AT522636B1 - Medical detector and method for extra pulse detection - Google Patents

Abstract

The invention provides a medical detector (1) and a method for automatically detecting an extra pulse (ex) which can indicate a medical condition such as diplophony. The medical detector (1) comprises an input unit (10) which is configured to receive an input signal (d), such as a phonation signal, the input signal (d) being an oscillation of a medical measured value and / or an oscillation of a human body part such as a glottis area; a cyclic pulse train estimation unit (20) configured to estimate a cyclic pulse train component (c1) of the input signal (1); a cyclic pulse train subtracting unit (20) configured to subtract the cyclic pulse train component (c1) from the input signal (d) to thereby obtain a cyclic model error signal (e1); and an extra pulse detector unit (6, 60) configured to detect an extra pulse (ex) in the input signal (1) based on the cyclic model error signal (e1); and upon detection of the extra pulse (ex) to output an output signal (o) which indicates the detected extra pulse (ex).

Description

description

The present invention relates to the field of medical devices and, more particularly, to a medical detector and method for detecting extra pulses.

Extra pulses that represent a deviation or irregularity in an otherwise cyclical vibration of a medical reading, such as a medical reading that indicates vibration of a human body part, can indicate a medical condition. For example, a human phonation process involves a cyclical oscillation of a glottic area or glottis between two vocal folds. Research has shown that auditory characteristics of the hoarseness and roughness of a human voice, such as crackling, scratching or diplophony, can be associated with extra pulses that occur in quasi-closed phases of an oscillation cycle of the glottic area. The diagnosis and treatment of voice disorders can benefit from knowledge of the presence of such extra pulses. However, there is a risk that extra impulses will be overlooked or ignored in clinical practice.

It is therefore an object of the present invention to provide a medical detector for detecting extra pulses which is capable of automatically detecting the presence of an extra pulse in an oscillation of a medical measurement.

Accordingly, a medical detector for detecting extra pulses comprises an input unit configured to receive an input signal indicative of a vibration of a medical reading; a cyclic pulse train estimation unit configured to estimate a cyclic pulse train component of the input signal; a cyclic pulse train subtracting unit configured to subtract the cyclic pulse train component from the input signal to thereby obtain a cyclic model error signal; and an extra pulse detector unit configured to detect an extra pulse in the input signal based on the cyclic model error signal and, upon detection of the extra pulse, to output an output signal indicative of the detected extra pulse.

The proposed medical detector can thus be described as being configured to use a waveform-based modeling approach to model, by estimation, a cyclic pulse train component of the input signal indicative of an oscillation of a medical reading. The cyclic pulse train component is subtracted or removed from the input signal and the extra pulse is automatically detected based on the remaining cyclic model error signal. The proposed medical detector offers the advantage that it enables reliable and automatic detection of an extra pulse in the oscillation of the medical measured value. Such automatic recognition offers the advantages of being faster than visual recognition, of being cheaper when using larger amounts of data, and of not requiring any expert knowledge on the part of the operator.

The medical reading can be a medical reading that indicates vibration of a human body part. The human body part can, for example, be a glottis or a vascular muscle. The medical reading can also be any other type of oscillating medical reading, such as a blood sugar level, a body temperature, or an EEG, for example over the course of a day.

An extra pulse can be a pulse that occurs between the times of the cyclical pulses of the input signal. The cyclical pulses can be pulses that occur at regular intervals and / or with regular amplitudes. "Regular intervals" also potentially includes some cycle duration modulation, and "regular amplitudes" also potentially includes pulse height modulation. A large number of such extra pulses can occur randomly or in a regular pattern.

The input signal may be a phonation signal (medical reading) which is a vibration of a glottic area between two vocal folds in the human larynx

(human body part). For example, a signal value or an instantaneous amplitude of the input signal can correspond to an instantaneous amount of the opening of the glottic area. The extra pulse can be a pulse or a peak that occurs during a quasi-closed phase of the vibration of the glottic area, i.e. a phase in which the glottic area is expected to be closed. Further examples of the input signal include a glottal width measured at a predefined sagittal position (medical reading), in particular the glottal width measured using a kymogram of a vocal fold vibration or a phonovibrogram, ie a spatial-temporal representation of a movement of edges of the vocal folds, a movement path of the vocal fold edge, in particular a Movement path measured using a kymogram of vocal folds vibration or a phonovibrogram, or an electroglottogram (EGG), ie an electrical impedance of the larynx over time, which is proportional to the contact area of the vocal folds.

In another example, the input signal may be a heart signal indicating vibration of a human heart or a myocardium (human body part), and the extra pulse may correspond to an ectopic beat of the heart or myocardium.

The input signal can be an acoustic, electrical or optical signal. The input signal can be an analog signal. The input signal can preferably be a digital signal or a sample which is recorded with a predetermined sampling frequency and quantized with a predetermined quantization resolution. The predetermined sampling frequency may preferably be higher than 1 and lower than 100 kHz. Particularly preferred predetermined sampling frequencies are 4 kHz, 44.1 kHz, 48 kHz and 50 kHz. The predetermined quantization resolution can preferably be higher than 1 and lower than 32 bits. Particularly preferred predetermined quantization resolutions are 8 bits, 16 bits and 24 bits.

In general, the terms "signal", "(signal) component" and "pulse shape" as used herein can each correspond to a respective discrete or continuous time series of signal values.

The input unit can receive the input signal by reading a digital record of the vibration of the human body part, such as a digital record stored in a file in a memory of a computer or a digital record obtained through communication with an external unit is received over a network. Alternatively, the input unit can receive the input signal directly from a detection device such as a microphone or a camera, for example an endoscopic camera as used in vocal fold videolaryngoscopy.

The cyclic pulse train estimation unit can estimate the cyclic pulse train component of the input signal by extracting a basic frequency from the input signal and generating the cyclic pulse train component as a sequence of cyclic pulses with the extracted basic frequency.

The cyclic model error signal achieved by the cyclic pulse train subtraction unit can be described as what is left of the input signal after the estimated cyclic pulse train component has been subtracted from the input signal. In other words, the cyclic model error signal can comprise a non-cyclic component of the input signal that is not covered (taken into account, described) by the estimated (modeled) cyclic pulse train component. In particular, the cyclical model error signal can include an extra pulse, but not the cyclical pulses. Using the cyclic model error signal, the extra pulse detection unit can therefore automatically detect the extra pulse with high accuracy without interference from the cyclic pulses.

[0015] The extra pulse detector unit can be configured, for example, in such a way that it detects an extra pulse by selecting maximum values (“peak picking”) based on one or more predetermined threshold values. In a variant, the extra pulse can be used

using a waveform-based approach, as described in more detail below.

The medical detector can be configured to carry out the detection of an extra pulse repeatedly in succession or in parallel in order to detect a plurality of extra pulses in the input signal.

The output signal that is output by the extra pulse detector unit when the extra pulse is detected can be any output signal that is suitable for displaying the detected extra pulse to an operator of the medical detector or an external device. The output signal can in particular be analog or digital and can be an electrical, optical, acoustic or visual signal. For example, the output signal can be light that is emitted by a light-emitting diode provided in the medical detector, or it can be an electrical signal that is provided at an output port of the medical detector.

It is noted that the output signal may indicate the fact of the presence of the extra pulse (at least one extra pulse or a train or series of extra pulses) in the input signal. In some examples, the output signal can indicate further information about the extra pulse, such as the pulse width, pulse height and / or phase (time of occurrence relative to a cyclical pulse) of the extra pulse. In a further example, the output signal can be a signal that is configured to cause a display unit to display a graphic representation of one or more signals and / or components that are involved in a modeling carried out by the medical detector, such as in particular Input signal, the cyclic pulse train component, the cyclic model error signal and / or the recognized extra pulse.

According to an embodiment of the medical detector, the cyclic pulse train estimation unit comprises a fundamental frequency extraction means which is configured to extract a fundamental frequency from the input signal; cyclic unit pulse train generator means configured to generate a cyclic unit pulse train having the extracted fundamental frequency; cyclic pulse shape estimating means configured to estimate an average cyclic pulse shape of the cyclic pulse train component of the input signal based on the input signal and the unit cyclic pulse train; and a cyclic pulse train synthesizing means configured to synthesize the cyclic pulse train component of the input signal estimated by the cyclic pulse train estimation unit based on the cyclic unit pulse train and the average cyclic pulse shape.

The fundamental frequency extraction means can be configured to extract the fundamental frequency from the input signal, for example using an average amount difference function (AMDF), an input autocorrelation function (AC), an autocorrelation function of a centrically clipped signal (ACC), an autocorrelation function of an inversely filtered signal (IFAC), a signal cepstrum (CEP), a harmonic product spectrum (HPS) of the input signal, a waveform fitting function (WM), a signal spectrum (e.g. FFT, STFT), a hidden Markov model (HMM), a model error-based candidate selection or like that. In particular, the fundamental frequency can be extracted by tracing candidates, first using an HMM that uses heights and frequencies of spectral peaks as input observation sequence as input, second by candidate waveform Fourier synthesis using instantaneous phases of unit pulse trains with candidate fundamental frequencies and Fourier Average pulse shape coefficients as inputs and thirdly by selecting a particular fundamental frequency candidate that minimizes modeling error.

According to a variant, several simultaneous fundamental frequencies can be extracted, and several cyclic unit pulse train components can accordingly be obtained and taken into account in the waveform modeling in the same way as is described by way of example for a cyclic unit pulse train.

The cyclical unit pulse train can be a train of unit pulses or delta pulses at regular time intervals which are defined by the extracted fundamental frequency.

The cyclic pulse train synthesizing means may be configured to synthesize the cyclic pulse train component by pulse shape filtering the cyclic unit pulse train with the average cyclic pulse shape using, for example, Fourier synthesis, convolution in the time domain and / or multiplication in the frequency domain.

The medical detector according to the present embodiment can offer the advantage of being able to estimate a realistic model of the cyclic pulse train component (modeled cyclic pulses) with high reliability.

According to a further embodiment of the medical detector, the cyclic pulse shape estimating means is configured to estimate the average cyclic pulse shape of the cyclic pulse train component of the input signal by cross-correlation between the input signal and the cyclic unit pulse train.

Cyclic pulses of the input signal can correlate with the cyclic unit pulse train and are allowed to pass the cross-correlation, while extra pulses of the input signal do not correlate with the cyclic unit pulse train and are therefore averaged out. As a result, the medical detector can advantageously estimate an average cyclic pulse shape that is not influenced by the shape of an extra pulse or noise, and reproduce the shape of the cyclic pulses of the input signal more precisely.

According to a further embodiment of the medical detector, the cyclic pulse train synthesizing means of the medical detector is configured to synthesize the cyclic pulse train component of the input signal by means of Fourier synthesis using an instantaneous phase of the cyclic unit pulse train and Fourier coefficients of the estimated average cyclic pulse shape.

