EP0644527A2 - Terminal for mobile radio - Google Patents

Abstract

The invention relates to a mobile radio terminal with a speech processing device for processing speech signals (x(i)) comprising noise and speech signal components. An estimating device (7) for continuous formation of estimated values (SNR(i)) of the signal-to-noise power ratio of the speech signals (x(i)) is provided through means for - determining the power values of speech signal sample values, - smoothing the power values, - determining in each case the minimum of a group of L consecutive smoothed power values (Px,s(i)), the groups following on continuously from one another and containing at least enough smoothed power values (Px,s(i)) to ensure that in each case all smoothed power values (Px,s(i)) associated with any given phoneme of the speech signal (x(i)) can be captured by a single group, - forming a current estimated value (SNR(i)) of the signal-to-noise power ratio from the current smoothed power value and the last determined minimum. <IMAGE>

Description

The invention relates to a mobile radio terminal with a speech processing device for processing speech signals consisting of noise and speech signal components.

In the field of speech processing, noise signals are often contained in speech signals to be processed, which leads to a reduction in the speech quality and thus in particular to a deteriorated speech intelligibility. This problem occurs in particular in the case of mobile radio terminals which are used in motor vehicles and have a hands-free device. Speech signals received by microphones of the hands-free device arranged in the motor vehicle contain, on the one hand, speech signal components which are generated by the respective user (speech source) of the mobile radio terminal within the motor vehicle and, on the other hand, noise signal components which result from other ambient noises of the mobile radio device. During a journey, the ambient noise essentially consists of engine and driving noises.

From "Proceedings of the IEEE, VOL.75, No. 2, February 1987" a device with several microphones is known, in which, with the exception of one microphone signal, all other delay signals that can be set are supplied to adjustable delay elements. The microphone signals offset in time with the aid of the delay elements are added up and then subjected to postprocessing. The useful signal components of the microphone signals essentially originate from a single acoustic source which is at different distances from the microphones.

This results in different transit times to the spatially separated microphones for an acoustic signal generated by the acoustic source. The acoustic signal causes time-delayed and otherwise approximately the same useful signal components of the microphone signals. The useful signal components are thus strongly correlated. Noise signal components of the microphone signals are at most weakly correlated with a suitable arrangement of the microphones. The output signal of the device or its signal / noise ratio is improved by suitably setting the delay elements as a function of the position of the acoustic source.

Such a device provides satisfactory results only when the signal-to-noise ratio of the microphone signals to be processed is above a threshold, i.e. the useful signal components must be sufficiently large compared to the noise signal components. In particular, the noise signal components must not be larger than the useful signal components. For this reason, an estimate of the respective signal / noise power ratio must be available from at least one microphone signal each time the delay elements are reset, in order to be able to avoid malfunctions of the speech processing device when the signal / noise power ratio is insufficient.

Previous devices for determining the signal / noise power ratio of a voice signal consisting of noise and voice signal components determine a value for the noise signal power in each case in a speech pause, where only noise signal components are present. The detection of a speech pause is based, for example, on a statistical evaluation of the speech signal using histograms or the evaluation of the short-term power of the noisy speech signal.

Such a determination of the signal / noise power ratio as a function of speech pauses is, on the one hand, susceptible to faults due to the necessary speech pause detection and, on the other hand, slow, since the signal / noise power ratio can only be updated when a speech pause occurs, and the power of the noise signal component may have changed between the speech pauses .

The invention is therefore based on the object of specifying a mobile radio terminal with a speech processing device of the type mentioned in the introduction, in which an estimate of the signal / noise power ratio of the speech signals is improved.

• Ermittlung der Leistungswerte von Abtastwerten der Sprachsignale,
• Glättung der Leistungswerte,
• Ermittlung jeweils des Minimums einer Gruppe von L aufeinanderfolgenden geglätteten Leistungswerten, wobei die Gruppen lückenlos aufeinanderfolgen und mindestens so viele geglättete Leistungswerte enthalten, daß jeweils alle einem beliebigen Phonem des Sprachsignals zugehörigen geglätteten Leistungswerte von einer einzigen Gruppe erfaßbar sind,
• Bildung eines aktuellen Schätzwertes des Signal-/ Rauschleistungsverhältnisses aus dem aktuellen geglätteten Leistungswert und dem zuletzt ermittelten Minimum enthält.

