EP1728322A1 - Circuit arrangement and signal processing device - Google Patents

Abstract

The circuit arrangement comprises a number of filter stages of a filter bank as well as a number of resonator circuits. The circuit arrangement also contains a resonator control circuit for controlling or regulating the quality of the resonator circuits. This resonator control circuit is configured so that it controls or regulates the quality of at least one resonator circuit according to the amplitude of the input signal and/or of the output signal of the resonator circuit.

Description

description

Circuit arrangement and signal processing device

The invention relates to a circuit arrangement and a signal processing device.

In the context of automatic speech recognition, it is known to apply a Fast Fourier Transformation (FFT) to a digitized input speech signal for spectral analysis of the input speech signal. Features used in the context of automatic speech recognition are derived from the performance spectrum formed by means of the Fast Fourier Transformation (cf. [1]).

With such a Fast Fourier Transformation, a time window of a predetermined length is usually used, a partial signal, which is represented by the respective time window, being subjected to a Fast Fourier transformation. This leads to a limited frequency resolution and time resolution.

If, as is customary in speech recognition, only the power spectrum and thus the magnitude spectrum of the respective partial signal is used, the time resolution is limited by the length of the time window used. This limitation of time resolution is a factor that limits the performance of previously known speech recognition systems. Problematic with sound processing systems and the use of such

The time window of a fixed, predetermined size is that when the power spectrum changes after a transformation back into a time representation, an error is formed which is based on the finiteness of the time window. According to alternative approaches to automatic speech recognition, filter banks are used as described in [2], [3] and [4]. These filter banks are designed to simulate the properties of a person's inner ear.

Furthermore, a model of inner hair cells (IHZ) and the auditory nerve is described in [5], the basilar membrane being simulated using different filters and a compression unit and a reinforcement unit. Furthermore, a model of a vesicle pool is described in [5].

[6] also describes a model of signal processing in the human ear, an acoustic sound signal being used as the input signal according to this model. The input signal is shifted into several frequency channels in the same way as the frequency-location transformation in the inner ear, with half-wave rectification and low-pass filtering and adaptation for amplifying sudden changes in the input signal and damping components of the input signal which are essentially constant over time, are provided for each channel.

[7] describes the phenomenon referred to as "recruitment" and the so-called dynamic compression to compensate for the recruitment phenomenon

Dynamic compression becomes a large range of sound levels, which occurs in the acoustic environment, "compressed" into a range that is perceptible to humans.

[8] describes a digital simulation of a one-dimensional long-wave model of the basilar membrane of a human using a wave digital filter structure. [9] describes a growth function in which the vibration of the basilar membrane is shown in relation to the sound pressure, measured in front of the eardrum of a test animal. The biological structure of the inner hair cells is also described in [10]. [11] describes amplifier circuits in which the amplifiers are used for

Improved intelligibility for a public address system. The voltage-controlled amplifiers described in [11] are controlled by an amplification control signal which is provided by means of a buffer amplifier and which is formed by a combination network from the partial signals of different frequency channels supplied by means of a filter bank. The voltage-controlled amplifiers described in [11] process four different signals in four different frequency channels. The aim of the circuit arrangement described there is that the output signal provided by the respective voltage-controlled amplifiers is brought to the level of the amplitude of the signal in the baseband channel.

The invention is based on the problem of specifying a circuit arrangement and a signal processing device for providing features for describing one of the circuit arrangement or the signal processing

Device supplied signal, wherein the features are more robust against noise.

The problem is solved by a circuit arrangement and by a signal processing device with the features according to the independent claims.

Preferred embodiments of the invention result from the dependent claims.

A circuit arrangement has a filter bank with a plurality of filter stages and a filter bank input, to which an input signal can be fed. Furthermore, the circuit arrangement has a plurality of resonator circuits for generating a respective partial output signal from the input signal, each resonator circuit being assigned to at least one filter stage of the plurality of filter stages and being coupled to an output of the respective filter stage. Each resonator circuit has a capacitance, an inductance and a resonator output, at which the respective partial output signal can be provided. In addition, at least one resonator control circuit for controlling or regulating the quality of at least one resonator circuit is provided in the circuit arrangement, the resonator control circuit being set up in such a way that it depends on the quality of the at least one resonator circuit controls or regulates the time course of the signal amplitude of the input signal and / or the partial output signal of the at least one resonator circuit.

A signal processing device has a circuit arrangement described above and a further processing unit for further processing the signal provided by the circuit arrangement.

The combination of a filter bank, preferably a linear filter bank with the resonator circuits, which clearly form non-linear compression stages, very well simulates the non-linear vibration behavior of the inner ear of mammals.

Clearly, as output signal of each resonator circuit, features are provided within the scope of a feature extraction of an input speech signal, which features are more robust against noise and in particular within the scope of a speech recognition system for an improved word error rate by means of a device according to the invention provided feature extraction system (formed by the circuit arrangement).

In particular, with increasing noise, the word error rate increases more slowly than in the context of classic preprocessing using a Fast Fourier Transformation, applied to time window split power spectra.

Another advantage which is achieved by the invention is to be seen in the extensive preservation of the fine time structure of the spoken voice signal, generally the analog input signal fed to the circuit arrangement, while in the power spectrum using a Fast Fourier Transformation only a time resolution of the Features in the range of the window length of the window used is achieved.

In this way, according to the invention, the biological structure and in particular the essential properties of the human hearing system are reproduced in an improved manner as compared to the prior art as part of the feature extraction of features in an acoustic signal. This leads to a more robust speech recognition system.

The invention can thus be clearly seen in the extraction of features for, for example, automatic speech recognition, that is to say for an automatic speech recognition system. In particular, features are provided from an input signal supplied to the circuit arrangement, which enable greater robustness against background noise than can be achieved according to the prior art.

Alternatively, the invention can be used in particular for a hearing aid, for example for a cochlear implant Patients with hearing loss due to inner ear can be used very advantageously.

According to the invention, the quality of the respective resonator circuit is set based on the amplitude of the input or partial output signal. Has one of these ^signals' a very high amplitude, so the quality can be of a resonator circuit can be so greatly reduced at least by means of the resonator type control circuit that the signal is highly attenuated. In contrast, in the case of a signal of a low amplitude, the quality can be increased in such a way that the amplitude of the signal is amplified at the output of the respective resonator circuit.

Clearly, according to the invention, the fact that a resonator circuit acts as a stable amplifier near its resonance frequency (resonance magnification) is used to carry out dynamic compression.

The quality of the resonator circuit used here means the ratio of the amplitude of the output signal at the resonance frequency of the resonator circuit to the corresponding amplitude of the input signal. The quality of a resonator circuit depends on its ohmic resistance, so that the quality can be determined, for example

Controlling or Regeins the ohmic resistance of the resonator circuit is adjustable.

In a scenario in which the quality of the resonator circuit based on the amplitude of the in the

Resonator circuit introduced input signal is set, the functionality of the control circuit can be referred to as a "control". If, on the other hand, the quality of the resonator circuit is set based on the amplitude of the output signal, the resonator control circuit fulfills a “regulation” functionality, since it carries out a feedback adjustment of the quality. The circuit arrangement according to the invention enables safe and effective dynamic compression of an input signal in the time domain without the disadvantages of a Fourier transformation occurring. In particular, the problems with a finite time window that occur in a Fourier transformation according to the prior art are eliminated. In addition, according to the invention, a dynamically compressed output signal is generated which, for example, has significantly less disruptive signal distortion compared to the inverse transformation of the logarithmic Fourier spectrum.

According to the invention, a sufficiently strong and intensity-selective (e.g. non-linear) attenuation of an input signal is made possible by selectively reducing the quality of the resonator circuit.