By using Fourier synthesis, the cyclic pulse train component of the input signal can be synthesized with favorable properties, i.e., with fewer artifacts, thereby enabling detection with higher accuracy. In particular, pulse shapes can vary according to an instantaneous frequency in Fourier synthesis, e.g. the pulse width can expediently decrease with increasing frequency in Fourier synthesis. In this way, glitch-like artifacts can be avoided, which would occur when synthesizing using FIR filtering or convolution, if pulses of adjacent cycles overlap in time if the frequency exceeds a reciprocal value of the pulse width. The Fourier synthesis can also enable an appropriate band limitation, which prevents aliasing even if a modulation noise is not band limited.

The Fourier coefficients can be real and imaginary coefficients, or sizes and phases, of a discrete Fourier transform of the estimated average cyclic pulse shape.

According to a further embodiment of the medical detector, the cyclic pulse train estimation unit further comprises a parameter optimization means which is designed to minimize the cyclic model error signal by means of a parameter variation of an amplitude and a deviation from the fundamental frequency for each of a plurality of pulses of the cyclic unit pulse train.

Here, the parameter optimization means can minimize the cyclic model error signal by minimizing its error level. The error level of the cyclic model error signal can be any measure that is suitable for describing the amount of error, deviation, or lack of correspondence between the input signal and the cyclic pulse train component. For example, error measurements in the time domain or in the frequency domain can be used:

By varying the amplitude of the cyclic pulses of the cyclic unit pulse train for each individual pulse, the cyclic waveform model, which is generated by the cyclic pulse train

component is represented which is synthesized using the cyclic unit pulse train causing fluctuations in the height of each pulse, which, within predetermined limits, is a normal phenomenon in vibration of a medical measurement value, such as a medical measurement value indicating vibration of a human body part. By varying a deviation from the basic frequency for each individual pulse, which can also be described as varying a cycle length of the pulses, the cyclic waveform model can cause so-called jitter or variations in the timing of the pulses, which is also a normal phenomenon within predetermined limits. Modulation noise such as jitter and amplitude modulation can be characterized as a random process and can be described and modeled, for example, by its probability density functions (PDF), its autocorrelation function (ACF), its power spectral density (PSD), parameters of an auto-regressive (AR) model and its stationarity properties.

By taking modulation noise or micro-variations such as jitter and flicker into account, a more realistic and precise modeling of the cyclic pulses of the input signal is made possible. As a result, artifacts in the cyclic model error signal are reduced and the medical detector can detect the extra pulse with greater accuracy.

According to a further embodiment of the medical detector, the cyclic pulse shape estimation means is further configured to linearize an instantaneous phase of the input signal and the cyclic unit pulse train in such a way that deviations from the fundamental frequency in the cyclic unit pulse train resulting from the parameter optimization means are compensated performed parameter variation results before the average cyclic pulse shape of the cyclic pulse train component of the input signal is estimated.

It is noted that "linearizing an instantaneous phase" can also be described as "balancing an instantaneous frequency". It is noted that "balancing an instantaneous frequency of the cyclic unit pulse train" also means linearizing an instantaneous frequency of the cyclic pulse train component which is synthesized based on the cyclic unit pulse train.

In particular, the cyclic pulse shape estimating means may be configured to determine the average cyclic pulse shape by cross-correlation between the input signal and the cyclic unit pulse train to compensate for pulses such as extra pulses that do not correlate with the cyclic unit pulse train. If in this case the cyclic unit pulse train is jittered, ie if the unit pulses occur in the cyclic unit pulse train with deviations from the fundamental frequency as a result of the parameter variation carried out by the parameter optimization means, the cross-correlation between the input signal and the jittered cyclic unit pulse train can artifacts in the estimated average cyclic Cause pulse shape. That is, it may be the case that, due to the jitter, some sections of the average cyclic pulse shape are incorrectly averaged, while certain other sections that do not belong to the average cyclic pulse shape and are to be averaged are actually not averaged. According to the present embodiment, by linearizing an instantaneous phase and / or compensating for an instantaneous frequency of the cyclic unit pulse train, such artifacts in the average cyclic pulse shape can advantageously be avoided, and the quality of the average cyclic pulse shape can be improved.

According to a further embodiment of the medical detector, the parameter optimization means is set up to determine the error level of the cyclic model error signal as a decibel ratio of a root mean square value of the cyclic model error signal and a root mean square value of the input signal.

Here, the error level can be an energy content of the cyclic model error signal relative to the energy content of the input signal. Thus, a cyclic pulse train component can be viewed as an optimized model of the cyclic pulses of the input signal if it minimizes the energy content of the cyclic model error signal.

According to another embodiment, the medical detector further comprises an extra pulse train estimation unit configured to estimate an extra pulse train component of the input signal based on the cyclic model error signal; and an extra pulse train subtracting unit configured to subtract the extra pulse train component from the cyclic model error signal to thereby obtain a residual error signal; and wherein the extra pulse detection unit is configured to detect the extra pulse by comparing the cyclic model error signal with the residual error signal.

In particular, the extra pulse detector can detect the extra pulse by comparing an error level of the cyclic model error signal with an error level of the residual error signal.

The medical detector according to the further embodiment can thus be described so that it is configured so that it uses not only a waveform-based modeling method to model the cyclic pulse train component of the input signal by estimation, but also a waveform model-based one Approach to model the extra pulse occurring in the input signal, which indicates the oscillation of the medical measured value.

The model cyclic error signal can be described as what remains of the input signal after only the modeled cyclic pulse train component has been subtracted or removed. The residual error signal can be described as what remains of the input signal after both the estimated cyclic pulse train component and the estimated extra pulse train component have been subtracted or removed. The model detector can detect the extra pulse when the error level of the residual error is lower than the error level of the cyclic model error signal. That is, the model detector can advantageously detect the extra pulse based on the determination that the input signal can be better modeled by including the extra pulse train component in the model.

As a result, the medical detector according to the further embodiment can use a waveform model-based approach in order to advantageously detect the extra pulse with higher accuracy and higher reliability in the correctness of the model.

According to a further embodiment of the medical detector, the fundamental frequency extraction means of the cyclic pulse train estimation unit is configured such that it extracts the fundamental frequency from the input signal by minimizing the cyclic model error signal and / or the residual error signal.

The medical detector according to the present embodiment can thus be described in such a way that it carries out an iterative optimization process in order to extract the fundamental frequency advantageously with higher accuracy.

According to a further embodiment of the medical detector, the parameter optimization means of the cyclic pulse train unit is further configured to minimize the residual error signal by means of the parameter change in the amplitude and the deviation from the basic frequency for each of the plurality of pulses of the cyclic unit pulse train.

In other words, the parameter optimization means can not only minimize the cyclical model error signal, but also simultaneously or alternatively minimize the residual error signal while a parameter variation is being carried out. This can offer the advantage of compensating for jitter and / or amplitude modulation, which are contained in the input signal, with greater accuracy.

According to a further embodiment of the medical detector, the extra pulse detector unit is set up to determine the error level of the cyclic model error signal as a decibel ratio of a root mean square value of the cyclic model error signal and a root mean square value of the input signal; determine the error level of the residual error signal as a decibel ratio of a root mean square value of the residual error signal and the root mean square value of the input signal; and to determine the extra pulse if the error level of the residual error signal is less than a predetermined threshold value.

n is less than the error level of the cyclic model error signal.

The respective error level can be the energy content of the respective error signals relative to the input signal. The extra pulse detector unit can determine that the oscillation of the medical measured value can be modeled better by a combination of the estimated cyclic pulse train component and the estimated extra pulse train component than by the cyclic pulse train component alone, if the inclusion of the estimated extra pulse train component in the model is suitable for the energy content of the residual error signal minimize.

Furthermore, the medical detector according to the present embodiment is advantageously able to reduce statistical artifacts and to detect the extra pulse with higher accuracy and reliability. That is, by using the predetermined threshold, which may be a predetermined non-zero threshold, the extra pulse detector unit can detect a case where the error level of the residual error signal is lower than the error level of the cyclic model error only for statistical reasons, such as because of additional Degrees of freedom of the model, and can only detect the extra pulse if the error level of the residual error signal is significantly below the error level, in a way that cannot be explained by such statistical reasons alone. The predetermined threshold can be determined or calibrated based on clinical data. In one example, the predetermined threshold can be chosen to be 1 dB.

According to a further embodiment, the extra pulse detector unit can be configured to compare the cyclic model error signal with the residual error signal using a statistical test.

For example, the extra pulse detector may be configured to calculate a p-value of the cross correlation performed by the cyclic pulse shape estimator and / or the extra pulse shape estimator. When calculating the p-value of the cross-correlation, the extra heart rate detector can detect the extra heart rate if the p-value is less than a predetermined threshold, such as 0.05.

By using a statistical test to compare the cyclical model error signal with the residual error signal, the decision as to whether an extra pulse is detected can advantageously be made with greater reliability.

According to a further embodiment of the medical detector, the extra pulse train estimation unit further comprises an extra pulse triggering means configured to estimate one or more extra pulse trigger times from the cyclic model error signal; extra unit pulse train generator means configured to generate an extra unit pulse train including an extra unit pulse at each estimated extra pulse trigger time; an extra pulse shape estimating means configured to estimate an average extra pulse shape of the extra pulse train component of the input signal based on the cyclic model error signal and the extra unit pulse train; and an extra pulse train synthesizing means configured to synthesize the extra pulse train component of the input signal estimated by the extra pulse train estimation unit based on the extra unit pulse train and the average extra pulse shape.

Each extra pulse trigger time can be a point in time or a position on a time axis of a pulse in the cyclic model error signal, which can indicate an extra pulse in the input signal that occurs at the corresponding extra pulse trigger time.

In an example of the present embodiment, the extra pulse triggering means may be a peak pickup means configured to pick up a peak value with at least a minimum peak prominence, the peak prominence or relative peak height as the difference between a local maximum signal and an adjacent local minimum signal value is defined, and / or receives a peak with a minimum absolute peak height. That is, the peak acquisition means can be configured to capture one or more peaks based on a comparison of a relative peak height and / or an absolute peak

to select zenhöhe with a respective predetermined threshold. The respective predetermined threshold values can be predetermined or calibrated based on clinical data.

According to a variant of the example, a number of peaks within a cycle of the cyclical model error signal can be limited to a maximum number. If there are multiple peaks within a cycle of the cyclic model error signal, only the highest peak or peaks can be selected. For example, the maximum number can be one.

In another example of the present embodiment, the extra pulse triggering means may be configured to estimate the one or more extra pulse trigger times from the cyclic model error signal by minimizing the residual error signal. The minimization of the residual error signal can in particular include the minimization of an energy level of the residual error signal. That is to say, the extra pulse triggering means can be configured to carry out an iterative optimization process by varying a candidate extra pulse triggering time and to select, as the estimated extra pulse triggering time, the candidate extra pulse triggering time for which the resulting residual error signal output by the extra pulse train subtracting unit of the medical detector is a minimum. This iterative approach has the advantage that no peak acquisition device is required and errors and uncertainties associated with peak recording are avoided. In particular, when selecting the peaks, it may be difficult to distinguish between a peak representing an extra pulse and a peak caused by noise. Advantageously, by eliminating the need for tip selection, the medical detector according to the present example of the present embodiment may be able to detect the extra pulse with greater accuracy and reliability.

