Method for optimising a multi-stage filter bank and corresponding filter bank and hearing aid

p0001The present invention relates to a method for optimizing a structure of a multi-stage filter bank having an input side and an output-side filter bank based on a given filter type, the input-side filter bank has a variable first parameter, the input channel number and a variable second parameter an oversampling. Moreover, the present invention relates to a multi-stage filter bank whose structure is obtained with the aid of an above method. The present invention further relates to a hearing apparatus having such a multi-stage filter bank. The term "hearing apparatus" here is any portable in or on the ear, sound-emitting device, in particular a hearing aid, a headset, headphones and the like, understood.

p0002Hearing devices are portable hearing devices used to support the hard of hearing. To meet the numerous individual requirements, different types of hearing devices such as behind-the-ear (BTE) hearing aid with an external receiver (RIC: receiver in the canal) are and In-the-ear (ITE) hearing, as well as concha hearing aids or canal hearing aids (ITE, CIC). The hearing by way of example are worn on the outer ear or in the ear canal. In addition, be on the market are bone, implantable or vibrotactile hearing aids available. The stimulation of the damaged hearing is either mechanically or electrically.

p0003Hearing devices principally as essential components an input transducer, an amplifier and an output transducer. The input transducer is generally a sound receiver, eg. As a microphone, and / or an electromagnetic receiver, eg. As an inductor. The output transducer is usually implemented as an electroacoustic transducer, eg. As a miniature loudspeaker, or as an electromechanical transducer, eg., Bone conduction hearing aid. The amplifier is usually integrated into a signal processing unit. This basic structure is in<figref idrefs="f0001">1 shows</figref> illustrated by the example of a behind-the-ear hearing aid. In a hearing aid housing 1 to be worn behind the ear, one or more microphones 2 for recording the ambient sound are built. A signal processing unit 3, which is also integrated in the hearing aid housing 1, processes the microphone signals and amplifies them. The output of the signal processing unit 3 is transmitted to a loudspeaker or receiver 4, which outputs an acoustic signal. The sound is optionally transmitted through a sound tube which is fixed with an earpiece in the ear canal to the eardrum of the wearer. The power supply of the hearing device and especially to the signal processing unit 3 by a likewise integrated into the hearing aid housing 1 battery. 5

p0004Sound signals, which are received from one or more microphones of a hearing aid are usually disassembled for further processing in sub-band signals. For this purpose, usually, use is made of one or more frequency-selective digital analysis filter banks (AFB), whereby K> 1 subband signals are obtained. After the decomposition of sub-band specific signal manipulations can be performed. A subsequent re-synthesis of the manipulated sub-band signals can be carried out by means of a digital synthesis filter bank (SFB).

p0005On quality filter banks in hearing certain requirements are imposed. Thus, a channel bandwidth of about 250 Hz is required for example in the lowermost bands. Otherwise the bandgap should align roughly to the Bark scale. Furthermore, a number of channels of at least 22 is desirable. Spurious aliasing should certainly be below 60dB. Due to the intensive sub-band processing (especially the high gain required to compensate for the hearing loss) hearing aids are conventional methods for eliminating aliasing not effective. The filter banks are therefore in principle "not critical" scan. Furthermore, should the group delay (both for AFB and SFB) are well below 5 ms and the group delay distortions do not exceed a certain extent. Especially for high frequencies is the group delay to be kept as low as possible, which is a major limiting factor for the filter bank.

p0006To date, filter banks already are multistage tree structure and uneven division (Bark scale), but used without optimized stop band rejection and thus lack or inadequacy of downsampling (solution A). Moreover, it is also known, a single-stage (eg. Complex modulated) filter bank to use (solution B). Finally, multi-stage filter banks are used (solution C) for the separation of acoustic signals. Such a filter bank, for example, equipped with K channels with at least two different bandwidths, the bandwidth based on the constraints of the Bark scale, but have identical channel groups bandwidth. This results in advantages over the solutions A and B.

p0007Furthermore, due to the multi-stage filter bank, an increased number of sub-bands possible in order to achieve improved resolution compared to the solution A can. At the same time, however, no obligation exists for the realization of an excessive number of sub-bands, which is considered an improvement over the solution B.