The circuit arrangement clearly has a filter circuit, based on the value of the inductance L and the capacitance C of the resonator circuit.

Circuit is the frequency range for which the resonator circuit is permeable. Thus, by dividing the values L, C, a simple possibility is created to set the frequency center of gravity of the transmissible interval of the resonator circuit. The width of the resonance curve of the resonator circuit can be adjusted in particular by dividing its quality. The resonator circuit can be regarded as a filter with nonlinear damping, with which in principle an arbitrarily high dynamic compression can be achieved. Due to a sufficiently narrow-band processing, distortions, which can arise due to excessive non-linearity, can also be kept sufficiently low.

The circuit arrangement can include a second-order resonator circuit, the attenuation increasing nonlinearly with increasing sound level. With a passive Realization of the circuit arrangement, that is to say when using passive components (coil L, capacitor C, ohmic resistor R), a stable circuit can be obtained (in contrast to systems which require an active, feedback amplifier).

The resonator circuits can have an ohmic resistance which can be controlled (or regulated) by means of the resonator control circuit. Such a controllable or regulable ohmic resistance is a simple circuit component, by means of which the functionality of regulating the quality of the resonator circuit can be fulfilled with little effort and accurately and stably.

It should be noted that the resonator control circuit can be formed from a plurality of partial resonator control circuits, with one partial resonator control circuit controlling the quality of a resonator circuit assigned to it.

The input signal can be provided between a first connection of the ohmic resistor and a first connection of the capacitance. The output signal can be provided between the first connection of the capacitance and a second connection of the capacitance. A second connection of the ohmic resistor can be coupled to a first connection of the inductance and a second connection of the inductance can be coupled to a second connection of the capacitance.

The resonator control circuit can be set up in such a way that it controls the quality of the at least one resonator circuit based on a Boltzmann function in which the amplitude of the output signal is contained as a parameter. With a suitable choice of the parameters contained therein, a Boltzmann function is well suited to approximate the sensitivity curve of the outer hair sensory cells in the human inner ear. A particularly good one This biological dependency can be described by a second-order Boltzmann function. This makes it possible to approximate the sensitivity curve in the human ear, which is advantageous for applications of the circuit arrangement in the medical field (for example for a hearing aid).

The resonator control circuit can be set up in such a way that it controls the quality of the at least one resonator circuit as a function of the amplitude of the

Output signals based on a sensitivity characteristic determined for a human ear. In order to emulate the sensitivity characteristic in the inner ear of a person particularly well by means of a circuit arrangement according to the invention, one can be determined, for example, experimentally or theoretically

Sensitivity characteristics of the human ear can be stored in the form of a file or table accessible to the control circuit. In this case, the resonator control circuit can control or regulate the quality of the at least one resonator circuit in such a way that the biological sensitivity characteristic stored therein is approximated.

The resonator control circuit can be set up in such a way that the lower the quality of the at least one resonator circuit, the higher the amplitude of the respective partial output signal of the respective resonator circuit.

The resonator control circuit can also be set up in such a way that it adjusts the quality of the at least one resonator circuit in a non-linear dependence on the amplitude of the respective partial output signal. This means that signal areas of large amplitude are disproportionately damped compared to signal areas of small amplitude. So even with one extreme high range of sound levels in an input signal, compression to a sufficiently narrow range can be achieved in the output signal.

The resonator control circuit can be set up in such a way that it adjusts the quality of the at least one resonator circuit in such a way that the amplitude of the respective partial output signal is within a predetermined interval. For certain applications it may be advantageous to keep the amplitude of a partial output signal within a predetermined interval in any case. This can be important in the context of data compression, for example, if a signal with a high intensity fluctuation is to be recorded with as few quantization levels as possible. In this case, the resonator control circuit can be set up in such a way that it controls or regulates the quality of the resonator circuit in such a way that the respective partial output signal lies within the predetermined interval.

The circuit arrangement can have a plurality of resonator circuits connected in series, an output signal of a resonator circuit connected upstream in each case being able to provide the resonator circuit connected downstream of it as an input signal.

According to this particularly advantageous embodiment, a filter bank with a series connection of several resonator circuits is clearly created, whereby the dynamic compression to an even greater

Dynamic range can be expanded. In principle, a sufficiently strong dynamic compression (eg 60 dB) can already be achieved with a filter stage (ie with a resonator circuit) with a very high quality Q (eg Q = 1000, which is reduced to a quality of Q = 1 at high levels) respectively. However, such a circuit arrangement is very narrow-band (for example 0.1% of the resonance frequency of the resonator Circuit). By cascading several filter stages

(e.g. three filter stages connected in series) with a 3 relatively low quality factor Q (e.g. Q = 10, so that Q = 1000), a sufficiently strong dynamic compression (e.g. of 60 dB) can also be achieved according to the invention. The not too high individual quality of each of these filters has the advantageous effect that, due to the increased bandwidth of the individual filters resulting from the lower quality, a larger frequency range of the filters is covered and at the same time the impulse behavior of the

Filter is improved, i.e. the system's settling and decay time is considerably shorter.

The resonator circuits connected in series can clearly be directly coupled to one another in such a way that the output voltage of an upstream resonator circuit is equal to the input voltage of the resonator circuit connected downstream of it and that the output current of an upstream resonator circuit (which is generally different from zero in operation) Circuit equal to that

Input current of the resonator circuit connected downstream. For this purpose, the circuit arrangement is generally free of an intermediate element between upstream and downstream resonator circuits. This can be realized by means of a circuit arrangement in which the second connection of the coil of an upstream resonator circuit is coupled to the first connection of the ohmic resistance of the resonator circuit connected downstream of the upstream resonator circuit.

Alternatively, the series-connected resonator circuits can clearly be free of direct coupling, i.e. be decoupled from each other in a certain way, in particular by interposing an intermediate element between the output of an upstream and

Input of a downstream resonator circuit. This is preferably implemented in such a way that the output voltage of an upstream resonator circuit is equal to the input voltage of the resonator circuit connected downstream of it and that the output current of an upstream resonator circuit is zero. The input current of the downstream resonator circuit essentially results only from the impedance of this resonator circuit. In such a circuit arrangement, an operational amplifier (as an impedance converter) is preferably provided as an intermediate element between an upstream resonator circuit and the resonator circuit downstream of it. A first input of the operational amplifier is coupled to the second connection of the coil of the upstream resonator circuit. A second input of the operational amplifier is connected to an output of the

Operational amplifier fed back and is coupled to the first connection of the ohmic resistance of the resonator circuit connected downstream of the upstream resonator circuit.

To reduce the computing power, the quality of all resonator circuits connected in series can be set identically. In this case, the computing power used by the resonator control circuit is kept particularly low, since a common quality is determined and set for all resonator circuits, i.e. all filter parameters are identical. If a circuit arrangement with a particularly high quality requirement is required, the quality of different resonator circuits connected in series can alternatively be set differently for the purpose of optimization. With such a circuit arrangement, the quality of each of the series-connected resonator circuits is thus set individually.

The circuit arrangement preferably has a plurality of branches connected in parallel, each branch has a resonator circuit or a plurality of resonator circuits connected in series. In this case, the quality of a respective resonator circuit can be controlled or regulated by means of the respective resonator control circuit.

According to this particularly advantageous development of the invention, a plurality of branches of resonator circuits connected in parallel are clearly provided, it being possible for a plurality of resonator circuits to be connected in series in each branch.