According to another example of the present embodiment, the extra pulse triggering means may be configured to estimate the extra pulse trigger time as a time with a fixed phase shift or a fixed average phase shift with respect to the cyclic pulses of the cyclic pulse train component. The fixed phase shift can be 180 degrees. According to the present example, the medical detector may be able to avoid the erroneous detection of noise as an extra pulse based on a theoretical model for medical readout that predicts the occurrence of extra pulses at a certain fixed phase shift with respect to the cyclic pulses.

According to yet another example of the present embodiment, the extra pulse triggering means can be configured to estimate the one or more extra pulse trigger times such that the one or more extra pulse trigger times occur during one or more cycles at a frequency that is different from the fundamental frequency of the cyclic Pulse train component is different. This can be advantageous in diplophony, for example, when two fundamental frequencies are detected simultaneously by the fundamental frequency extraction means.

The extra pulse train synthesizing means may be configured to synthesize the extra pulse train component by pulse shape filtering the extra unit pulse train with the average extra pulse shape estimated by the extra pulse shape estimating means using, for example, Fourier synthesis, convolution in the time domain and / or multiplication in the frequency domain becomes.

The extra pulse triggering means, the extra pulse train generator means, the extra pulse shape estimating means and the extra pulse train synthesizing means advantageously enable the medical detector to estimate a realistic model of the extra pulse train component of the input signal, thereby improving the accuracy of the extra pulse detection.

According to a further embodiment of the medical detector, the extra pulse shape estimating means of the medical detector is configured to estimate the average extra pulse shape of the extra pulse train component of the input signal by cross correlation between the cyclic model error signal and the extra unit pulse train.

Extra pulses contained in the cyclic model error signal can correlate with the extra unit pulse train and are allowed to cross-correlate, while other components of the cyclic model error signal, such as noise and the like, may not correlate with the extra unit pulse train and can be averaged out. Thereby, the medical detector can advantageously estimate an average extra pulse shape that is not influenced by noise such as Gaussian noise and the like.

According to a further embodiment of the medical detector, the extra pulse train estimation unit further comprises an extra pulse phase determining means configured to determine an extra pulse phase during which the extra pulse is to be expected, and an error signal suppression means configured to be suppressed deriving the cyclic model error signal from the cyclic model error signal by muting the cyclic model error signal during the times outside the extra pulse phase, and wherein the extra pulse triggering means is configured to estimate the one or more extra pulse trigger times from the suppressed cyclic model error signal. In particular, the extra pulse triggering means can estimate the one or more extra pulse trigger times only from the suppressed cyclic model error signal and not from the non-suppressed cyclic model error signal.

The extra pulse phase can be a phase or period of time during which no pulse is normally expected and in particular no cyclic pulse is expected. If the oscillation of the medical reading indicates an oscillation of an area of the glottis, the extra pulse phase may be a phase during which the glottis of a human vocal apparatus is expected to be closed or quasi closed.

The extra pulse phase determining means may determine the extra pulse phase as a phase between the cyclic pulses of the input signal, i.e. a phase other than the cyclic pulse phase which is a phase at and around the times when the cyclic pulses occur. The extra pulse phase determining means can, for example, determine the cyclical pulse phase by determining an interval of the cyclical pulses and determining the cyclical pulse phase as a phase which lasts for a predetermined fraction of the determined interval and is temporally related to the times of the cyclical pulses, and can determine the extra pulse phase as a phase that is mutually exclusive with the cyclic pulse phase. A temporal relationship between the cyclical pulse phase and / or the extra pulse phase and the times of the cyclical pulses can be determined or calibrated in advance based on clinical data. For example, the extra pulse phase can be centered between the times of the cyclic pulses.

The extra pulse phase suppressing means of the model detector according to the present embodiment can suppress the cyclic model error signal during times outside of the extra pulse phase by suppressing the cyclic model error signal during the cyclic pulse phase. This means that all pulses occurring in the vicinity of the cyclic pulses during the phase of the cyclic pulse, such as jittered cyclic pulses or noise, cannot be recorded as an extra pulse. As a result, the medical detector according to the present embodiment can advantageously reduce the number of false positive detections of the extra pulse, as a result of which the accuracy of the detection of an extra pulse is further improved.

According to a further embodiment of the medical detector, the extra pulse phase determining means is configured such that it determines the extra pulse phase by adapting a parametric pulse shape model to the average cyclic pulse shape of the cyclic pulse train component of the input signal.

For example, the parametric pulse shape model can be an inverted parabola, a Chen model, a Rosenberg model, a Liljencrants-Fant model, a Fujisaki-Ljungqvist model, a Klatt model, a weighted sum of sine components, a Schoentgen model Be a model or a Rank Kubin model. The adaptation can be done e.g. by least squares approximation.

For example, the extra pulse phase determining means may have respective zeros or one

determine any other predetermined level of the parametric pulse shape model as the beginning and end of the cyclical pulse phase and can determine the extra pulse phase as the phase outside the cyclical pulse phase or as the phase mutually exclusive with the cyclical pulse phase.

Alternatively, the cyclic pulse phase can be calculated from parameters of the parametric pulse shape model, which offers the advantage that a predetermined level, such as the zeros, does not have to be required.

By determining the extra pulse phase on the basis of an adapted parametric pulse shape model, the extra pulse phase determination means can advantageously determine the extra pulse phase exclusively on the basis of data that are intrinsic in the input signal and / or the cyclic model error signal or can be derived therefrom without a predetermined one Fraction of an interval needs to be set or calibrated in advance.

According to a further embodiment of the medical detector, the extra pulse train synthesizing means may be configured to synthesize the extra pulse train component of the input signal using Fourier synthesis in a manner similar to the operation of the cyclic pulse train synthesizing means.

According to a further embodiment of the medical detector, the extra pulse train synthesizing means is configured to synthesize the extra pulse train component of the input signal by convolution of the extra unit pulse train and the estimated average extra pulse shape.

The convolution operation can be performed directly in the time domain. As a result, the medical detector can advantageously avoid a conversion into the frequency domain, as a result of which expensive and complex processes, such as, for example, Fourier analysis and Fourier synthesis, are eliminated.

According to a further embodiment of the medical detector, the cyclic pulse train estimation unit further comprises a first crosstalk pulse shape estimating means configured to estimate a first crosstalk pulse shape of a first crosstalk component by cross-correlation between the cyclic unit pulse train and the extra pulse train component; first crosstalk synthesizing means configured to synthesize the first crosstalk component based on the cyclic unit pulse train and the first crosstalk pulse shape; and a first crosstalk subtracting unit configured to subtract the first crosstalk component from an output of the cyclic pulse train synthesizing means, which is a distorted cyclic pulse train component, to obtain the cyclic pulse train component.

The first crosstalk pulse shape estimating means, the first crosstalk synthesizing means and the first crosstalk subtracting unit can also be collectively referred to as "first crosstalk compensation means".

With regard to the advantages of the present embodiment, it is noted that the medical detector according to the first aspect can be described as operating on the assumption that the cyclic pulses and the extra pulses are not correlated. As described above for certain embodiments, the cyclic pulse shape estimating means can, for example, estimate the average cyclic pulse shape by cross-correlation between the input signal and the cyclic unit pulse train, whereby artifacts of non-cyclic pulses such as the extra pulse are averaged out. The assumption here is that the extra pulse train component is not correlated with the cyclic unit pulse train. In the case of input signals that are representative of oscillating medical readings, however, there may be crosstalk between the cyclical unit pulse train and the extra pulse train component, i.e. there may be a certain correlation between the extra pulse train component and the cyclical unit pulse train. Therefore, the cyclic pulse shape estimating means may erroneously estimate a distorted average cyclic pulse shape which

is distorted to some extent by the influence of crosstalk between the unit cyclic pulse train and the extra pulse train component (i.e., the average cyclic pulse shape can be influenced to some extent by the shape of the extra pulse). Consequently, the cyclic pulse train component output from the cyclic pulse train synthesizing means can similarly be influenced, i.e., corrupted or distorted, by the extra pulse train component. The amount of falsification, influencing or distortion of the distorted cyclic pulse train component, which is due to crosstalk between the cyclic unit pulse train and the extra pulse train component, or alternatively an estimate or a model thereof, are both referred to as the first cross-talk component.

Therefore, according to the preset embodiment, the medical detector comprising the above-described first crosstalk compensation means may advantageously be able to synthesize the (non-distorted) cyclic pulse train component with higher accuracy by calculating by subtracting the (estimated or modeled) first crosstalk component of the distorted cyclic pulse train component compensates for effects caused by crosstalk between the cyclic unit pulse train and the extra pulse train component.

It is noted that the undistorted cyclic pulse train component output from the first crosstalk subtracting means (first crosstalk compensation means) is input to the cyclic pulse train subtracting unit to obtain the cyclic model error signal which is input to the extra pulse train estimating unit and the extra pulse train subtracting unit to obtain the extra pulse train component, which in turn is used as an input to the first crosstalk canceller. Therefore, the first crosstalk compensation means can be configured to operate iteratively repeatedly, e.g., as a feedback loop, and can be configured to gradually approach an optimal solution.

According to a further embodiment of the medical detector, the extra pulse train estimation unit further comprises a second crosstalk pulse shape estimating means configured to estimate a second crosstalk pulse shape of a second crosstalk component by cross-correlation between the extra unit pulse train and the cyclic pulse train component; a second crosstalk synthesizing means configured to synthesize the second crosstalk component based on the extra unit pulse train and the second crosstalk pulse shape; and a second crosstalk subtracting unit configured to subtract the second crosstalk component from an output of the extra pulse train synthesizing means which is a distorted extra pulse train component to obtain the cyclic pulse train component.

The second crosstalk pulse shape estimating means, the second crosstalk synthesizing means and the second crosstalk subtracting unit can also be collectively referred to as "second crosstalk compensation means".

The features and effects which have been described for the embodiment of the medical detector with the first crosstalk compensation means also apply mutatis mutandis to the medical detector according to the present embodiment which comprises the second crosstalk compensation means. In particular, the second crosstalk compensation means can be configured to operate iteratively repeatedly, e.g. as a feedback loop. The medical detector, which comprises the second crosstalk compensation means, can advantageously be able to compensate for the effects caused by crosstalk between the extra unit pulse train and the cyclic pulse train component by subtracting the second crosstalk component from the distorted extra pulse train component, thereby increasing the accuracy of the extra pulse train component to improve.

According to a further embodiment of the medical detector, the first crosstalk pulse shape estimating means is configured such that it shows a current phase of the extra pulse

sequence component and the cyclic unit pulse train are linearized in such a way that deviations from the basic frequency are compensated for in the cyclic unit pulse train resulting from the parameter variation carried out by the parameter optimization means before the cross-correlation is carried out. That is, the first crosstalk pulse shape estimating means may be configured to match an instantaneous frequency of the unit cyclic pulse train before performing the cross-correlation.