The object is achieved in that the speech processing device for processing speech signals consisting of noise and speech signal components has an estimation device for continuously forming estimates of the signal / noise power ratio of the speech signals by means of

• Determination of the power values of samples of the speech signals,
• Smoothing the power values,
• Determination of the minimum of a group of L successive smoothed power values, the groups successively following one another and containing at least so many smoothed power values that all smoothed power values associated with any phoneme of the speech signal can be detected by a single group,
• Formation of a current estimated value of the signal / noise power ratio from the current smoothed power value and the last determined minimum contains.

The course of the smoothed power values of speech signals consisting of noise and speech signal components shows peaks between two speech pauses (for example the pauses between two words), ie areas of briefly high power, between which areas of lower power lie. The smoothed power values between the peaks are used to estimate the noise signal power. At least one peak of the curve of the smoothed power values is assigned to a phoneme of a speech signal. A phoneme is the smallest meaning-distinguishing unit of language and a sound that is formed on the one hand by vowels or on the other hand by single or several consonants. If the groups with L successive smoothed power values are so large that any phoneme and thus also any peak in the course of the smoothed power values can be completely detected, it is ensured that at least one value of a region of lower conduction lying next to a peak can be detected by each group . This avoids the case that a group contains only smoothed power values belonging to a peak. The minimum of a group can thus be used to estimate the noise signal power. The scaling factor serves to improve the estimate. The groups can be adjacent to each other or partially overlap. In the event that the groups are adjacent to one another, the minimum interval between two updates of the weighted minimum L used to estimate the noise signal power is sampling intervals of the speech signals. If the groups overlap so that at least one smoothed performance value belongs to several groups, the minimum period between two updates of the weighted minimum can be reduced. Through the ongoing of The speech processing device can also adapt to changes in the noise signal power between two speech pauses, independently of speech pauses, by forming estimates of the signal / noise power ratio. No speech pause is required to update the noise signal power estimate.

aufeinanderfolgenden geglätteten Leistungswerten und zur Ermittlung des Minimums der Minima von jeweils W aufeinanderfolgenden Untergruppen zur Ermittlung des Minimums der zugehörigen Gruppe vorgesehen, wobei W eine natürliche Zahl darstellt und W Untergruppen eine Gruppe bilden.In one embodiment of the invention, means for forming adjoining subgroups are provided $\text{M = L / W}$

successive smoothed power values and for determining the minimum of the minima of W successive subgroups for determining the minimum of the associated group, where W represents a natural number and W subgroups form a group.

Both adjacent and overlapping groups can be realized with little effort. In the case of adjacent groups, a new estimated value for the noise signal power is determined after every L sampling intervals from the minimum of the minima of W successive subgroups. In the case of overlapping groups, the noise signal power is re-estimated by the minimum of the minima of W consecutive sub-groups after M sampling intervals.

In this embodiment, means for using the last determined minimum of a subgroup instead of the last determined minimum of a group with a predeterminable number of monotonically increasing minima of subgroups can also be provided in order to estimate a current value of the signal / noise power ratio.

In this way, an estimate of the noise signal power is updated after M sampling intervals, with only M past smoothed power values being included in the estimate. With the help of the updates of the estimated values of the noise signal power, which are thus faster and better adapted to the course of the smoothed power values, improved estimated values of the signal / noise power ratio of the speech signals result.

The invention can be further embodied in that means are provided for using the current smoothed power value instead of a recently determined minimum of a group or subgroup for estimating a current value of the signal / noise power ratio in the event that the current smoothed power value is less than the last one determined minimum is.

Regardless of the size and arrangement of groups or subgroups, the minimum determined last is immediately replaced by the current smoothed power value if the power values are correspondingly small. In this case, there is an instantaneous update of an estimate of the noise signal power by the current smoothed power value.

In another embodiment, speech processing means are provided for processing the speech signals as a function of estimated values of the signal / noise power ratio.

It is prevented that the speech processing means operate incorrectly when the signal / noise power ratio of the speech signals to be processed is inadequate and in particular deliver output signals, their speech quality is very low. For example, if the signal-to-noise power ratio is too low, the settings of the speech processing means previously determined, ie if the signal-to-noise power ratio is sufficiently high, can be kept constant until there is again a sufficiently high signal-to-noise power ratio.