The at least one resonator circuit of a respective branch is preferably set up in such a way that it is permeable to a respective frequency range of the input signal in such a way that the branches are together permeable for a coherent frequency interval. The frequency range for which human hearing is sensitive is approximately between 20 Hz and 20 kHz. In order to cover this hearing frequency range, the parallel

Arrangement of resonator circuits in different channels, the frequency ranges of transmissible signals usually different. The frequency range of transmissible signals in a resonator circuit is a distribution curve around the resonance frequency with a certain half-width. The resonance frequency is clearly possible by dividing the values L, C of the resonator circuit, the half width is by. Adjustment of the respective quality adjustable. If the different frequency passbands of the different branches of resonator circuits are put together, a preferably coherent frequency interval results, by means of which the sensitivity range of the human ear or another frequency range of interest can be determined. The frequency ranges for which different branches are permeable are preferably at least partially overlapping one another. In this case, it is ensured that all frequencies are recorded and the signal components of individual branches can be combined.

The frequency range for which a respective branch is permeable can preferably be predetermined by setting the value of the capacitance and / or the inductance of the at least one resonator circuit of the branch. This is due to the fact that the resonance frequency of a resonator circuit depends on the values of the inductance and the capacitance.

The circuit arrangement of the invention is preferably set up to process an acoustic signal as an input signal. In this case, the circuit arrangement of the invention is suitable for use in a speech processing system. Such can be based, for example, on pulsating neural networks which rely on a reduction in the dynamic range. Other areas of application are systems for sound processing and (audio) data compression if signals with high amplitudes are to be recorded with as few quantization levels as possible. There are also applications in medical

Area, especially as a hearing aid for patients with noise and hearing loss.

The circuit arrangement according to the invention can be implemented in digital or analog circuit technology.

At least part of the circuit arrangement, in particular the filters, the control or regulation functionality of the resonator control circuit, can be implemented as a computer program. The invention can be implemented both by means of a computer program, ie software, and by means of one or more special electrical circuits, ie in hardware or in any hybrid form, ie using software components and hardware components.

Software implementation, in particular of the control circuit, can take place, for example, in "C ++". It can be implemented on any processor or DSP (digital signal processor), as well as on an FPGA module. An FPGA ("Field Programmable Gate Array") is an integrated programmable circuit which generally has a large number of programmable cells on a chip.

According to another embodiment of the invention, it is provided that the, preferably linear, filter bank is designed as a linear digital will filter.

Furthermore, a plurality of high-pass filters can be provided, with at least one high-pass filter being assigned to each filter stage, one high-pass filter in each case at the

Output of a respective resonator circuit is coupled. These high-pass filters simulate the fluid coupling of the hair bundles of the sensory cells in the inner ear to the vibration of the basilar membrane.

With at least one high-pass filter per filter stage, coupled to the output of the respective resonator circuit, the otherwise relatively flat, high-frequency filter edge of the filter bank is achieved. At least some of the high-pass filters, preferably all high-pass filters, are preferably designed as first-order high-pass filters. According to a development of the invention, the cutoff frequency of at least some of the first-order high-pass filters is selected such that it corresponds to the frequency of the maximum sensitivity of a basilar membrane vibration of an inner ear of a mammal. According to another embodiment of the invention, a plurality of rectifier circuits, wherein a rectifier circuit is assigned to one of the filter stages and a high-pass filter and is coupled to an output of a respective high-pass filter, and preferably a plurality of low-pass filters, each with a low-pass filter assigned to a rectifier circuit and coupled to an output of a respective rectifier circuit. According to these configurations, a very good approximation of the formation of the receptor potential U. in the human hearing system is achieved.

Furthermore, a plurality of activation circuits can be provided, an activation circuit being assigned to one of the filter stages, each

Activation circuit is set up to amplify a change over time in a signal supplied to the activation circuit and to attenuate components which are essentially constant over time in the signal supplied to the activation circuit.

Furthermore, each activation circuit preferably has a vesicle pool circuit with a multiplicity of vesicle circuits.

The signal processing device according to the invention, which has a circuit arrangement according to the invention, is described in more detail below. Refinements of the signal processing device also apply to the circuit arrangement and vice versa.

In the signal processing device, the further processing unit can be a speech recognition device or a hearing aid.

When the processing unit is implemented as a hearing aid, an application in which dynamic compression is performed to compensate for disturbances in the volume perception of the hearing impaired. In the disturbed hearing, the outer hair cells can be affected, which increases the sensitivity at low sound levels. The hearing then always works vividly with the sensitivity provided for high sound levels. This means that the usable range of sound levels between the hearing threshold (very quiet) and the inconvenience threshold (very loud) becomes smaller (recruitment). To compensate for this phenomenon, a dynamic compression can be carried out by means of the circuit arrangement of the signal processing device according to the invention, which clearly compresses the large sound level range of the acoustic environment onto the area to be perceived by the patient.

The signal processing device can also form the input for a speech recognition system, in particular in a pulsating neural network architecture.

The signal processing device can be set up as an analog or digital filter bank.

The following principles are clearly to be emphasized within the scope of the invention:

a) There is a frequency analysis of the input signal provided to the circuit arrangement with non-linear dynamic compression;

b) an information reduction is provided by “soft” one-way rectification of the frequency channels formed with threshold value and saturation;

c) there is an emphasis on language-relevant modulation frequencies by simulating the neural Adaptation achieved in the hearing system of a mammal, in particular a human.

Exemplary embodiments of the invention are shown in the figures and are explained in more detail below.

Show it

1 shows a circuit arrangement according to a preferred exemplary embodiment of the invention,

FIG. 2 shows a resonator circuit according to an exemplary embodiment of the invention,

FIG. 3 shows a realization of the resonator circuit shown in FIG. 2 as a wave digital filter,

FIGS. 4 and 5 are diagrams to illustrate the functionality of the circuit arrangement according to FIG. 1,

FIG. 6a shows a partial circuit arrangement according to another exemplary embodiment of the invention,

FIG. 6b shows a realization of the resonator circuits shown in FIG. 6a as a wave digital filter,

7a shows a partial circuit arrangement according to another exemplary embodiment of the invention,

FIG. 7b shows a realization of the resonator circuits shown in FIG. 7a as a wave digital filter,

FIG. 8 shows a block diagram of a speech recognition system according to an exemplary embodiment of the invention, FIG. 9 shows a circuit diagram of a linear filter bank and a plurality of individual filter stages of the resonator circuits assigned to the filter bank according to an exemplary embodiment of the invention,

FIG. 10 shows a diagram in which an excitation pattern of a nonlinear basilar membrane model for a 1 KHz tone is shown.

FIG. 11 shows a circuit diagram of a subcircuit, which is in each case connected in series with a respective resonator circuit for forming the circuit arrangement according to an exemplary embodiment of the invention,

FIG. 12 shows a representation of modeled nerve action potentials, and

FIG. 13 shows a diagram in which different speech recognition rates for a speech recognition system according to an exemplary embodiment of the invention are shown compared to a speech recognition system according to the prior art.

8 shows a speech recognition system 800 according to an exemplary embodiment of the invention.

As part of the speech recognition system, a

Feature extraction system 801 used, which in the context of the actual automatic speech recognition (in

8 represented by a speech recognition block 802) features extracted from a supplied analog speech signal.

The feature extraction system 801 has in particular

Components for pre-filtering, a filter bank and components for non-linear feature extraction. The feature extraction system 801 clearly reproduces the signal processing strategy and the signal processing structure of the human hearing system. It is scaled into physical units analogous to the signal processing system of the human hearing system.

The feature extraction system 801 is supplied with an input signal 803 (input signal) in analog form as a sound pressure signal (measured in pascals).

A first component 804 of the mermal extraction system 801 forms a model of the auditory canal, which is, however, optional and is neglected in the preferred implementation.