By linearizing an instantaneous phase and / or compensating for an instantaneous frequency of the extra unit pulse train component and the cyclic unit pulse train, artifacts in the first crosstalk pulse shape caused by jitter that is introduced into the cyclic unit pulse train by parameter variation carried out by the parameter optimization means can advantageously be avoided becomes, and the accuracy of the first crosstalk pulse shape can be improved.

According to a further aspect, the invention relates to a method for detecting extra pulses, the method comprising: obtaining an input signal indicating an oscillation of a medical reading; Estimating a cyclic pulse train component of the input signal; Subtracting the cyclic pulse train component from the input signal to thereby obtain a cyclic model error signal; Detecting an extra pulse in the input signal based on the cyclic model error signal; and upon detection of the extra pulse, outputting an output signal which indicates the detected extra pulse.

The embodiments and features described with reference to the medical detector of the present invention apply mutatis mutandis to the method for detecting extra pulses of the present invention.

According to a further aspect, the invention relates to a computer program product which comprises a program code for executing the above-described method for detecting extra pulses when it is executed on at least one computing device.

The computing device can be a personal computer (PC). The computing device can also be a device that includes a digital signal processor (DSP). The computing device can also be a handheld device, a medical device, a smartphone, a Smartwaltch, a tablet PC, a laptop computer or a desktop computer.

A computer program product, such as, for example, a computer program means, can be implemented as a memory card, USB stick, CD-ROM, DVD or as a file or a file collection that can be downloaded from a server in a network. For example, such a file or file collection can be provided by transmitting the file or file collection comprising the computer program product from a wireless communication network.

Further possible implementations or alternative solutions of the invention also include combinations - which are not explicitly mentioned here - of features which have been described above or below in relation to the embodiments. Those skilled in the art can also add individual or isolated aspects and features to the most basic form of the invention.

Further embodiments, features and advantages of the present invention will become apparent from the following description and the dependent claims in conjunction with the accompanying drawings, in which:

Fig. 1 shows a block diagram of the medical detector according to a first exemplary embodiment;

Figure 2 shows a block diagram of the medical detector in accordance with a second exemplary embodiment;

Figure 3 shows a block diagram of the medical detector in accordance with a third exemplary embodiment;

Figure 4 shows a block diagram of the medical detector in accordance with a fourth exemplary embodiment;

Fig. 5 shows a block diagram of the cyclic pulse train estimation unit of the medical detector according to the fourth exemplary embodiment;

Fig. 6 shows a block diagram of the extra pulse sequence estimation unit of the medical detector according to the fourth exemplary embodiment;

Fig. 7 is a graph showing a clinically observed input signal, an estimated cyclic pulse train component of the input signal, and an estimated extra pulse train component of the input signal; and

Fig. 8 shows a flow chart illustrating steps of a method for detecting an extra pulse in accordance with the first exemplary embodiment.

In the figures, the same reference symbols denote the same or functionally equivalent elements, unless otherwise specified.

1 shows a block diagram of the medical detector 1 according to a first exemplary embodiment. The medical detector 1 comprises an input unit 10, a cyclic pulse sequence estimation unit 20, a cyclic pulse sequence subtraction unit 30 and an extra pulse detector unit 6. The input unit 10 is connected to the cyclic pulse sequence estimation unit 20 and the cyclic pulse sequence subtraction unit 30. The cyclic pulse train estimation unit 20 is connected to the input unit 10 and the cyclic pulse train subtraction unit 30. The cyclic pulse train subtraction unit 30 is also connected to the extra pulse detector unit 6.

The operation of the medical detector 1 will now be described in connection with Figs. 1 and 8, wherein Fig. 8 shows a flow chart illustrating steps of a method for detecting extra pulses according to the first exemplary embodiment.

In step S1, the input unit 10 receives an input signal d. In the present exemplary embodiment, the input signal d is representative of a clinically obtained phonation signal indicative of a vibration of a glottic region of a human vocal apparatus that has been digitally sampled or resampled at a sampling rate of 50 kHz. The input unit 10 receives the input signal d by receiving it through an input terminal (not shown) connected to an external device (not shown) such as a computer on which the digital sample of the phonation signal is stored or a detection device that supports the Live phonation signal captured and sampled.

The signal values of the input signal d indicate an opening amount of the glottic area. Under normal conditions, the input signal d is expected to contain a series of cyclic pulses that represent normal phonation pulses and occur when the glottis is open. Under normal conditions, it is expected that the input signal d does not contain any pulses in a phase that is between the cyclic pulses where the Gilottis region should be in a closed phase. Under certain conditions, however, the input signal d can contain so-called extra pulses during the closed phase (also referred to as the quasi closed phase or extra pulse phase). The presence of such an extra pulse can indicate a physical condition, i.e. the emission of additional phonation pulses during a quasi-closed phase of the glottic area, which can be a manifestation, a symptom or a sign of a voice disorder, such as diplophony. It is therefore desirable to automatically detect the presence or absence of an extra pulse in the input signal d.

The input unit 10 supplies the input signal d to the cyclic pulse train estimation unit 20 and to the cyclic pulse train subtracter 30.

In step S2, the cyclic pulse train estimation unit 20 estimates a cyclic pulse train component c1 of the input signal d. The estimated cyclic pulse train component c1 of the input signal d can be understood as a model which models the cyclic component of the phonation signal of the human tuner.

In order to estimate the cyclic pulse train component c1, the cyclic pulse train estimation unit 20 extracts a basic frequency f0 from the input signal d and synthesizes the cyclic pulse train component c1 of the input signal d as a series of cyclic pulses which occur at the basic frequency f0. The cyclic pulse train estimation unit 20 then supplies the cyclic pulse train component c1 to the cyclic pulse train subtraction unit 30.

In step S3, the cyclic pulse train subtraction unit 30 receives the input signal d and the cyclic pulse train component c1 as input signals, subtracts the cyclic pulse train component c1 from the input signal d to obtain the cyclic model error signal e1, and delivers the cyclic model error signal e1 to the extra.

The cyclic model error signal e1 is therefore the remainder of the subtraction of the cyclic pulse train component c1 from the input signal d. That is to say, the cyclic model error signal e1 only contains signal components which are not precisely modeled by the cyclic pulse train component c1. Such signal components can be noise, for example, but can also include an extra pulse that is not part of the modeled cyclic pulse train component c1. In other words, the cyclic model error signal e1 does not contain any cyclic pulses, but can contain an extra pulse.

In step S4, the extra pulse detector unit 6 detects whether the cyclic model error signal e1 contains an extra pulse. The extra pulse detector unit 6 uses peak picking to detect the extra pulse. In particular, the extra pulse detector unit 6 checks whether the cyclic model error signal e1 contains a peak signal value which is higher than a predetermined first threshold, or whether the cyclic model error signal e1 contains a peak signal value which is more pronounced or has a higher prominence than a predetermined second threshold value. Prominence, as used herein, refers to the difference between the relative or local maximum signal value at the peak and an adjacent relative or local minimum signal value divided by an average signal value.

If the extra pulse detector unit 6 detects the presence of an extra pulse in the cyclic model error signal e1, the extra pulse detector unit 6 delivers an output signal in step S5 which indicates the extra pulse detected.

The medical detector 1 according to the first exemplary embodiment uses a waveform-based modeling approach to modeling a cyclic phonation signal of a human speech device and has the advantage that it can automatically detect an extra pulse present in the phonation signal based on the model error that remains after subtracting the model from the actual input signal.

FIG. 2 shows a block diagram of a more complex variant of the medical detector 1 according to a second exemplary embodiment.

The medical detector 1 according to the second exemplary embodiment differs from the medical detector 1 according to the first exemplary embodiment in that it further comprises an extra pulse train estimation unit 40 and an extra pulse train subtraction unit 50. The extra pulse train estimation unit 40 is connected to the cyclic pulse train subtracting unit 30 and the extra pulse train subtracting unit 50. The extra pulse train subtraction unit 50 is connected to the extra pulse train estimation unit 40, the cyclic pulse train subtraction unit 30 and an extra pulse detector unit 60. The extra pulse detector unit 60 of the medical detector 1 according to the second embodiment differs from the extra pulse detector unit 6 of the medical detector 1 according to the first embodiment in that it is also connected to the extra pulse train subtraction unit 50. Further functional differences are described below.

In operation, the cyclic pulse train subtraction unit 30 of the medical detector 1 according to the second exemplary embodiment supplies the cyclic model error signal e1 not only to the extra pulse detector unit 60, but also to the extra pulse train estimation unit 40. The extra pulse train estimation unit 40 estimates an extra pulse train component c2 of the input signal d in the cyclic model error signal e1 remains after the modeled

(estimated) cyclic pulse train component c1 has been subtracted from the input signal d. The extra pulse train component c2 can thus be understood to be a model based on waveforms which models one or more extra pulses of the phonation signal of the human vocal device.

In order to estimate the extra pulse train component c2, the extra pulse train estimation unit 40 estimates one or more extra pulse triggering times from the cyclic model error signal e1 using peak picking in a similar manner as was described for the extra pulse detector unit 6 of the first embodiment. That is, the times of the selected peaks are used as the estimated extra pulse release times. The extra pulse train estimation unit 40 then synthesizes the extra pulse train component c2 of the input signal d as a series of pulses occurring at the estimated extra pulse triggering times of the extra pulses, and supplies the extra pulse train component c2 to the extra pulse train subtracting unit 50.

The extra pulse train subtraction unit 50 receives the cyclic model error signal e1 and the extra pulse train component c2, subtracts the extra pulse train component c2 from the cyclic model error signal e1 to obtain the residual error signal e2, and delivers the residual error signal e2 to the extra pulse detector unit 60.

The residual error signal e2 is therefore the remainder of the subtraction of the extra pulse train component c2 from the cyclic model error signal e1, which in turn is the remainder of the subtraction of the cyclic pulse train component c1 from the input signal d. The residual error signal e2 therefore only contains signal components which are not modeled precisely either by the cyclic pulse train component c1 or by the extra pulse train component c2. Such signal components can be due to model errors or noise, for example.

The extra pulse detector unit 60 receives the residual error signal e2 and the cyclic model error signal e1 as inputs. The extra pulse detector unit 60 compares an error level of the cyclic model error signal e1 with an error level of the residual error signal e2. In particular, the extra pulse detector unit 60 evaluates the following formula (1):

RMS (e, (n)) 20:10 RMS (e "(n)) RMS (d (n)) NO RMS (d (n)) (1)

Here, RMS refers to the root mean square value of signal values of a respective signal. n is the time index. e: (n) represents the values of the cyclic model error signal e1, d (n) represents the values of the input signal d, e> (n) represents the values of the residual error signal e2 and a is a predetermined non-zero threshold.

In other words, the extra pulse detector unit 60 evaluates whether a relative energy content of the residual error signal e2 (decibel ratio of the root mean square value of the residual error signal e2 and the root mean square value of the input signal d) is smaller than a predetermined non-zero threshold a by at least one Energy content of the cyclic model error signal e1 (decibel ratio of the root mean square value of the cyclic model error signal e1 and the root mean square value of the input signal d).