p0008Furthermore, with the multi-stage filter bank a reduction of the sampling rate of the sub-band signals as far as possible, as the avoidance of interference permits (improvement over solution A). However, the reduction of the sampling rate can be carried out only to the extent that by manipulating the sub-band signals are no interference by spectral About folds (aliasing). This leads to an improvement over prior art filter bank process for aliasing compensation as they are described for example in<nplcit id="ncit0001" npl-type="b"><text>Heinz Göckler and Alexandra Groth: "multirate systems" Schlembach Verlag, Wilburgstetten 2004</text></nplcit>,

p0009From the article by <nplcit id="ncit0002" npl-type="s"><text>Brennan, R .; Schneider, T .: A flexible filterbank structure for extensive signal manipulations in digital hearing aids; In. IEEE, Proceedings of the International Symposium on Circuits and Systems May 31-June 3 1998, 1998. ISCAS '98. Volume 6, pp 569-572</text></nplcit> discloses a flexible filter bank structure for hearing aids. This filter bank structure is made up of an analysis filter bank and a synthesis filter bank, between which a channel-dependent processing is performed. It is an oversampled filter bank, and the aim is to achieve the shortest possible time delay and a high efficiency computer.

p0010In addition, the article<nplcit id="ncit0003" npl-type="s"><text> Yu Shao; Chip Hong Chang: A Generalized Time-Frequency Subtraction Method for Robust Speech Enhancement Based on Wavelet Filter Banks Modeling of Human Auditory System. In: IEEE TRANSACTIONS ON SYSTEMS, MAN AND CYBERNETICS. PART B: CYBERNETICS, Vol 37, No.. 4, August 2007, 2007, pp 877-889</text></nplcit> a filter bank modeling of a human ear. To improve speech intelligibility, the signal-to-noise ratio is improved.

p0011The object of the present invention is to be able to provide a multi-stage filter bank, which has the smallest possible group delay in the sub-bands and can be operated with as little energy expenditure.

p0012This object is achieved by a method for optimizing a structure of a multi-stage filter bank having an input side and an output-side filter bank based on a given filter type, the input-side filter bank has a variable first parameter, the number of channels and as a variable second parameter an oversampling, comprising the steps of: determining a respective group delay for several pairs of values ​​of the first and second parameters, determining in each case an operation rate for several pairs of values ​​of the first and second parameters, selecting one of the plurality of pairs of values, in which the associated group delay and the associated operation rate a predetermined criterion meet, and configure the input-side filter bank with the channel number and the oversampling factor corresponding to the selected pair of values.

p0013Advantageously, it is possible to use a filter bank system that has numerous parameters such Nutzkanalzahl the input-side filter bank, oversampling of the input-side filter bank, the number of output-side filter banks, Nutzkanalzahl the output-side filter banks, oversampling of the output-side filter banks, tree "full tree" or "reduced tree ", etc., to optimize successfully use very few, specifically selected parameters. So it's not a complicated optimization process in a multidimensional space required.

p0014If the multi-stage filter bank, ie the cascaded filter bank system, the entire bank width is verschiedenstufig can, each in addition to the plurality of pairs of values, a data rate that represents a sampling rate cumulated over all sub-bands of the multi-stage filter bank is determined, and this data rate for selecting one of the several pairs of values ​​are used. This results in an additional evaluation criterion, with the result of space can be accurately limited.

p0015In a further development, instead of commonly used "perfect reconstruction" filter banks (PR filter bank) an "approximate perfectly reconstructed" filter bank (NPR filter bank) can be used. An NPR filter bank is usually sufficient for an intensive sub-band processing. With an NPR filter bank can be further expenses and group delay savings.

p0016Further, it may be at any part-filter bank of the filter bank system at an oversampled, complex modulated regular (DFT) or frequency-shifted (GDFT) polyphase filter bank. Such filter banks have extensive variation possibilities, but are easy to optimize the process of the invention.

p0017The predefined criterion for optimization may be that the group delay and the operation rate is as low as possible. If both parameters are not at one and the same pair of values ​​a minimum, so a suitable compromise for the parameter values ​​and the corresponding pair of values ​​is to be selected (input channel number and oversampling).

p0018Advantageously, a multi-stage filter bank can now be found with the inventive method, the structure of which is optimized for group delay and power consumption. In particular, such a multi-stage filter bank for input signal processing in hearing devices, especially in hearing aids is suitable.