The signal 805 formed by the model of the auditory canal 804 from the input signal 803 is fed to a middle ear model component 806.

As shown in FIG. 9, the middle ear model component 806 has a parallel connection of an ideal spring component 901 and an ideal damping component 902 (realized in the form of an electrical coil, that is to say inductance, for the ideal spring or in the form of an ohmic resistor as a damping component ). The

Middle ear model component 906 is set up in such a way that the language-relevant region of the spectrum of the input signal is emphasized, that is to say is amplified. In analogy to the human hearing system, the signal 807 provided by the middle ear model component 806 is fed to an inner ear model component 808.

The signals 809 formed by the inner ear model component 808 can optionally be used directly for speech recognition as extracted speech recognition features or can be supplied to a sensor cell model component 810, which will be described in more detail below. The one from the Signals 811 generated by sensor cell model component 810 can likewise be used directly as feature components in the context of automatic speech recognition or can be further processed and can be fed as part of this further processing to a synaptic model component 812 which simulates the synaptic mechanism of the human hearing system. The signals 813 formed by the synaptic model component 812 are also used according to the invention as features in the context of automatic speech recognition.

In this context, it should be noted that, according to alternative embodiments of the invention, one, two all or three of the feature signals 809, 811, 813 described above can optionally be used in the context of speech recognition.

Within the framework of the human hearing system, the sound pressure impinging on the eardrum (the eardrum has a surface of -6 2 A _ec j = 55x10 m) is converted into a mechanical deflection of the

Middle ear bone (expressed in m) converted.

As described above, the middle ear model component 806 has a low-pass filter, in this exemplary embodiment a first-order low-pass filter, preferably with a low-pass filter corner frequency of 1 kHz.

The spring constant simulated by means of inductance 901 is approximately 1500 N / m and the damping component represented by means of ohmic resistor 902 is closed

Dimensioned 0.25 Ns / m.

The signal 807 generated by the middle ear model component 806 is coupled into the inner ear model component 808, which is shown in detail in FIG. 9. The

Inner ear model component 808 is as a filter bank 903 formed, according to this embodiment in the form of a linear wave digital filter model.

The filter bank 903 has a multiplicity of filter stages 904, 905, 906 and an ohmic terminating resistor 907.

Each filter stage 904, 905, 906 is formed by a series connection of an inductor 904a, 905a, 906a, an ohmic resistor 904b, 905b, 906b and a capacitance 904c, 905c, 906c.

The speed of the basilar membrane vibration in the inner ear of a person corresponds to the current in a filter stage 904, 905, 906. The deflection of the basilar membrane can therefore be calculated by integrating the speed. In order to avoid numerical problems with the integration, however, the deflection is sensibly calculated in a different way:

For a spring represented by capacitances 904c, 905c, 906c, the instantaneous deflection x can be calculated as the product of the spring force and the spring constant. This deflection x forms the entrance, i.e. the input signal for the dynamic compression stages 101, in each case one according to this exemplary embodiment of the invention

Series connection of two resonator circuits 101, as will be explained in more detail below.

The resonator circuits 101 clearly, as will be explained in more detail below, form compression stages

Use of second order resonators.

In an alternative preferred embodiment, four resonator circuits 101 are provided for each filter stage 904, 905, 906 in a series connection. As also explained in more detail below, the quality is of the resonator circuits 101, each instantaneously in a range of 1 to a maximum quality value Q _ax (depending on the position in the human inner ear) as a function ^'of the output signal of each filter stage 904, 905 , 906 modulated.

The structure of the resonator circuits 101 is explained in more detail below with reference to FIG.

Circuit arrangement 100 includes a variety of

Resonator circuits 101, each of which has a capacitance and an inductance (not shown in FIG. 1), as well as an input at which an input signal can be provided and an output at which an output signal can be provided. In each case three of the resonator circuits 101 are connected in series along a respective line of the matrix arrangement, so that a respective output of an upstream resonator circuit 101 is coupled to a respective input of a resonator circuit 101 connected downstream of it. The values of the

Inductance and the capacitance of the resonator circuits 101 of a row are each selected such that the respective row can transmit a signal of a corresponding frequency interval in a surrounding area of the resonance frequency of the resonator circuits 101 of the row. The

Resonator circuits 101 of different rows each have different values for L, C, so that, taken together, the individual rows or branches of resonator circuits 101 cover a coherent frequency interval which corresponds to the sensitivity range of the human ear (approximately 20 Hz to 20 kHz).

A resonator control circuit 111 is in communication with all resonator circuits 101, ie the control circuit 111 is coupled with all resonator circuits. The quality of each of the resonator circuits 101 is medium of the control circuit 111 for controlling or regulating the quality of the resonator circuits 101 can be set, the control circuit 111 being set up in such a way that the quality of the resonator circuits 101 is dependent on the amplitude of an output signal of the last resonator circuit 101 sets the respective line. For example, the quality of the resonator circuits R, R12 _. R-13 is set by means of the resonator control circuit 111 based on the amplitude of a signal at the output of the resonator circuit R.

A sound source 103 is also shown in FIG. 1, which emits an acoustic signal as a global input signal 102. This is provided to the inputs of the resonator circuits 101 (Rn. R-21 / • • • 1 ^R kl / • • ■ / / ^R nl) of the first column of resonator circuits 101.

The resonator circuit 101 Rn arranged in the first row and the first column of resonator circuits is considered below. The global input signal 102 of the sound source 103 is provided to this at an input. The resonator circuit 101 Rn passes a frequency component of the global input signal 102 which is dependent on the values L and C assigned to it and which is provided at an output of the resonator circuit Rn as the first local output signal 104. Furthermore, due to the functionality of the resonator circuit 101 Rn, the global input signal 102 is changed in amplitude depending on its (current) quality Q. The quality Q of the resonator circuit 101 Rn is regulated by means of an ohmic resistor (not shown in FIG. 1) of the resonator circuit 101 Rn, the control circuit 111 providing this controllable ohmic resistor with a corresponding control signal, whereby the resistance to a predetermined value is set. As a result, the quality of the resonator circuit 101 is set so that in one subsequent processing cycle according to this value of quality, an input signal is attenuated more or less. Since the circuit arrangement 100 is set up for dynamic compression of the global input signal 102, signal areas of high amplitude are clearly weakened more than signal areas of low amplitude.

The first local output signal 104 is provided as the first local input signal 105 to the resonator circuit 101 R12 connected downstream of the resonator circuit 101 Rn. The first local input signal 105 passes through the resonator circuit 101 Ri2, the second local output signal 106 being provided at an output. The second local output signal 106 serves as a second local input signal 107 of the resonator circuit 101 Rn connected downstream

Resonator circuit 101 R13. A third local output signal 108 is provided at its output 108-. This is together with the output signals of the last resonator circuits arranged in a row, each relating to a separate frequency interval

101 (R13, R-23 r ••• / ^R k3 / ••• / ^R n3) combined ^{to form} a global output signal 109 (added).

For each of the resonator circuits 101 of a respective row of resonator circuits (R] ci, 2 ^R k3) the

The quality of all resonator circuits 101 of the line is regulated by means of the resonator control circuit 111 based on the amplitude of the output signal at the output of the last resonator circuit (in the kth line resonator circuit R] 3).

The composite global output signal 109 is thus subjected to dynamic compression with respect to the global input signal 102. The resonator circuit 101 from FIG. 1 is described below with reference to FIG.