If the result of this comparison is true, this can be an indication that a combined model, which includes both the cyclic pulse train component c1 and the extra pulse train component c2, is a better model for the input signal d than just the cyclic pulse train component c1 comprehensive cyclic model. If the result of the comparison is true, the extra pulse detector unit 60 of the second exemplary embodiment detects the presence of the extra pulse and, in a similar manner to the extra pulse detector unit 6 of the first exemplary embodiment, outputs an output signal o which indicates the detected extra pulse or the detected extra pulses.

It is noted that in the above comparison the predetermined non-zero threshold value a is used to distinguish between a case in which the error level of the combined model is due solely to the presence of an additional degree of freedom.

that in the combined model is slightly lower than the error level of the cyclic model, and a case in which the error level of the combined model is substantially lower than the error level of the cyclic model, which cannot be explained by the additional degree of freedom alone, but an indication on an extra pulse. The predetermined non-zero threshold a can be, for example, 1 dB.

The medical detector 1 according to the second exemplary embodiment uses a waveform-based modeling approach to model both a cyclic component of a phonation signal of a human vocal apparatus (cyclic model) and an extra pulse component of the phonation signal (combined model). The medical detector 1 then compares the residual error level of the combined model with the error level of the cyclic model. This step-by-step approach offers the medical detector 1 of the second exemplary embodiment the advantage of being able to automatically detect an extra pulse present in the phonation signal with greater accuracy than the medical detector 1 of the first exemplary embodiment.

3 shows a block diagram of the medical detector 1 according to the third exemplary embodiment. The medical detector 1 of the third exemplary embodiment differs from the medical detector 1 of the second exemplary embodiment in that the cyclic pulse train subtraction unit 30 is configured to feed back the cyclic model error signal e1 to the second input i22 of the cyclic pulse train estimation unit 20. The cyclic pulse train estimation unit 20 uses the cyclic model error signal e1, which is fed back to its second input i22, for parameter optimization in order to take into account jitter that occurs in the cyclic pulses of the input signal d.

The medical detector 1 of the third exemplary embodiment also differs from the medical detector 1 of the second exemplary embodiment in that the cyclic pulse train estimation unit 20 is further configured to calculate an average cyclic pulse shape r1 from its second output 022 to the second input i42 of the extra pulse train estimation unit 40 to assist the extra pulse train estimation unit 40 in estimating an average extra pulse shape.

Further details on the above-mentioned features of the third embodiment are described in more detail below in connection with the fourth embodiment.

4 shows a block diagram of the medical detector 1 according to the fourth exemplary embodiment.

The medical detector 1 of the fourth exemplary embodiment differs from the medical detector 1 of the third exemplary embodiment in that the extra pulse train subtraction unit 50 is further configured to feed the residual error signal e2 back to a third input i23 of the cyclic pulse train estimation unit 20. The cyclic pulse train estimation unit 20 also uses the residual error signal e2 fed back to its second input i23 for parameter optimization, as will be described in more detail below.

The medical detector 1 of the fourth exemplary embodiment also differs from the medical detector 1 of the third exemplary embodiment in that the extra pulse train subtraction unit 50 is further configured to feed the residual error signal e2 back to the third input i43 of the extra pulse train estimation unit 40 in order to the extra pulse train estimation unit 40 to further assist in estimating an average extra pulse shape, as described in more detail below.

The medical detector 1 of the fourth exemplary embodiment also differs from the medical detector 1 of the third exemplary embodiment in that the extra pulse train estimation unit 40 is configured to feed back the extra pulse train component c2 to the fourth input i24 of the cyclic pulse train estimation unit 20 in order to obtain the to help cyclic pulse train estimation unit 20 to compensate for a first crosstalk component b1, which is generated by a correlation of the cyclic unit pulse train u1 with the extra pulse

sequence component c2 is caused, as described in more detail below. The first crosstalk component is a crosstalk component which results from a crosstalk of the extra pulse train component c2 in the cyclic pulse train component c1.

The medical detector 1 of the fourth exemplary embodiment further differs from the medical detector 1 of the third exemplary embodiment in that the cyclic pulse train estimation unit 20 is configured to feed the cyclic pulse train component c1 to the fourth input i44 of the extra pulse train estimation unit 40 in order to obtain the To assist extra pulse sequence estimation unit 40 in compensating for a second crosstalk component b2, as will be described in more detail below. The second crosstalk component is a crosstalk component which results from crosstalk of the cyclic pulse train component c1 in the extra pulse train component c2.

Further details will now be described in connection with the medical detector 1 of the fourth exemplary embodiment. Reference is made to FIGS. 4 to 6.

Fig. 5 shows a block diagram of the cyclic pulse train estimation unit 20 of the medical detector 1 according to the fourth exemplary embodiment. The cyclic pulse train estimating unit 20 comprises the fundamental frequency extraction means 21, the cyclic unit pulse train generator means 22, the cyclic pulse shape estimating means 23, the cyclic pulse train synthesizing means 24, the parameter optimizing means 25 and a first crosstalk compensation means, comprising a first crosstalk unit 28, a first crosstalk pulse shape estimating means 27 and a first crosstalk pulse shape estimating means 26, a first crosstalk pulse shape estimating means 26.

The cyclic pulse train estimation unit 20 receives the input signal d received from the input unit 10 at its first input 121. The input signal d is fed to the fundamental frequency extraction means 21 and the cyclic pulse shape estimation means 23 and, optionally, although not shown in the figures, to the parameter optimization means 25. The cyclic pulse train estimation unit 20 receives the cyclic model error signal e1 at its second input i22 and the residual error signal e2 at its third input i23. The cyclic model error signal e1 and the residual error signal e2 are each fed to both the parameter optimization means 25 and the fundamental frequency extraction means 21. The cyclic pulse train estimation unit 20 also receives the extra pulse train component c2 at its fourth input i24, which is fed to the first cross-talk pulse shape estimation means 26.

The fundamental frequency extraction means 21 according to the fourth exemplary embodiment is an f0 extractor. The f0 extractor 21 extracts a fundamental frequency fü from the input signal d and supplies the fundamental frequency f0 to the cyclic unit pulse train generator means 22.

The f0 extractor 21 uses the cyclic model error signal e1 and the residual error signal e2, which are supplied by the second and third inputs 122, 123 of the cyclic pulse train estimation unit 20 for error feedback. This error feedback enables the f0 extraction to be carried out in conjunction with the cyclic pulse train estimation carried out by the cyclic pulse train estimation unit 20 and the extra pulse train estimation carried out by the extra pulse train estimation unit 40. As a result, a higher level of accuracy can advantageously be achieved. In particular, the f0 extractor 21 can run through a large number of candidates of the fundamental frequency f0 and select as the fundamental frequency f the candidate of the fundamental frequency f0 which results in the smallest model error level of the cyclic model error signal e1 and / or the residual error signal e2.

The cyclic unit pulse train generator means 22 generates a cyclic unit pulse train u1 according to the following formula (2):

u, (n) => oiln - u MN] u (2)

where n is the time index, U is a unit pulse number, ö is the delta function and N + is a cycle

is length expressed as the number of time samples that can be calculated as N + 1, = fs / fo, where fs is the sampling frequency of the input signal d and fo is the fundamental frequency extracted by the f0 extractor 21.

The cyclic unit pulse train generator means 22 supplies the cyclic unit pulse train ul to the cyclic pulse shape estimating means 23, the cyclic pulse train synthesizing means 24, the first crosstalk pulse shape estimating means 26 and the first crosstalk synthesizing means 27.

The cyclic pulse shape estimating means 23 estimates an average cyclic pulse shape r1 of the cyclic pulses of the input signal d by performing normalized cross-correlation between the input signal d and the unit cyclic pulse train u1 according to the following formula (3):

1 OÖ ED LO daD (3)

[00143] Here | a pulse shape index that goes from -N; / 2 + 1 to + N + / 2 - 1, n is the discrete time index, u4 (n) is the cyclic unit pulse train and d is the input signal.

It is noted that according to a variant of the fourth embodiment, the instantaneous phases of the input signal d and the cyclic unit pulse train u1 before the cross-correlation can be linearized by interpolating the input signal d according to a time grid, the jitter of the unit pulse train u1 compensated, and by balancing the instantaneous frequency of the cyclic unit pulse train u1. This has the advantage of reducing the distortion induced by jitter when estimating the average cyclic pulse shape r1.

It should be noted that according to a further variant of the fourth embodiment, the average cyclic pulse shape r1 with a parametric pulse shape model, such as an inverted parabola, a Chen pulse model [1], a Rosenberg model [2], a Liljencrants -Fant model, a Fujisaki-Ljungqvist model, a Klatt model, can be modeled as a weighted sum of sine components, e.g. a Schoentgen model or a Rank model.

The normalized cross-correlation of the formula (3) makes it possible that cyclic pulse components of the input signal d, which are expected to correlate with the unit pulses of the cyclic unit pulse train u1, "get through", ie are summed, while other signal components that do not correlate with the cyclic unit pulse train u1, be averaged. The resulting waveform r1 can be understood as the average cyclic pulse shape r1 of the cyclic pulses contained in the input signal d. A first distortion, which is attributable to the correlation of the cyclic unit pulse train u1 with the extra pulse train component c2, can, however, also occur through the cross-correlation. The first distortion can be described as crosstalk of the extra pulse train component c2 in the cyclic pulse train component c1.

The cyclic pulse shape estimating means 23 supplies the average cyclic pulse shape r1 to the cyclic pulse train synthesizing means 24 and also to the second output 022 of the cyclic pulse train estimating unit 20.

The cyclic pulse train synthesizing means 24 synthesizes the cyclic pulse train component, i.e. the cyclic model of the cyclic pulses contained in the input signal, based on the cyclic unit pulse train u1 and the average cyclic pulse shape r1, according to the following formula (4):

P cin) = Aln) (| ao + la. Cos (p. 6 (n)) + Dp -sin (p. 6 (n))] Dp = 1

(4)

Here, ap and b, with p = 0..P, are the real and imaginary parts of the discrete Fourier transformation of the average cyclic pulse shape r; (I). 8 (n) is the instantaneous or momentary phase, which is defined, for example, as mt: [2 - WU + 1] at the pulse locations n = u: N +, the cyclic unit pulse sequence u1, and is interpolated between the pulse locations, e.g. with a spline . A (n) is an amplitude modulation function which is defined as A (n) = u {(n) at the pulse locations n = u + N + of the cyclic unit pulse train u1 and is obtained by interpolation, such as a shape-retaining cubic interpolation between the pulse locations.

That is, the cyclic pulse train synthesizing means 24 synthesizes the cyclic pulse train component c1 (n) by Fourier synthesis using the current phase of the cyclic unit pulse train u1 and the real and imaginary Fourier coefficients of the estimated average cyclic pulse shape r1.

It should be noted that since a first crosstalk attributable to the correlation of the cyclic unit pulse train u1 with the extra pulse train component c2 may have passed the cross-correlation performed by the cyclic pulse shape estimating means 23, as described above, the synthesized cyclic pulse train component c1 (n) is actually a distorted cyclic pulse train component c1b which includes the first crosstalk as a component.