p0019The present invention will now be explained in more detail with reference to the accompanying drawings, in which:<dl id="dl0001"><dt>1 shows</dt><dd>the basic design of a hearing device according to the prior art;</dd><dt>FIG 2</dt><dd>a tree structure of a to be optimized based filter bank concept;</dd><dt>FIG 3</dt><dd>a diagram of the group delay of the frequency group of low frequencies as a function of K_On and U_On; </dd><dt>FIG 4</dt><dd>a diagram of the operation rate of multiplication of the entire filter bank as a function of K_On and U_On;</dd><dt>FIG 5</dt><dd>a diagram of the data rate of all sub-bands (total number of samples per unit time) depending on K_On and U_On; and</dd><dt>FIG 6</dt><dd>a tree structure of an optimized filter bank concept.</dd></dl>

p0020The more detail below embodiments represent preferred embodiments of the present invention. For the optimization method according to the invention of a filter bank approach is in the following example of a tree structure is considered as in <figref idrefs="f0001">FIG 2</figref> is shown. Such a base filter bank concept includes an input-side filter bank 10 and a plurality of output-side filter banks 11₁ to 11_m and 12₁ until 12_n, The input-side filter bank 10 has a number of channels K_On = N + m. The sampling rate at the input of the filter bank 10 is f_{A, A},

p0021The basic filter bank concept is designed in two stages only in this example. It has in addition to the input-side filter bank 10 in the first stage only a second stage with the filter banks 11₁ to 11_m and 12₁ until 12_nRepresenting each conventional complex modulated filter banks. The filter banks 11₁ to 11_m have Nutzkanalzahl K_{Off, 1}While the filter banks 12₁ until 12_n the channel number K_{Off, 2} have. The actual number of channels corresponding to the product U_On · K_{Off 1.2}Where U_On the oversampling of the input-side filter bank represented.

p0022The output-side filter banks sit in the example chosen so together only two different filter bank types. It is in a "full tree" of the filter bank type with the channel number K_{Off, 1} S₁fold and the one with the channel number K_{Off, 2} S₂fold represented. The sampling rate at the output of the input-side filter bank 10 is f_{A, Zw} and sampling at the outputs of the output side filter banks 11₁ to 11_m and 12₁ until 12_n be f_{A, TB1} or f_{A, TB2},

p0023If only a "reduced tree" in front, so the output-side channels are partially formed by the input-side filter bank 10th Accordingly, for example, on the output side filter banks 12₁ until 12_n (in <figref idrefs="f0001">FIG. 2</figref> dashed lines) can be dispensed.

p0024The aim is now, the structure of the filter bank system of <figref idrefs="f0001">FIG 2</figref> in particular to optimize in terms of an application in hearing aids. It is the smallest possible group delay in each channel and a total of the lowest possible energy consumption, ie as little filtering operations can be achieved. It has been shown, that can be reduced to two main parameters to optimize the entire parameter space of the filter bank system, namely, the input channel number K_On and the oversampling factor U_On the input-side filter bank 10th

p0025First, the group delay is a critical band of low frequencies, depending on the two parameters K_On and U_On examined. For the study, as well as for all subsequent investigations filter according to the Remez-standard design (linear phase FIR filters) are used. The three-dimensional graphic of<figref idrefs="f0002">FIG 3</figref> are the group delay τ_G on the parameters U_On and K_On again. There is a significant increase in group delay for the frequency group of low frequencies in the right corner, ie at high input channel count K_On and low oversampling factors U_On, Differential Group Delay times are so rather the front and rear portions of the graphic, very low running times in the left corner of the graph, ie at high oversampling factors U_On and low input channel numbers K_On, For higher frequencies, the group delay varies depending on U_On and K_On only marginally.

p0026High oversampling factors U_On but lead to a high number of necessary operations, whereby the power consumption increases. Consequently, a compromise between group delay and number of operations or operation rate to find. Therefore, in accordance with<figref idrefs="f0002">FIG 4</figref> the operation rate f_{OP, mult} (For multiplication) the entire filter bank as a function of K_On and U_On determined. The multiplication operation rate can be considered representative of the entire operation rate of the filter bank here. In the three-dimensional graphics<figref idrefs="f0002">FIG 4</figref> shows that the operation rate f_{OP, mult} In the left corner, ie at a high oversampling factor U_On and a small number of channels K_On, Is relatively large. In the front and rear corner of the graph, the operation rates are significantly lower and in the right corner at least, that is at a high number of channels K_On and low oversampling factor U_On, As part of the optimization, however, both optimization criteria, namely the group delay and the operation rate must be considered simultaneously. It turns out - as mentioned above - found that the group delay of the frequency group of high frequencies is hardly critical and need not therefore be used for the optimization. Accordingly, where the<figref idrefs="f0002">Figures 3 and 4</figref> considered together, then give optimum value pairs of the parameters U_On and K_On the front and rear region of the plane, these two parameters U_On and K_On in accordance with the <figref idrefs="f0002">Figures 3 and 4</figref> span.