An input signal 200 is symbolized as voltage source U in FIG. Furthermore, an output signal 204 is symbolized as voltage U. The input signal 200 is provided between a first connection of an ohmic resistor 203 and a first connection of a capacitance 201. The output signal 204 is provided between the first connection of the capacitance 201 and a second connection of the capacitance 201. Furthermore, a second connection of the adjustable ohmic resistor 203 is coupled to a first connection of an inductor 202, and a second connection of the inductor 202 is coupled to the second connection of the capacitance 201.

The value of the ohmic resistor R 203 can be set by means of the control circuit 111. The resonator circuit 101 from FIG. 2 thus clearly represents a filter with controllable damping.

In the circuit arrangement 100 according to the invention, three (or generally N) resonator circuits 101 are connected in series as filter elements without feedback in each row. The time-dependent output signal Uc (t), where t is the time, of an upstream filter defines the input signal U 200 of the filter downstream of the upstream filter.

The resistor R 203 can be changed as a non-linear function of the output voltage U _c (t) (regulation), as a function of Uc (t) of the upstream filter (control), or for all filters simultaneously as a function of Uc (t ) the last filter stage in a row.

It is described below on the basis of which calculation rule according to the exemplary embodiment described the value R of a respective ohmic resistor R 203 is set.

For this purpose, a quality factor Q to be set is first calculated.

According to the exemplary embodiment described, the quality Q of the filter is damped according to a Boltzmann function:

Q (t) = (Qo - Q _mi „Xi - [i) l + exp {-SAT | U _c (t) |} -iD + Q. (

In equation (1) Q (t) is the dependence of the quality Q on time t. QQ is a predeterminable maximum quality of the resonator circuit 101 (for example QQ = 10). Q _m i _n is a predetermined minimum quality of the resonator circuit (for example, in Q = 1) • SAT is a predetermined saturation threshold, i.e., a

Parameter with which the time dependency of the quality can be set (e.g. SAT = 1).

The Boltzmann function (1) approximates the sensitivity curve of the outer hair cells in the inner ear. If required, the function can be replaced by a Boltzmann function of the second order, which allows an even more precise adaptation by introducing a further parameter. A simple first-order Boltzmann function is used in equation (1), since it has only one free parameter (namely SAT) and can therefore be processed with little numerical effort.

The value of the non-linear resistance R to be set is calculated from the quality Q of the filter: The time-dependent value of the ohmic resistance R (t) thus depends on the value of the inductance L and the capacitance C and the time-dependent quality factor Q (t).

Clearly form equations (1) and (2)

Regulation for setting the value R of the ohmic resistor 203 by means of the control circuit 111.

The filter formed by the resonator circuit 101 shown in FIG. 2 is linear at very low amplitudes Uc (t) (with Q → QQ for Uc (t) → 0). It is also approximately linear for very large amplitudes Uc (t) (Q → Qmin for Uc (t) → ∞). The dynamic compression K takes place in the area of the saturation threshold (SAT) and is K = Qo / Qmin - ^{in the} case of N = 4 filter stages connected in series (in Fig.l, however, only three filter stages are provided by means of three resonator circuits in one line) and that

Values QQ = 10 and Q _m i _n = 1 a strong compression around N 80dB (K _N = (Qo / Qmin)) can be realized.

In order to cover the entire human hearing range, a filter bank with resonance frequencies in the range of approximately 20 Hz to approximately 20 kHz is implemented, which is implemented by typically fifty to one hundred rows of resonator circuits 101 (i.e. n = 50 to n = 100). According to the exemplary embodiment described, the value of the inductance is set to L = 1 H. The respective value C is then calculated for each line of resonator circuits 101 in accordance with the filter frequency fg covered by this line from the resonance frequency of the corresponding LC element:

C = (4π ² f ² L) - ¹ (3)

It should be noted that the nonlinear quality factor Q for each filter frequency fg, ie for each line of resonator Circuits 101 is calculated independently. Referring to FIG. 1, this means that each row of oscillator circuits 101 is assigned a corresponding filter frequency fg, for which the value of the quality Q (t) is calculated.

A wave digital filter 300 as a realization of the resonator stage 101 shown in FIG. 2 is described below with reference to FIG.

A wave digital filter represents a class of digital filters with particularly favorable properties. They are modeled on traditional filters from the classic components of telecommunications and are operated using modern, integrated digital circuits. According to the

Technology of a wave digital filter can vividly implement an analog model digitally (for example using a computer).

The components of the

Wave digital filter 300 from FIG. 3 clearly assigned to the components of the resonator circuit 101 from FIG. 2 and the corresponding variables defined.

A first block 301 of the wave digital filter 300 contains a reflection-free serial coupler with the impedances R11 and R13. Rll clearly represents the adjustable ohmic resistance R 203, based on a reference resistance. R12 represents a corrected resistance (impedance) of the coil L 202 with respect to one

Base frequency. A second block 302 contains a parallel coupler which reproduces the parallel connection of the capacitance 201, the conductance values G21, G22, G23 being shown in the second block. G21 is an input conductance of the second block (G12 = I / R13) 302, G23 is an output conductance of the second block 302. The resistance of the capacitance C 201 is modeled by means of the conductance G22. A third block 303 represents a memory or a filter register for the capacitance 201 and a fourth block 304 represents a memory or a filter register for the coil 202.

The variables shown in FIG. 3 are defined below. The parameters for a wave digital filter for each filter frequency are:

R11 = R / R_B (4)

R12 = 2 ιτ F_B L / (R_B tan [7T F_B / f_s]) (5)

R13 = R11 + R12 (6) G21 = R13 ^"1 (7)

G22 = 2τr F_B C R_B / tan ("/ r F_B / f_s) (8)

G23 = G21 + G22 (9)

R is the ohmic resistor 203 and R_B is a predefinable reference resistance. F_B is a predefinable reference frequency. The values R_B and F_B are used for scaling. Since the implementation according to the exemplary embodiment described is implemented with double precision float variables, this normalization is not relevant, but it is relevant if integer arithmetic is used. L is the inductance of the coil 202. The value f_s is a sampling frequency of the sampled time signal. The variables R11, R12, R13 are ohmic resistors, whereas the variables G21, G22 and G23 are guide values, that is to say inverse ohmic resistors.

Filter coefficients gl, g2 result in: g2 = G2l / G23 (11)

The initial values of the filter registers ZI (fourth block 304) and Z2 (third block 303) are initialized to zero.

The signals at the individual ports can be calculated successively. For the "forward wave" of the signal, that is to say the coefficients on the arrows oriented to the right in accordance with FIG. 3, the following results: bl3 = - (U + Zl) (12) b20 = -g2 (Z2-bl3) (13) b23 = b20 + Z2 (14)

The quantity U in equation (12) is the input signal 200.

The coefficients for the "backward wave", that is to say the arrows oriented to the left in FIG. 3, are: b22 = b20 + b23 (15) b21 = b22 + Z2-bl3 (16) a0 = b21 -bl3 (17) bll = U-gl aO (18) bl2 = - (bll + b21) (19)

The output signal U _c 204 is then calculated as: U _c = (b22 + Z2 [sec]) / 2 (20) The filter registers (blocks 303, 304) are updated as follows:

Zl = -bl2 (21)

Z2 = b22 (22)

The output signal U _c 204 is transferred as an input signal U 200 to the filter stage 101 downstream of the filter stage 101 under consideration. Based on the output signal U _c 204 of the last filter stage 101 of a row of filter stages 101, the quality of the filters 101 connected in series is determined anew in accordance with equation (1). From the value for the quality Q determined in this way, the value of the resistance R determining the damping is calculated in accordance with equation (2). With the changed value of the ohmic resistor R 203, the filter resistances (R11, R12, R13, G21, G22, G23) and filter coefficients (gl, g2) are recalculated according to equations (4) to (11). After this step, the output signal is calculated for a next time slice. In other words, the time spectrum can be broken down into several time slices, which are successively calculated numerically.