In a variant of the fourth exemplary embodiment that does not include the first crosstalk compensation device, the cyclic pulse train synthesizer 24 supplies the synthesized (modeled, estimated) distorted cyclic pulse train component c1b directly to the first output 021 of the cyclic pulse train estimation unit 20.

In a variant of the fourth exemplary embodiment, which does not include the first crosstalk compensation device, the distorted cyclic pulse train component c1b is supplied to the first crosstalk subtraction unit 28, and the cyclic pulse train component c1 is obtained by subtracting a first crosstalk component b1 from the distorted cyclic pulse train component c1b. The first crosstalk component b1 is obtained as follows:

The first crosstalk pulse shape estimating means 26 estimates the pulse shape rb1 of the first crosstalk component b1, which is an estimate of the first crosstalk contained in the distorted cyclic pulse train component c1b, and is associated with the correlation of the cyclic unit pulse train u1 and the extra pulse train component c2 . In particular, the first crosstalk pulse shape estimating means 26 correlates the cyclic unit pulse train u1 with the extra pulse train component c2 in a manner similar to the working mode of the cyclic pulse shape estimating means 23 to obtain a first crosstalk pulse shape rb1 and supplies the first crosstalk pulse shape rb1 to the first crosstalk synthesizing means 27. According to an optional variant, the current phases of the extra pulse train component c2 and the unit pulse train u1 can be linearized before the cross-correlation, namely by interpolating the extra pulse train component c2 according to a time grid that compensates for jitter in the cyclic unit pulse train u1, and by adapting the momentary frequency of the cyclic unit pulse train u1, which gives the advantage of reducing the jitter induced distortion in the estimation of the first crosstalk pulse shape.

The first crosstalk synthesizing means 27 synthesizes the first crosstalk component b1 by Fourier synthesis using the current phase of the cyclic unit pulse train u1 and the Fourier coefficients of the first crosstalk pulse shape rb1. The first crosstalk synthesizing means 27 supplies the resulting synthesized (modeled, estimated) first crosstalk component b1 to the first crosstalk subtracting unit 28.

The first crosstalk subtracting unit 28 subtracts the first crosstalk component b1 from the distorted cyclic pulse train component c1b and feeds the (undistorted) cyclic pulse train component c1 to the first output 021 of the cyclic pulse train.

estimation unit 20 too.

Similar to the first, second and third exemplary embodiment, the cyclic pulse train subtraction unit 30 receives the cyclic pulse train component c1 from the first output 021 of the cyclic pulse train estimation unit 20 and subtracts the cyclic pulse train component c1 from the input signal d in order to obtain the cyclic model error signal e1, which is fed to the first input i41 of the extra pulse sequence estimation unit 40.

According to the fourth exemplary embodiment, the cyclic model error signal e1 is also fed back to the second input i22 of the cyclic pulse train estimation unit 20, from where it is fed to the fundamental frequency extraction means 21 and the parameter optimization means 25. The parameter optimization means 25 performs a parameter optimization by varying an amplitude modulation parameter s (u) and a jitter parameter j (uW) for each of the unit pulses of the cyclic unit pulse train u1 within a respective predetermined parameter range. The parameter optimization means 25 supplies the amplitude modulation parameters s (uW) and the jitter parameters j (uW) to the cyclic unit pulse train generator means 22. Upon receipt of the amplitude modulation parameters s (uW) and the jitter parameters j (uW), the cyclic unit pulse train generator means 22 generates the cyclic unit pulse train u1 according to the following modified formula (2):

u, (n) =) SW oil - HN - JG] u (2) As can be seen from formula (2 '), the amplitude modulation parameter s (uW) takes into account deviations in the amplitude of each individual pulse, and the jitter parameter j (u) takes into account deviations in the location (time of occurrence) of each unit pulse.

After the amplitude modulation parameters s (u) and the jitter parameters j (uW) are supplied to the cyclic unit pulse train generator means 22, the parameter optimization means 25 determines the relative error level of the resulting cyclic model error signal e1, which is afflicted with jitter and amplitude modulation, using the following formula ( 5):

(5)

The parameter optimization means 25 then continues to vary the amplitude modulation parameters s (W) and the jitter parameters j (W) within their respective predetermined ranges until the relative error level of the cyclic model error signal e1 is minimized. The parameter optimizing means 25 can use a method such as trial and error, a halving method, a downhill simplex method, a gradient method or a conjugate gradient method and the like to calculate a combination of the amplitude modulation parameters s (W) and the jitter parameters j (uW) to determine which minimize the error level of the cyclic model error signal e1.

In a manner similar to that described for the cyclic model error signal e1, the parameter optimization means 25 also receives the residual error signal e2 and is configured in such a way that it also takes into account the minimization of the error level of the residual error signal e2 when the amplitude modulation parameters s (u) and the jitter parameters j ( uW) can be determined. By also taking into account the residual error signal e2, which results from the extra pulse estimation carried out by the extra pulse train estimation unit 40, the accuracy of the parameter optimization can be further improved.

The medical detector 1 according to the fourth exemplary embodiment with the parameter optimization means 25 can include in the cyclic model represented by the cyclic pulse train component c1 with jitter and amplitude modulation pulses that would otherwise have been averaged out by the cyclic pulse shape estimating means 23. As a result, such jitter and amplitude modulation pulses can also be caused by the cyclic pulse train sub-

traction unit 30 can be subtracted from the input signal d and can thus be excluded from the cyclical model error signal e1. The medical detector 1 according to the third and fourth exemplary embodiments therefore offers an advantage of preventing a situation in which a jitter pulse is erroneously detected by the extra pulse detector unit 60 as an extra pulse.

Fig. 6 shows a block diagram of the extra pulse sequence estimation unit 40 of the medical detector 1 according to the fourth exemplary embodiment. The extra pulse train estimation unit 40 comprises an extra pulse phase determining means 45, an error signal suppression means 46 and an extra pulse triggering means 41, an extra unit pulse train generator means 42, an extra pulse shape estimating means 43 and an extra pulse train synthesizing means 44 and a second crosstalk compensation means, which comprises a second crosstalk sub-synthesizing means 49, a second crosstalk pulse shape estimating means 47 .

At its first input i41, the extra pulse sequence estimation unit 40 receives the cyclic model error signal e1, which is supplied by the cyclic pulse sequence subtraction unit 30, and feeds the cyclic model error signal e1 to the error signal suppression means 46. The extra pulse sequence estimation unit 40 also receives at its second input 142 the average cyclic pulse shape r1 of the cyclic pulses of the input signal d, which is supplied by the second output 022 of the cyclic pulse sequence estimation unit 20, and supplies the average cyclic pulse shape r1 to the extra pulse phase determination means 45.

The extra pulse train estimation unit 40 also receives the residual error signal e2 from the output of the extra pulse train subtraction unit 50 at its third input i43 and supplies the residual error signal e2 to the extra pulse triggering means 41. The extra pulse train estimation unit 40 also receives the cyclic pulse train component c1 from the first output 021 at its fourth input 144 cyclic pulse train estimation unit 20 and feeds the cyclic pulse train component c1 to the second cross-talk pulse shape estimation means 47.

The extra pulse phase determining means 45 determines an extra pulse interval, the extra pulse interval being a time interval in which an extra pulse is to be expected, based on the average cyclical pulse shape r1. The extra pulse phase determining means 45 determines the extra pulse phase as an interval in which the cyclic pulses contained in the estimated (modeled) cyclic pulse train component c1 do not occur. For this purpose, the extra pulse phase determination means 45 adapts a parametric pulse shape model, for example an inverted parabola, to the average cyclic pulse shape r1 using the least squares method and determines the extra pulse phase and the cyclic pulse phase from parameters of the parametric pulse shape model. For example, the earlier zero point of the inverted parabola is identified as the time when the cyclic pulse phase begins and the extra pulse phase ends, and the later zero point of the inverted parabola is identified as the time when the cyclic pulse phase ends and the extra pulse phase begins. The extra pulse phase determination means 45 then supplies the discrete time indices of the later and the earlier zeros of the inverted parabola fitted by the least squares method to the error signal suppression means 46, specifically as information which identifies the extra pulse phase.

The error signal suppressing means 46 suppresses, i.e. sets it to zero, the cyclic model error signal e1 during times outside the extra pulse phase, i.e. during the cyclic pulse phase, based on the information identifying the extra pulse phase received from the extra pulse phase determining means 45.

By muting the cyclic model error signal e1 during times outside the extra pulse phase, the extra pulse train estimation unit 40 ensures that pulses that occur in the input signal d in the immediate vicinity of the cyclic pulses of the cyclic pulse train component c1 are not detected as extra pulses. Such pulses in the immediate vicinity of the cyclic pulses can, for example, be jittered cyclic pulses, artifacts of cyclic pulses that occur, for example, as a result of cycle-to-cycle pulse shape modulations.

that are not taken into account in the model, or cyclic pulses with an unusual shape or noise. As a result, the medical detector 1 according to the fourth exemplary embodiment with the extra pulse sequence estimation unit 40 with the error signal suppression means 46 can advantageously suppress false positive detections and detect the extra pulse with higher accuracy.

The error signal suppression means 46 supplies the suppressed cyclic model error signal e1m to the extra pulse triggering means 41 and to the extra pulse shape estimating means 43.

The extra pulse triggering means 41 estimates one or more extra pulse triggering times, which are each represented by a time index n. For example, the extra pulse triggering means selects peak values in the suppressed cyclic model error signal e1m based on peak prominence and / or peak height in a manner similar to that described above with respect to the extra pulse detector unit 6 of the first exemplary embodiment. Alternatively, the extra pulse triggering means 41 estimates the extra pulse trigger times n, in an iterative process which is aimed at minimizing the residual error signal e2 or a level thereof. The extra pulse triggering means 41 supplies information on the estimated extra pulse triggering times n to the extra unit pulse train generator means 42.

The extra unit pulse train generator means 42 generates an extra unit pulse train u2, which is a series of unit pulses that are nv at each extra pulse triggering time.

The extra unit pulse train u2 can be represented by the following formula:

u> (n) =) oiln— nm] "(6)

where n is the discrete time index, v is an extra unit pulse number, δ is the delta function and n is the discrete time index of the v-th extra unit pulse, i.e. the extra pulse triggering time.

The unit pulse train u2 is referred to as the extra unit pulse train u2, based on the assumption that any pulse (peak) that occurs at one of the estimated extra pulse trigger times n in the suppressed cyclic model error signal e1, which is the result of subtracting the cyclic pulse train component c1 from is the input signal d and therefore probably does not contain any cyclic pulses, is an extra pulse.

The extra unit pulse train generator means 42 supplies the extra unit pulse train u2 to the extra pulse shape estimating means 43 and the extra pulse train synthesizing means 44.

