IDFT into an artifact-reduced output signal aSR.
The method steps shown can be implemented in accordance with the invention in a digital signal processor of a hearing device. In this way a high naturalness of an amplified sound signal with simultaneous noise reduction can be achieved. The cepstral modification transmits the fine structures present in the original noise-affected signal into the enhanced output signal and/or into the estimation of the speech power and/or into the estimation of the noise power, so that an enhanced naturalness is achieved and/or non-stationary noises are better mapped. The option of estimating rapidly changing noise makes this method extraordinarily interesting. Previously known methods merely achieve a reduction of the spectral fluctuations, but simultaneously reduce the fine timing structure.
| Number | Date | Country | Kind |
|---|---|---|---|
| 10 2008 031 150.2 | Jan 2008 | DE | national |