4, a diagram 400 is explained in which the functionality of the circuit arrangement according to the invention according to a preferred exemplary embodiment of the invention is shown. Diagram 400 relates to a circuit arrangement with N = 4 resonator circuits connected in series. Qo = 10 is assumed as the maximum Q value, Q _m in = 1 is assumed as the minimum Q value.

The frequency of a signal normalized to a reference frequency fo is plotted along an abscissa 401 of the diagram 400 in a logarithmic representation. Along an ordinate 402 is the logarithmic representation System response to an input signal of a certain intensity is shown. First to eighth curves 403 to 410 represent the frequency response (ie here the respective value of the maximum amplitude of the filter output) of the circuit arrangement according to the invention for different signal amplitudes (based on a reference amplitude). The first curve 403 corresponds to an amplitude of IxlO ^"9 , the second curve 404 corresponds to an amplitude of IxlO ^"4 , the third curve 405 corresponds to an amplitude of IxlO ^{" 3} , the fourth curve 406 corresponds to an amplitude of IxlO ^"2 , the fifth curve 407 corresponds to an amplitude of IxlO ^{" 1} , the sixth curve 408 corresponds to corresponds to an amplitude of IxlO ¹ and the eighth curve 410 corresponds to an amplitude of lxl0 ^e an amplitude of 1x10 °, the seventh curve 409th Further, it is assumed as the input signal is a sine oscillation which is windowed with a cos ² window. Curves 403 to 410 result for an entire filter bank from N = 4 feedback-free connected resonator circuits.

It can first be seen from diagram 400 that the

Attenuation of the input signal is stronger, the higher the signal intensity or signal amplitude. At very small amplitudes, the filters are linear and the resonance increase is approximately 80dB. The response of the filter bank decreases very steeply at high low frequencies, since the filters are implemented as a low pass (cf. FIG. 2). The high-frequency response of the filters drops at around 6dB per octave, due to the scaling of the filter parameters with f ₀ . The curves in Fig. 4 simulate the highly asymmetrical frequency selectivity of the human ear in a good approximation.

The relationship between amplitudes of the input signal and the output signal of a circuit arrangement according to the invention is described below with reference to FIG.

In the diagram 500, a sound pressure level Aι along an abscissa in a logarithmic representation in _n dB plotted, based on a sound pressure of 20μPa. The strength of an output signal A ₀ uτ in dB is plotted along an ordinate 502 in arbitrary units. Curves 503 to 507 show the growth function of a filter cascade of four resonator circuits (series connection of N = 4 filters) at the resonance frequency f ₀ for different scenarios. Qmin - is assumed as the minimum quality.

A first curve 503 shows a linear growth function. A second curve 504 shows a growth function of the inner ear, i.e. the speed of the basilar membrane in relation to the sound pressure measured in front of the eardrum. The data of the second curve 504 are taken from [2]. A third curve 505 shows the curve course for a quality factor Q = 2, a fourth curve 506 shows the course for Q = 4 and a fifth curve 507 shows the course for Q = 10.

5 shows the growth function of a filter output for f = fg with the filter quality Q as a parameter. With very large and very small amplitudes, the growth functions are approximately linear. What is striking is the large compression range (especially with a large Q), which extends over more than four decades. The large dynamic range of the input signal (100 dB) is compressed to 40 dB- (for Q = 10). Due to the excessive resonance, quiet signals are "amplified" in a frequency-specific manner. The growth function replicates vibrational responses measured on the living hearing system very well (see curve 504). An approximately technical simulation of the nonlinear preprocessing in the inner ear is therefore realized with the circuit arrangement according to the invention.

A circuit arrangement 600 according to another preferred exemplary embodiment of the invention is described below with reference to FIG. The circuit arrangement 600 is formed from a first resonator circuit 601 and a second resonator circuit 602, each of which is constructed like the resonator circuit 101 shown in FIG. 2. The second resonator circuit 602 is the first resonator Circuit 601 connected downstream.

The circuit arrangement 600 can clearly be seen as a directly coupled implementation of two (N = 2) series-connected resonator circuits 601, 602.

As shown in FIG. 6A, the second connection of the coil 202 of the upstream resonator circuit 601 is coupled to the first connection of the ohmic resistor 203 of the downstream second resonator circuit 602.

According to the exemplary embodiment shown in FIG. 6A, resonator circuits coupled directly to one another

601, 602, the output voltage Uci of the upstream resonator circuit 601 is equal to the input voltage of the following resonator circuit 602. Furthermore, the output current of the first resonator circuit 601 is equal to the input current of the second resonator circuit 602.

It should be noted that the values of the resistors R1 and R2, the inductors L1 and L2 and the capacitances C1 and C2 of the resonator circuits 601, 602 can be different from one another or can be set / regulated differently.

A realization of the resonator circuits 601, 602 shown in FIG. 6A as a wave digital filter 650 is described below with reference to FIG. 6B. Clearly, the wave digital filter 650 is formed from a first component 651, which represents the first resonator circuit 601, and from a second component 652, which represents the second resonator circuit 602. According to the coupled configuration of the resonator circuits 601, 602 according to FIG. 6A, the two components 651, 652 are directly coupled to one another in the manner shown in FIG. 6B. The internal structure of each of the components 651, 652 essentially corresponds to that of the wave digital filter 300 from FIG. 3.

A circuit arrangement 700 according to yet another exemplary embodiment of the invention is described below with reference to FIG. 7A.

The circuit arrangement 700 is formed from a first resonator circuit 701 and a second resonator circuit 702, which are connected in series. The resonator circuits 701, 702 are clearly connected in series in a decoupled configuration, i.e. an intermediate element is connected between the resonator circuits 701 and 702.

Each of the resonator circuits 701, 702 is essentially constructed like the resonator circuit shown in FIG

101. Furthermore, an operational amplifier 703 is provided between the first resonator circuit 701 and the second resonator circuit 702, a non-inverting input 703a of the operational amplifier 703 being coupled to the second connection of the coil 202 of the upstream first resonator circuit 701. Furthermore, an inverting input 703b of the operational amplifier 703 is fed back with its output 703c and is coupled to the first connection of the ohmic resistor 203 of the second resonator circuit 702 connected downstream of the first resonator circuit 701. According to this configuration, the output voltage of the upstream resonator circuit 701 Uci 204 is equal to the input voltage of the second resonator circuit 702 downstream of the first resonator circuit 701. The output current of a respective resonator circuit is zero. The input current of the second resonator circuit 702 connected downstream of the first resonator circuit 701 is based solely on the impedance of the second resonator circuit 702 connected downstream. As shown in FIG. 7A, these circumstances can be realized in analog technology by means of an impedance converter which can Output voltage of the upstream resonator circuit 701 impresses the input of the downstream resonator circuit 702.

A wave digital filter 750 as a realization of the circuit arrangement 700 from FIG. 7A is described below with reference to FIG. 7B.

The wave digital filter 750 is divided into a first component 751 and a second component 752, the first component 751 representing the first resonator circuit 701 and the second component 752 representing the second resonator circuit 702. Due to the functionality of the operational amplifier 703, the two components 751, 752 are clearly coupled to one another. The internal structure of each of the components 751, 752 essentially corresponds to the configuration shown in FIG. The input signal of the first component 751 is U, the input signal of the second component 752 is Uci.