The extra pulse shape estimating means 43 estimates an average extra pulse shape r2 of the extra pulses which are to form the extra pulse train component c2 of the input signal d to be estimated by the extra pulse train estimation unit 40 by performing a normalized cross-correlation operation between the cyclic model error signal e1 and the extra unit pulse train u2 according to the formula ( 7) performs:

1 (= to 2 'en -D (7)

Similar to the cross-correlation of the cyclic pulse shape estimating means 23, the cross-correlation of the extra pulse shape estimating means 43 allows extra pulses of the cyclic model error signal e1, which correlate with the extra unit pulses of the extra unit pulse train u2, to run, while the other content of the cyclic model error signal e1, such as Gaussian error signal e1 Noise and the like, is averaged out.

Optionally, the instantaneous phases of the cyclic model error signal e1 and the extra unit pulse train u2 can be linearized before the cross-correlation by interpolating the suppressed error signal e1m according to a time grid that compensates for jitter in u2, and the

Compensation of the instantaneous frequency of u2, which has the advantage of reducing the distortion caused by jitter when estimating the average extra pulse shape r2.

The extra pulse train synthesizing means 44 synthesizes an extra pulse train component c2, which may contain a distortion which is attributable to the correlation of the extra unit pulse train u2 with the cyclic pulse train component c1 based on the extra unit pulse train u2 and the average extra pulse shape r2 and is therefore also referred to as a distorted extra pulse train. according to the following formula:

can) = kun) 720m - DI L (8) That is, the extra pulse train synthesizing means 44 synthesizes the (distorted) extra pulse train component c2 (c2b) of the input signal d by convolving the extra unit pulse train u2 and the average extra pulse shape r2 directly in the time domain. This has the advantage that a computationally intensive conversion into the frequency domain and back is avoided.

The (distorted) extra pulse train component c2, c2b, which is synthesized by the extra pulse train synthesizing means 44, forms an estimated, based on a waveform model of extra pulses that are present in the cyclic model error signal e1 and thus also in the input signal d.

In a variant in which no crosstalk compensation is used, the distorted extra pulse train component c2b is fed directly as an extra pulse train component c2 to the output 041 of the extra pulse train estimation unit 40 and from there to the extra pulse train subtracter 50.

According to the fourth embodiment, however, the extra pulse train estimation unit 40 comprises the second crosstalk compensation device which is configured to obtain an (undistorted) extra pulse train component c2 by subtracting a second crosstalk component from the distorted extra pulse train component c2b, and the operation of which is described hereinafter.

That is, the second crosstalk pulse shape estimating means 47 estimates the pulse shape rb2 of the second crosstalk component b2 which is an estimate of the second crosstalk included in the distorted extra pulse train component c2b and is associated with the correlation of the extra unit pulse train u2 and the cyclic pulse train component c1. In particular, the second crosstalk pulse shape estimating means 47 correlates the extra unit pulse train u2 with the cyclic pulse train component c1 in a manner similar to the cyclic pulse shape estimating means 23 and / or the first crosstalk pulse shape estimating means 26 to obtain the second crosstalk pulse shape rb2, and leads the second crosstalk pulse shape 48 to the second crosstalk pulse shape rb2. According to an optional variant, the instantaneous phases of the cyclic pulse train component c1 and the extra unit pulse train u2 can be linearized before the cross-correlation by interpolating the cyclic pulse train component c1 according to a time grid, compensating the jitter in the extra unit pulse train u2, and by adjusting the momentary frequency of the extra unit pulse train u2 , whereby the advantage is achieved that the distortion induced by jitter is reduced in the estimation of the second cross-talk pulse shape.

The second crosstalk synthesizing means 48 synthesizes the second crosstalk component b2 by convolution of the extra unit pulse train u2 and the second crosstalk pulse shape rb2 in a manner similar to that of the extra pulse train synthesizing means 44. The second crosstalk synthesizing means 48 supplies the synthesized (modeled, estimated) second crosstalk sub-unit 49 to the second crosstalk unit 49 .

The second crosstalk subtracting unit 49 subtracts the second crosstalk component b2 from the distorted extra pulse train component c2b and feeds the resulting (undistorted) extra pulse train component c2 to the first output 041 of the extra pulse train sequence.

z unit 40 to and from there to the extra pulse sequence subtracting unit 50.

The extra pulse train subtraction unit 50 and the extra pulse detector unit 60 according to the fourth embodiment operate in a manner similar to the corresponding units of the medical detector 1 according to the second and third embodiment. That is, the medical detector 1 according to the fourth exemplary embodiment also detects an extra pulse when the error level of the combined model, which includes both the cyclic pulse train component c1 and the extra pulse train component c2, is at least a predetermined threshold of, for example, 1 dB lower than that Error level of the cyclic model that contains only the cyclic pulse train component c1.

Fig. 7 shows a diagram with a graph of a clinically observed phonation signal (input signal) d and graphs of an estimated cyclical pulse train component c1 and an estimated extra pulse train component c2 of the clinically observed input signal d, as from the respective estimation means of the medical detector 1 according to the third exemplary embodiment appreciated.

In the diagram, the x-axis denotes the time and the y-axis denotes a signal value or an amplitude. The clinically observed phonation signal d is shown as a solid line. The modeled cyclic pulse train component c1 is shown as a dashed line. The modeled extra pulse train component c2 is shown as a dotted line. The arrows labeled ex1 to ex6 denote extra pulses which form extra pulses in the clinically observed phonation signal d.

As can be seen from the diagram, there is a good correspondence between the modeled cyclic pulse train component c1 and the actual cyclic pulses of the clinically observed phonation signal d. Furthermore, the extra pulse train component c2 shows peaks which correspond to the positions of the visible extra pulses ex1 to ex6. The extra pulse train component c2 also shows a peak at the time index 1.473 which does not correspond to an extra pulse.

During the further processing of the extra pulse train subtraction unit 50 and the extra pulse detector unit 60 of the medical detector 1, the combination of signal components shown in FIG. 6 leads to six correct automatic detections of an extra pulse and to a false positive detection.

In order to deal with false positive detections, according to a modification of the medical detector 1 according to one of the embodiments, the extra pulse detector unit 6, 60 do not output the output signal o immediately after the detection of a single extra pulse, but can be configured to count extra pulses detected in the cyclic model error signal e1 and to count cyclic pulses included in the cyclic pulse train component c1. The extra pulse detector unit 6, 60 according to the modification can further be configured to output the output signal as soon as it has been detected that a ratio of the number of extra pulses detected to the number of cyclic pulses counted is higher than a predetermined detection threshold. The predetermined detection threshold can be set to a value that has been found to be clinically relevant. In one example, the predetermined threshold can be chosen to be ten percent.

In addition, all parameters can be estimated that characterize the random process of the extra pulse generation, e.g. its probability density function (PDF), its autocorrelation function (ACF), its power spectral density (PSD), parameters of an autoregressive (AR) model and its stationarity properties. In particular, the generation of extra pulses can be modeled as a Bernoulli distribution.

While the present invention has been described in accordance with preferred embodiments, it will be apparent to those skilled in the art that modifications are possible in all of the exemplary embodiments.

For example, the medical detector 1 according to the third and fourth exemplary embodiment comprises as a means for eliminating jitter, amplitude modulation and noise

24/36

between both the parameter optimization means 25 of the cyclic pulse train estimation unit 20 and the extra pulse phase determination means 45 and the error signal suppression means 46 of the extra pulse train estimation unit 40. However, the medical detector can also comprise only the parameter optimization means 25 or only the extra pulse phase determination means 45 and the error signal suppression means 46 or none of them. Likewise, the medical detector 1 according to the first or second embodiment, which has no parameter optimization means 25, can also be provided with the parameter optimization means 25.

The extra pulse train synthesizing means 44 of the extra pulse train estimation unit 40 of the medical detector 1 according to the third exemplary embodiment synthesizes the extra pulse train component c2 by convolving the extra unit pulse train u2 and the average extra pulse shape r2. However, the extra pulse train synthesizing means 44 may also synthesize the extra pulse train component c2 using Fourier synthesis in a manner similar to the operation of the cyclic pulse train synthesizing means 24.

It has been described that the f0 extractor 21 of the cyclic pulse train estimation unit 20 uses error feedback with reference to the cyclic model error signal e1 and the residual error signal e2, which are fed back to the fO extractor 21 in order to perform the fO extraction optimize. The f0 extractor 21 can, however, also carry out the f0 extraction without an error signal being fed back to the f0 extractor 21.

It has been described that the parameter optimization means 25 is configured so that it receives and minimizes both the error level of the cyclic model error signal e1 and the residual error signal e2. The parameter optimization means 25 can, however, also be configured in such a way that it only receives and minimizes either the cyclic model error signal e1 or the residual error signal e2.

The extra pulse triggering means 41 has been described primarily as a peak recording means configured to estimate the one or more extra pulse triggering times by selecting maximum values ("peak picking"). As an alternative to peak picking, however, time instances of the extra pulses (the extra pulse triggering times) can be a fixed mean phase difference to the cyclic pulses, e.g. 180 degrees, and be subject to a subsequent selection process, e.g. model error minimization, such as minimizing the error level of the residual error signal e2. As an additional option, jitter and amplitude modulation parameters of the extra pulses can also be estimated by Minimizing the error level of the residual error signal e2. In addition, extra pulses can be triggered cyclically using an extra pulse base frequency that differs from the base frequency of the cyclic pulse train component, e.g. in diplophony, when two simultaneous fundamental frequencies are detected by the cyclic pulse train estimation unit 20.

The first crosstalk compensation means and the related functional steps described in connection with the fourth exemplary embodiment are all optional and can also be omitted. The second crosstalk compensation device, which was described in connection with the fourth exemplary embodiment, is likewise optional and can also be omitted.

The respective functionalities for linearizing an instantaneous phase and / or for balancing an instantaneous frequency of the input signal d, the cyclic unit pulse train u1, the cyclic model error signal e1, the extra unit pulse train u2 and / or the extra pulse train component c2 of the fourth embodiment are each optional and can also be used individually or be omitted altogether.

In the second to fourth exemplary embodiments, the comparison of the cyclical model error signal e1 and the residual error signal e2, which is carried out by the extra pulse detector unit 60, was mainly described in such a way that it can be determined by comparing relative error levels or decibel ratios of the corresponding error signals e1, e2 in With respect to the input signal d while using a predetermined threshold as a criterion to identify that the relative error level of the residual error signal e2 sig-

is significantly lower than the relative error level of the cyclic model error signal e1 in order to guarantee the detection of an extra pulse. However, the extra pulse detector unit 60 according to one of these exemplary embodiments can also be configured to compare the cyclic model error signal e1 and the residual error signal e2 using a statistical test. For example, the extra pulse detection unit 60 may be configured to calculate a p-value of the cross correlation performed by the extra pulse shape estimating means 43. When calculating the p-value of the cross-correlation, the extra heart rate detector unit 60 can detect the extra heart rate if the p value is less than a predetermined threshold, such as 0.05.

The output signal given by the extra pulse detector unit 6, 60 has mainly been described as a mere indication of the presence of at least one extra pulse in the input signal d, such as an audio signal, an acoustic indication or a visual indication, e.g. a flashing LED. However, the output signal can also include any other information that is available to the medical detector 1. In a variant, the output signal can be set up to cause an external display device to display a complete graphic representation of the input signal d, the cyclical pulse train component c1 and the extra pulse train component c2 or any other signal or a (signal) component involved in the modeling and capture, or processed versions thereof. If the input signal is a phonation signal, such information can be helpful to a physician in assessing mechanical and aerodynamic parameters of phonation which can be controlled in clinical treatment.