Fig. 1

eine Sprachverarbeitungsvorrichtung für zwei Sprachsignale,

Fig. 2

eine Steuervorrichtung zur Einstellung eines Zeitversatzes zwischen den beiden Sprachsignalen nach Fig. 1,

Fig. 3

eine Sprachverarbeitungsvorrichtung für drei Sprachsignale,

Fig. 4 und 5

Blockschaltbilder mit Steuervorrichtungen zur Einstellung von Zeitversätzen zwischen den drei Sprachsignalen nach Fig. 3,

Fig. 6 und 7

ein Blockschaltbild und ein Flußdiagramm zur Bestimmung des Signal-/ Rauschleistungsverhältnisses eines Sprachsignals,

Fig. 8

eine Einteilung von geglätteten Leistungswerten eines Sprachsignals in Gruppen und Untergruppen und

Fig. 9

ein Mobilfunkendgerät mit einer Sprachverarbeitungsvorrichtung nach Fig. 1 bis 8.

Embodiments of the invention are explained below with reference to the drawings. Show it:

Fig. 1

a speech processing device for two speech signals,

Fig. 2

1 a control device for setting a time offset between the two voice signals according to FIG. 1,

Fig. 3

a speech processing device for three speech signals,

4 and 5

3 block diagrams with control devices for setting time offsets between the three speech signals,

6 and 7

2 shows a block diagram and a flowchart for determining the signal / noise power ratio of a speech signal,

Fig. 8

a division of smoothed power values of a speech signal into groups and subgroups and

Fig. 9

a mobile radio terminal with a voice processing device according to FIGS. 1 to 8.

The speech processing device shown in FIG. 1 contains two microphones M1 and M2. These are used to convert acoustic to electrical voice signals, which are made up of speech and noise signal components. The speech signal components come from a single one Speech source (speaker), which is usually at different distances from the two microphones M1 and M2. The speech signal components are thus highly correlated. The noise signal components of the two speech signals received by the microphones M1 and M2 are ambient noises generated by the individual speech source, which can be assumed to be uncorrelated or only slightly correlated with suitable microphone spacings in the range from 10 to 60 cm if the microphones reverberated in a so-called Environment such as in the car or in an office. If the speech source and speech processing device are located in a motor vehicle, for example, the noise signal components are caused in particular by engine and driving noises.

The microphone signals generated by the microphones M1 and M2 are digitized by analog-digital converters 1 and 2. The resulting digitized and thus present as samples x1 (i) and x2 (i) microphone signals are evaluated by a control device 3, which is used to control and set a delay element 4. The sampled microphone signals x1 (i) and x2 (i) are referred to below as microphone or speech signals. The delay element 4 delays the microphone signal x1 with delay values T1 that can be set by the control device 3. An adding device 5 adds the microphone signal x1 (i) delayed by the delay element 4 and the microphone signal x2 (i) delayed by a delay element 16 with a constant time delay T _max . The delay element 16 is provided in order to be able to set both a leading and a lagging of the microphone signal x1 (i) relative to the microphone signal x2 (i). A sum signal X (i) present at the output of the adder 5 is a sampled speech signal, the signal / noise power ratio of which compared to the signal / noise power ratios of the speech signals x1 (i) and x2 (i) is increased. Through a suitable setting of the delay time T1 of the delay element 4, the addition by the adder 5 increases the power of the voice signal components of the two voice signals x1 (i) and x2 (i) by approximately a factor of 4 and increases the power of the noise signal components only approximately caused by a factor of 2. This results in an improvement in the power-related signal / noise power ratio of approximately 3 dB.

Die Sprachsignalschätzwerte x1_int(i) sind Werte, die sich aus einer Interpolation von Abtastwerten des Sprachsignals x1(i) ergeben. Die Bestimmung der Sprachsignalschätzwerte x1_int(i) wird später erläutert. i ist eine Variable, die ganzzahlige Werte annehmen kann und mit der einerseits Abtastzeitpunkte der Sprachsignale x1(i) und x2(i) und andererseits auch Programmzyklen der programmierbaren und Steuermittel aufweisenden Steuervorrichtung 3 indiziert werden, wobei in einem Programmzyklus jeweils ein neuer Abtastwert per Sprachsignal verarbeitet wird.In Fig. 2 the operation of the control device 3 is explained in more detail using a block diagram. From the speech signal x2 (i) and speech signal estimated values x1 _int (i), error values e 1 (i) result from difference formation

${\text{e₁₂ (i) = x1}}_{\text{int}}\text{(i) - x2 (i) (1)}$

The speech signal estimates x1 _int (i) are values that result from an interpolation of samples of the speech signal x1 (i). The determination of the speech signal estimates x1 _int (i) will be explained later. i is a variable which can take integer values and with which, on the one hand, sampling times of the speech signals x1 (i) and x2 (i) and, on the other hand, also program cycles of the programmable control device 3 having control means 3, are indicated, with one new sample value per speech signal in each program cycle is processed.