According to the invention, the combination of the linear filter bank 808 with the non-linear compression stages 101 forms the non-linear vibration behavior of the inner ear of the

Mammals very well, as explained in connection with Fig. 5 above. In particular, a large dynamic compression of sound levels in the range from 0 dBgp to 120 dBgp _L to a range from 1 nm to 100 nm (this corresponds to approximately 40 dB) is achieved (see FIG. 10), which is necessary for further processing as part of the feature extraction and speech recognition is very important.

10 shows in a diagram 1000, in which the cochlea position is shown along an abscissa 1001 and along the ordinate 1002 the deflection of the basilar membrane occurring at the respective cochlea position. Diagram 1000 thus clearly shows an excitation pattern (RMS value) of the nonlinear basilar membrane model for a 1 KHz tone. The curves shown

1003, 1004, 1005, 1006, 1007, 1008, 1009, are at position 1010 with the greatest sensitivity (shown in FIG. 10 as a dashed line) in the usually at the position of 21 mm in relation to the zero position of the cochlea, very strongly compressed. 10 also shows an excitation threshold 1011, above which the hearing of the human hearing system perceives a signal deflection.

That provided by the respective resonator circuits 101 at the end of a respective series connection

Basilar membrane signal X _ß M _l is supplied to a respective filter output circuit 1100 shown in FIG.

Each filter output circuit 1100 has a high-pass filter 1101, a rectifier circuit 1102 subsequently coupled on the output side, a low-pass filter 1103 subsequently coupled in the signal flow direction, an activation circuit 1104 subsequently coupled in the signal flow direction, and a vesicle pool circuit 1105 and a neurotransmitter circuit 1106 on. The respective basilar membrane signal XBMI _* X _ß Mi '••• _> ^x _BM n is high-pass filtered and scaled by means of the high-pass filter 1101 having a capacitance 1107 and an ohmic resistor 1108, so that only the speech-relevant dynamic range is used by means of a second-order Boltzmann function is extracted. The still relatively flat flank of the filter curves of the inner ear, as shown in FIG. 10, is tightened somewhat by means of the first-order high-pass filter 1101.

According to this exemplary embodiment of the invention, the cut-off frequency of the high-pass filter 1101 corresponds to the frequency of the maximum sensitivity of the basilar membrane oscillation.

The asymmetry of the Boltzmann function used in accordance with the invention causes the signal to be rectified (implemented according to this exemplary embodiment by means of the rectifier circuit 1102, which is low-pass filtered in the next stage by means of the low-pass filter 1103, so that a receptor potential signal UM is provided at the output of the low-pass filter 1103 becomes.

According to this exemplary embodiment of the invention, the ohmic resistance 1109 of the low-pass filter 1103 has a conductance g ^ = 60 nS in analogy to the cell membrane, and the capacitance 1110 of the low-pass filter 1103 has a capacitance of CM that simulates the cell membrane _. = 12 pF on.

The activation of the respective cell, simulated according to the invention by the vesicle pool circuit 1105 and the neurotransmitter circuit 1106, is calculated from the receptor potential U] ι using a first-order Boltzmann function.

Low pass filtering and rectification as described above have several effects: a) At low signal frequencies, there is exactly a maximum excitation of the sensory cells per cycle of acoustic excitation,

b) acoustic signals in the frequency range above the cut-off frequency of the inner hair cells lead to an activation according to their envelope, and

c) the sensitivity and saturation of the Boltzmann function cause the sound processing to focus on language-relevant information.

The receptor potential signal UM is further processed in this way, in other words, the vesicle pool circuit 1105 and the neurotransmitter circuit 1106 are set up in such a way that changes in time of the sound signal (that is to say the input signal) are emphasized and constant, essentially temporal constant signal components of the input signal are neglected (adapted).

In this way, stationary signals (e.g. noise) are effectively suppressed.

According to this exemplary embodiment of the invention, the adaptation is modeled by means of the vesicle pool circuit 1105, the replicated vesicle pool being continuously (but slowly) filled up to its desired value. A neurotransmitter current (according to this exemplary embodiment of the invention at a rate of) takes place from the vesicle pool circuit 1105, relative to the current vesicle pool size and a probability that is derived from the membrane potential of the inner hair cell with a Boltzmann function 28,000 / s). In the case of sound signals with a large amplitude, a large part of the vesicle pool is broken down, so later signal components only generate a small signal, that is to say a signal with a small amplitude.

The vesicle pool regenerates itself in phases with an input signal of low amplitude. In other words, the vesicle pool circuit 1105 emulates the functionality described above, which is set up in such a way that two time constants are implemented, namely a first time constant of τi = 140 ms and a second time constant% ₂ = 3 ms) ,

The neurotransmitter current flows into the “synaptic gap”, where it is reduced with a time constant T ₃ = 1 ras in that of the neurotransmitter, reproduced according to the invention by means of the neurotransmitter circuit 1106.

In addition to the vesicle function from the vesicle pool, another neurotransmitter current arises that only depends on the membrane potential of the inner hair cell, which is why the model chosen is based on an infinite vesicle pool size and a rate of 9,000 / s ,

The two neurotransmitter currents allow adequate coding of stationary and transient sound signals, i.e. an adequate coding of nerve action potentials.

The vesicle pool 1105 can be modeled both continuously and consisting of discrete vesicles. With discrete modeling, the neurotransmitter current results as a stochastic process. This procedure is chosen to encode the switching signal in discrete nerve action potentials. A nerve action potential is shown in FIG. 12 in a nerve action potential diagram 1200 and is triggered when the concentration of the respective neurotransmitter in the synaptic cleft exceeds a predetermined threshold value, according to this exemplary embodiment 1.0 vesicle.

12 shows the generated model nerve action potentials when stimulated with an artificial vowel "e".

Excitation occurs at both formant frequencies of the vowel “e”. Furthermore, the temporal structure (in particular with the second formant) that is modulated with the basic speech frequency (100 Hz corresponds to 10 ms) is achieved.

A very advantageous property of the feature extraction unit 801 according to the invention is that it uses the achievable recognition performance in the context of a speech recognition method of an automatic

Speech recognition system can be evaluated and optimized.

In FIG. 13, the speech recognition performance of a conventional Fast Fourier Transformation-based speech recognition method (word error rate curve 1301) with different processing stages according to the invention (inner ear model component only), word error rate curve 1302) and sensor cell word error rate curve 1303 is shown in a diagram 1300 different noise levels (plotted along the abscissa).

The ordinate 1305 shows the word error rate achieved in each case in FIG. 13.

As shown in Fig. 13, the detection performance is the usual fast without the existence of noise Fourier transformation-based methods of higher quality, which can be attributed in particular to the degree of maturity of the algorithms that have been developed for years, but the robustness of the features provided according to the invention is evident with increasing noise.

The following publication is cited in this document:

[1] E.G. Schukat-Talamazzini, Automatic Speech Recognition, Friedrich Vieweg & Sohn Publishing Company, Braunschweig-Wiesbaden, ISBN 3-528-05492-1, Chapters 1 to 3, 1995;

[2] S. Seneff, A Computational Model for the Perforal Auditory System: Application to Speech Recognition Research, Proceedings of IEEE ICASSP 1986, Tokyo, pages 1983 to 1986, Tokyo, 1986;

[3] M. Hunke and T. Holton, Training Machine Classifiers to Match the Performance of Human Listeners in a Natural Vowel Classification Task, ICSSLP, pages 574 to 577, 1996;

[4] S. Sandhu and 0. Ghitza, A Comparative Study of MEL Cepstra and EIH for Phone Classification under Adverse Conditions, IEEE, pages 409 to 412, 1995;

[5] C. Sumner et al. , A Revised Model of the Inner-Hair-Cell and Auditory Nerve Complex, Journal of Acoustic Society of America, Vol. 111, pages 2178 to 2188, May 2002;

[6] T. Dau, model of effective signal processing in the ear, insights, No. 29, pages 16 to 18, April 1999;

[7] V. Hohmann, signal processing in digital hearing aids, insights, No. 33, pages 24 to 26, June 2001;

[8] H. W. Strube, A Computationally Efficient Basilar-Membrane-Model, Acoustica, Vol. 58, pages 207 to 214, 1985;

[9] MA Ruggero et al. , Mechanical Basis of Frequency Tuning and Neural Exitation at the Base of the Cochlea Comparison of Basilar-Membrane-Vibration and Auditory Nerve-Fiber-Responses in Chinchilla, Proceedings of National Academy of Science USA, Vol. 97, No. 22, pages 11744 to 11750, October 2000;

[10] P. Dallos et al. , The Cochlea, ISBN 0387944494, Springer-Verlag, chapter 6, pages 318 to 385, 1998;

[11] DE 691 31 095 T2.