An added value of the described approach compared to black box data-driven approaches, which are popular in machine learning and artificial intelligence, is the explanatory power or expressiveness of the model. In particular, the visual and auditory comparison of signals that are involved in the modeling and detection process with natural stimuli provides convincing evidence of the validity of the approach presented. For example, videos of the vocal cords can be visually compared to the signals involved in the modeling and acquisition process. With regard to sonification and auditory comparison, all signals involved in the modeling and detection can be audibly checked as to whether they correspond to certain auditory currents perceptible in the input signal by being transmitted to a human listener via an electroacoustic device, e.g. headphones or one or more loudspeakers be reproduced.

The present invention has been described primarily in relation to a human phonation signal. However, the medical detector according to the present invention may be useful in a variety of scenarios that include an input signal indicative of a vibration of a medical reading, such as a vibration of a medical reading that is a vibration of a human body part such as a heart muscle or a oscillatory movement of the limbs as a sign of a neurological disorder, e.g. Parkinson's disease.

In addition, the invention can be applied to the analysis of any signal that is subject to the underlying signal model, i.e. a summation of a cyclic pulse train, an extra pulse train and noise, wherever such a signal can occur in nature or theoretically.

For example, in the medical field, fluctuations in blood sugar can first be analyzed for the presence of extra pulses. Second, vibrations of substances in the human body can be analyzed, e.g. the detection of irregular blood flows can be relevant for the control of artificial hearts.

The invention can also be relevant for the analysis of climate data, financial data or astronomy, namely for the detection of exoplanets, in particular in connection with the transit method in which cyclical changes in the light intensity of a star with a focus on the detection of exoplanets going through the star as seen by an observer.

26/36

REFERENCE MARK

1 medical detector

6 extra pulse detector unit

10 input unit

20 cyclic pulse train estimation unit

21 Fundamental Frequency Extractants

22 cyclic unit pulse train generating means 23 cyclic pulse shape estimating means

24 cyclic pulse train synthesis means 25 parameter optimization means

26 first crosstalk pulse shape estimating means 27 first crosstalk synthesizing means 28 first crosstalk subtracting unit

30 cyclic pulse train subtraction unit

40 extra pulse rate estimator

41 Extra pulse triggering means

42 extra unit pulse train generator means

43 extra pulse shape estimator

44 Extra Pulse Synthesizers

45 extra pulse phase determining means

46 error signal suppressing means

47 second crosstalk pulse shape estimating means 48 second crosstalk synthesizing means 49 second crosstalk subtracting unit

50 extra pulse train subtraction unit

60 extra pulse detector unit

70 output unit

d input signal

c1 cyclic pulse train component

cl1b distorted cyclic pulse train component b1 first crosstalk component

c2 extra pulse train component

c2b distorted extra pulse train component

b2 second crosstalk component

e1 cyclic model error signal

el1m suppressed cyclic model error signal

e2 residual error signal

for fundamental frequency

0 output signal

no average cyclic pulse shape rn average extra pulse shape

rb1 first crosstalk pulse shape

rb2 second crosstalk pulse shape

ul cyclic unit pulse train

u2 extra unit pulse train

i21 first input of the cyclic pulse train estimation unit i22 second input of the cyclic pulse train estimation unit 021 first output of the cyclic pulse train estimation unit 022 second output of the cyclic pulse train estimation unit i41 first input of the extra pulse train estimation unit

i42 second input of the extra pulse sequence estimation unit

041 Output of the extra pulse train estimation unit

ex1 .. ex6 extra pulses

My extra pulse trigger time

28/36

Claims (9)

Claims 1. A medical detector (1) for extra pulse detection, comprising: - an input unit (10) configured to receive an input signal (d) indicative of a vibration of a medical reading; - A cyclic pulse train estimation unit (20) which is configured to estimate a cyclic pulse train component (c1) of the input signal (d); - a cyclic pulse train subtraction unit (30) configured to subtract the cyclic pulse train component (c1) from the input signal (d) to thereby obtain a cyclic model error signal (e1); - and an extra pulse detector unit (6, 60) configured to detect an extra pulse (ex) in the input signal (d) based on the cyclic model error signal (e1); and to output an output signal (0) when the extra pulse (ex) is detected, which indicates the extra pulse (ex) detected, wherein the cyclic pulse train estimation unit (20) comprises: - a fundamental frequency extraction means (21) configured to extract a fundamental frequency (f0) from the input signal (d); - a cyclic unit pulse train generator means (22) which is configured to generate a cyclic unit pulse train (u1) with the extracted fundamental frequency (f0); - a cyclic pulse shape estimating means (23) configured to estimate an average cyclic pulse shape (r1) of the cyclic pulse train component (c1) of the input signal (d) based on the input signal (d) and the cyclic unit pulse train (u1); and - A cyclic pulse train synthesizing means (24) which is configured to synthesize the cyclic pulse train component (c1) of the input signal (d), which is generated by the cyclic pulse train estimation unit (20) based on the cyclic unit pulse train (u1) and the average cyclic pulse shape (r1 ) should be estimated; wherein the medical detector further comprises: - an extra pulse train estimation unit (40) configured to estimate an extra pulse train component (c2) of the input signal (d) based on the cyclic model error signal (e1); and - an extra pulse train subtracting unit (50) which is configured to subtract the extra pulse train component (c2) from the cyclic model error signal (e1) to thereby obtain a residual error signal (e2); and - wherein the extra pulse detector unit (60) is configured to detect the extra pulse (ex) by comparing the cyclic model error signal (e1) with the residual error signal (e2); wherein the extra pulse sequence estimation unit (40) further comprises: - an extra pulse triggering means (41) which is configured to estimate one or more extra pulse trigger times from the cyclical model error signal (e1); - an extra unit pulse train generator means (42) which is configured to generate an extra unit pulse train (u2) comprising an extra unit pulse at each estimated extra pulse triggering time; - an extra pulse shape estimating means (43) configured to estimate an average extra pulse shape (r2) of the extra pulse train component (c2) of the input signal (d) based on the cyclic model error signal (e1) and the extra unit pulse train (u2); and - An extra pulse train synthesizing means (44) designed to synthesize the extra pulse train component (c2) of the input signal (d), the extra pulse train component (c2) being estimated by the extra pulse train estimation unit (40) based on the extra unit pulse train (u2) and the average extra pulse shape (r2) shall be; 29/36 wherein the cyclic pulse train estimation unit (20) further comprises: - a first crosstalk pulse shape estimating means (26) configured to estimate a first crosstalk pulse shape (rb1) of a first crosstalk component (b1) by cross-correlation between the cyclic unit pulse train (u1) and the extra pulse train component (c2); - a first crosstalk synthesizing means (27) configured to synthesize the first crosstalk component (b1) based on the cyclic unit pulse train (u1) and the first crosstalk pulse shape (rb1); and - a first crosstalk subtracting unit (28) which is subtracted to subtract the first crosstalk component (b1) from an output of the cyclic pulse train synthesizing means (24), which is a distorted cyclic pulse train component (c1b), in order to obtain the cyclic pulse train component (c1), and or - the extra pulse sequence estimation unit (40) further comprises: a second crosstalk pulse shape estimating means (47) configured to estimate a second crosstalk pulse shape (rb2) of a second crosstalk component (b2) by cross-correlation between the extra-unit pulse train (u2) and the cyclic pulse train component (c1); - a second crosstalk synthesizing means (48) configured to synthesize the second crosstalk component (b2) based on the extra unit pulse train (u2) and the second crosstalk pulse shape (rb2); and - a second crosstalk subtracting unit (49) configured to subtract the second crosstalk component (b2) from an output of the extra pulse train synthesizing means (44) which is a distorted extra pulse train component (c2b) to obtain the extra pulse train component (c2). 2. The medical detector according to claim 1, wherein the cyclic pulse shape estimating means (23) is configured to calculate the average cyclic pulse shape (r1) of the cyclic pulse train component (c1) of the input signal (d) by cross-correlation between the input signal (d) and the cyclic unit pulse train ( u1) to appreciate. 3. Medical detector according to claim 1 or 2, wherein the cyclic pulse train synthesizing means (24) is configured to generate the cyclic pulse train component (c1) of the input signal (d) by means of Fourier synthesis using an instantaneous phase of the cyclic unit pulse train (u1) and Fourier Synthesize coefficients of the estimated average cyclic pulse shape (r1). 4. The medical detector according to any one of claims 1 to 3, wherein the cyclic pulse train estimation unit (20) further comprises: - A parameter optimization means (25) which is designed to minimize the cyclic model error signal (e1) by means of a parameter variation of an amplitude and a deviation from the fundamental frequency (fO) for each of a plurality of pulses of the cyclic unit pulse train (u1). 5. The medical detector according to claim 4, wherein the cyclic pulse shape estimating means (23) is further configured to linearize an instantaneous phase of the input signal (d) and the cyclic unit pulse train (u1) in a manner to avoid deviations from the fundamental frequency (f0) in to compensate for the cyclic unit pulse train (u1), which results from the parameter variation carried out by the parameter optimization means (25), before the average cyclic pulse shape (41) of the cyclic pulse train component (c1) of the input signal (d) is estimated. 30/36 6. The medical detector of claim 1, wherein the extra pulse detector unit (60) is configured to: - Determining the error level of the cyclic model error signal (e1) as a decibel ratio of a root mean square value of the cyclic model error signal (e1) and a root mean square value of the input signal (d); - Determining the error level of the residual error signal (e2) as a decibel ratio of a root mean square value of the residual error signal (e2) and the root mean square value of the input signal (d); and - Determination of the extra pulse (ex) when the error level of the residual error signal (e2) is smaller than the error level of the cyclic model error signal (e1) by at least a predetermined threshold. 7. The medical detector according to claim 1, wherein the extra pulse shape estimating means (43) is configured to the average extra pulse shape (r2) of the extra pulse train component (c2) of the input signal (d) by cross-correlation between the cyclic model error signal (e1) and the extra unit pulse train (u2) estimate. The medical detector of claim 1, wherein the extra pulse train estimation unit (40) further comprises: - an extra pulse phase determining means (45) configured to determine an extra pulse phase during which the extra pulse is to be expected; and - an error signal suppression means (46) which is configured in such a way that it derives a suppressed cyclic model error signal (e1m) from the cyclic model error signal (e1) by muting the cyclic model error signal (e1) during the times outside the extra pulse phase, - and wherein the extra pulse triggering means (41) is configured to estimate the one or more extra pulse trigger times from the suppressed cyclic model error signal (e1m). 9. The medical detector according to claim 8, wherein the extra pulse phase determining means (45) is configured to determine the extra pulse phase by adapting a parametric pulse shape model to the average cyclic pulse shape (r1) of the cyclic pulse train component (c1) of the input signal (d). In addition 5 sheets of drawings