p0027As an additional decision criterion for finding an optimum value pair U_On/ K_On in addition the so-called data rate "f_databe "used. It results from all sample cumulated over all subbands. As <figref idrefs="f0003">FIG 5</figref> However shows, it is (see, for a "full tree" of the filter system. <figref idrefs="f0001">FIG 2</figref>) constant. Only for a "reduced tree" there is a function of the parameters U_On and K_On, In the example shown was on the output side filter banks 12₁ until 12_n waived. This resulted in a dependence of the data rate f_data of the two parameters K_On and U_On, The three-dimensional graphics of<figref idrefs="f0003">FIG 5</figref> it would be in accordance with the best to use a reduced tree (high number of channels K_On) And the oversampling factor U_On to keep the input-side filter bank to a minimum. In this pair of values ​​range, however, the group delay is relatively high (same<figref idrefs="f0002">FIG 3</figref>).

p0028Overall, a compromise can be found, of all optimization criteria (low group delay, low operation rate and possibly low data rate) into account. In the present example, it was found that a filter bank with the tree under<figref idrefs="f0004">FIG 6</figref> can be regarded as optimal. 10 Thus has the input-side filter bank channel number K_On = 4 and an oversampling factor U_On = 2. The optimal structure leads to a full tree with four output-side filter banks 11₁, 12₁, 12₂ and 12₃, The output-side filter bank 11₁ for the low frequency range has the channel number K_Out = 12 and the oversampling factor U_Out = 5. For the higher frequencies, there are three output-side filter banks 12₁, 12₂ and 12₃ the same filter bank type. They each possess the channel number K_Out = 4 and an oversampling factor U_Out = 4. This opens up over the entire frequency range 24 channels, ie a channel number that is in accordance with the originally established claims at least 22nd The further requirement that the lowest bands a channel bandwidth is needed by about 250 Hz, with the structure of<figref idrefs="f0004">FIG 6</figref> be respected. In this optimized filter bank structure namely possess the lower channels has a bandwidth of 240 Hz and the upper channels have a bandwidth of 750 Hz with the optimization process of the invention it is thus possible, a multi-stage structure of oversampled, complex modulated. (Modified: GDFT) polyphase filter banks specifically in terms of applications in hearing devices, in particular hearing aids to optimize.

p0029In the case of establishment of frequency shifted complex modulated (GDFT) polyphase filter banks can present invention for each part-filter bank each is the cheapest solution for realizing the frequency shift are selected: In the first case, the input signal of each part-filter bank is itself frequency offset, wherein the prototype filter is still real coefficients having coefficient, but is fed with a complex signal (as realization DFT filter bank). The output signal has to be shifted in frequency again in the opposite direction. In the second case, the prototype filter is itself modulated complex, whereby said complex coefficient, but is supplied with a real signal. To preserve the symmetry of possibly present FIR filter coefficients, the zero phase of the frequency offset causing carrier wave is suitably selected.

p0030The choice of two options is as follows: If the input signal of each part-filter bank is real, then the second option is selected, it is complex, so the first. In any case arise so also GDFT-part filter banks in the realizable prototype filters only half complex multiplications. In a multistage filter bank may, where appropriate, the necessary in the first case and frequency shifts are grouped behind the DFT filter bank.

p0031A further improvement in terms of computational effort (energy consumption) and group delay can be achieved in that instead of conventional PR filter banks ( "perfect reconstructive filter banks") NPR filter banks ( "not perfect reconstructing filter banks") are used. The NPR filter banks can namely be designed so that they are sufficient for an intensive sub-band processing.

p0032As a composite of at least two cascaded stages filter bank or as a partially least two-stage (analysis) filter bank can to decompose the input signal into K sub-band signals are used with reduced sampling. A digital synthesis filter bank (SFB) for re-synthesis by manipulating the sub-band signals can be set symmetrically to the analysis filter bank (AFB) be constructed.

p0033Altogether with the optimized filter bank structure, a smaller group delay particularly in the broader band at higher center frequencies reach centric channels, which represents an improvement over the introduction mentioned solution B. A further advantage of the optimized filter bank system can be of reduced computational effort in the analysis-synthesis filter bank cascade mention what particular represents an improvement over a non-optimized solution C. Finally, resulting in an advantageous manner also a greatly reduced computational effort in the sub-band signal processing by downsampling of the sub-band signals and avoid excess channel number.

p0034The filter bank system described above can be used for example in a hearing aid or other hearing device.