LIST OF REFERENCE NUMBERS

100 circuit arrangement

101 resonator circuits

102 global input signal

103 (sound) signal source

104 first local output signal

105 first local input signal

106 second local output signal

107 second local input signal

108 third local output signal

109 global output signal 111 control circuit

200 input signal

201 capacity

202 inductance

203 adjustable ohmic resistance

204 ^" output signal

300 wave digital filters

301 first block (serial coupler)

302 second block (parallel coupler)

303 third block (storage element for capacity)

304 fourth block (storage element for inductance)

400 diagram

401 abscissa

402 ordinate

403 first curve

404 second curve

405 third curve

406 fourth curve

407 fifth curve 408 sixth curve 409 seventh curve

410 eighth curve

500 diagram

501 abscissa

502 ordinate

503 first curve

504 second curve

505 third curve 506 fourth curve 507 fifth curve

600 circuit arrangement

601 first resonator circuit

602 second resonator circuit

650 wave digital filter

651 first component

652 second component

700 circuit arrangement

701 first resonator circuit

702 second resonator circuit

703 operational amplifier

703a non-inverting input 703b inverting input 703c output

750 digital wave filters

751 first component

752 second component

800 speech recognition system

801 feature extraction system

802 speech recognition device 803 input signal

804 ear canal model component 805 signal

806 middle ear model component 807 signal

808 inner ear model component

809 signal

810 sensor cell model component

811 signal

812 Synaptic model component

813 signal

901 inductance

902 ohmic resistor 903 filter bank

904a inductance filter stage

904b Ohmic resistance filter stage

904c capacity filter stage

905 filter stage

905a inductance filter stage

905b Ohmic resistance filter stage

905c capacity filter stage

906 filter stage

906a inductance filter stage

906b Ohmic resistance filter stage

906c capacity filter stage

907 ohmic terminating resistor

XBM basilar membrane signal

1000 diagram

1001 abscissa

1002 ordinate

1003 excitation curve

1004 excitation curve

1005 excitation curve

1006 excitation curve

1007 arousal curve 1008 arousal curve 1009 arousal curve 1010 digit with the greatest sensitivity

1011 excitation threshold

1100 filter output processing circuit

1101 high pass

1102 rectifier circuit

1103 low pass filter

1104 activation circuit

1105 vesicle circuit

1106 neurotransmitter circuit

1107 capacity high pass filter

1108 ohmic resistance high pass filter

1109 ohmic resistance low pass filter

1110 capacity low pass filter

1200 diagram

1300 diagram

1301 word error rate curve

1302 word error rate curve

1303 word error rate curve

1304 abscissa

1305 ordinate

Claims

claims

1. Circuit arrangement • with a filter bank with a plurality of filter stages and a filter bank input to which an input signal can be fed, • with a plurality of resonator circuits for generating a partial output signal from the input signal, each resonator circuit, Circuit each assigned to a filter stage of the plurality of filter stages and coupled to an output of the respective filter stage, • wherein each resonator circuit has: - a capacitance, - an inductance, - a resonator output, at which the respective partial output signal can be provided ; At least one resonator control circuit for controlling or regulating the quality of at least one resonator circuit, the at least one resonator control circuit being set up in such a way that it controls the quality of the resonator circuit as a function of the time course of the signal amplitude of the input signal and / or the part output signal of the resonator circuit controls or regulates.

2. Circuit arrangement according to claim 1, • in which each resonator circuit has a plurality of series-coupled part-resonator circuits, and • wherein at least one of the part-resonator circuits is coupled to an output of the resonator circuit ,

3. Circuit arrangement according to claim 1 or 2, wherein the filter bank is designed as a linear digital will filter.

4. Circuit arrangement according to one of claims 1 to 3, in which at least one of the resonator circuits has a controllable ohmic resistance by means of the resonator control circuit.

5. Circuit arrangement according to one of claims 1 to 4, with a plurality of high-pass filters, each filter stage being assigned at least one high-pass filter, a high-pass filter being coupled to the output of a respective resonator circuit.

6. The circuit arrangement as claimed in claim 5, in which at least some of the high-pass filters are designed as first-order high-pass filters.

7. The circuit arrangement as claimed in claim 6, in which the cut-off frequency of at least part of the high-pass filters of the first order is chosen such that it corresponds to the frequency of the maximum sensitivity of a basilar membrane vibration of an inner ear of a mammal.

8. Circuit arrangement according to one of claims 5 to 7, with a plurality of rectifier circuits, wherein in each case a rectifier circuit is assigned to one of the filter stages and a high-pass filter and is coupled to an output of a respective high-pass filter.

9. Circuit arrangement according to claim 8, with a plurality of low-pass filters, each having a low-pass filter assigned to a rectifier circuit and coupled to an output of a respective rectifier circuit.

10. Circuit arrangement according to claim 9, with a plurality of activation circuits, each with an activation circuit being assigned to one of the filter stages, each activation circuit being set up to amplify a change over time in a signal supplied to the activation circuit and for damping components which are essentially constant over time Activation circuit supplied signal.

11. A circuit arrangement according to claim 10, wherein each activation circuit has a vesicle pool circuit with a plurality of vesicle circuits.

12. Circuit arrangement according to one of claims 1 to 11, wherein the resonator control circuit is set up such that it controls the quality of the at least one resonator circuit based on a Boltzmann function and / or its derivation, the Boltzmann function contains the amplitude of the respective partial output signal as a parameter.

13. Circuit arrangement according to one of claims 1 to 12, wherein the resonator control circuit is set up in such a way that it determines the quality of the at least one resonator circuit as a function of the amplitude of the respective part output signal based on one for one Sensitivity characteristic determined by a human ear.

14. Circuit arrangement according to one of claims 1 to 13, wherein the resonator control circuit is set up in such a way that it sets the quality of the at least one resonator circuit, the lower the higher the amplitude of the respective partial output signal ,

15. Circuit arrangement according to claim 14, in which the resonator control circuit is set up in such a way that it adjusts the quality of the at least one resonator circuit in a non-linear dependence on the amplitude of the respective partial output signal.

16. Circuit arrangement according to one of claims 1 to 15, wherein the resonator control circuit is set up in such a way that it adjusts the quality of the at least one resonator circuit in such a way that the amplitude of the respective partial output signal is within a predetermined interval is.

17. Circuit arrangement according to one of claims 1 to 16, set up for processing an acoustic signal as an input signal of the filter bank.

18. Signal processing device • with a circuit arrangement according to one of claims 1 to 17, and • with a further processing unit for further processing of the signal provided by the circuit arrangement.

19. Signal processing device according to claim 18, wherein the further processing unit is a speech recognition device or a hearing aid.