Das die Werte x2_H(i) von x2(i) liefernde Digitalfilter 6 ist ein FIR-Filter der Ordnung K, das Koeffizienten h(0), h(1), ..., h(K) aufweist. Im vorliegenden Ausführungsbeispiel ist K gleich sechzehn, so daß das Digitalfilter 6 siebzehn Koeffizienten aufweist. Das Digitalfilter 6 besitzt dem Betrage nach die Übertragungsfunktion eine Tiefpasses. Es erzeugt weiterhin eine Phasenverschiebung von 90 Grad. Die feste Phasenverschiebung von 90 Grad ist die entscheidende Eigenschaft des Digitalfilters 6, der Verlauf des Betrages der Übertragungsfunktion ist für das Funktionieren der Sprachverarbeitungsvorrichtung nicht entscheidend. So kann das Digitalfilter 6 auch mit Hilfe eines Differenzierers realisiert werden, was allerdings zu einer Unterdrückung von niederfrequenten Anteilen von x2(i) und damit zu einer verringerten Leistungfähigkeit der Sprachverarbeitungsvorrichtung führen würde.A digital filter 6 carries out a Hilbert transformation of the sample values x2 (i):

The digital filter 6 supplying the values x2 _H (i) of x2 (i) is an FIR filter of the order K, which has coefficients h (0), h (1), ..., h (K). In the present exemplary embodiment, K is sixteen, so that the digital filter 6 has seventeen coefficients. The digital filter 6 has a low pass in terms of the transfer function. It continues to produce a 90 degree phase shift. The fixed phase shift of 90 degrees is the decisive property of the digital filter 6, the course of the amount of the transfer function is not decisive for the functioning of the speech processing device. The digital filter 6 can thus also be implemented with the aid of a differentiator, which would, however, lead to a suppression of low-frequency components of x2 (i) and thus to a reduced performance of the speech processing device.

gebildet wird. N gibt die Anzahl der in die Berechnung eingehenden Abtastwerte von x1 an. N ist beispielsweise gleich 65. Die Multiplikation mit 1/P_x2(i) dient zur Vermeidung von Instabilitäten in der Steuervorrichtung 3 beim Steuern des Verzögerungsgliedes 4. Damit ergibt sich durch

ein auf die Kurzzeitleistung P_x2(i) normierter geschätzter Gradient grad(i) der Quadrate bzw. der Leistung der Fehlerwerte e₁₂(i) im Programmzyklus i.The output values x2 _H (i) are multiplied by the error values e₁₂ (i) and the reciprocal 1 / P _x2 (i) of a short-term power P _x2 (i), the short-term power P _x2 (i) after

${\text{P}}_{\text{x2}}{\text{(i) = P}}_{\text{x2}}\text{(i-1) + [x2 (i)] ² - [x2 (iN] ² (3)}$

is formed. N indicates the number of samples of x1 used in the calculation. N is, for example, equal to 65. The multiplication by 1 / P _x2 (i) serves to avoid instabilities in the control device 3 when the delay element 4 is controlled

an estimated on the short-term power P _x2 (i) graded gradient degree (i) of the squares or the power of the error values e₁₂ (i) in the program cycle i.

A function block 7 continuously forms estimated values SNR (i) of the associated signal / noise power ratio from the samples of the speech signal x2 (i), which are evaluated by a function block 8. An evaluation of the speech signal x1 (i) instead of the speech signal x2 (i) is also possible without the functionality of the speech processing device being restricted. The functioning of the function block 7 will be explained in more detail later with reference to FIGS. 6 to 8. Function block 8 carries out a threshold decision regarding the estimated values SNR (i). Only if the estimated values SNR (i) lie above a predefinable threshold is an intermediate memory 9 overwritten with the newly determined gradient estimated value grad (i). This case is symbolized by the closed position of a switch 11 which is controlled by the function block 8. The memory content (degree (i)) of the intermediate memory 9 is further processed by a functional unit 10. In the event that an estimated value SNR (i) lies below the predefinable threshold value, the buffer 9 is not overwritten with the newly determined gradient estimated value grad (i) and it retains its old memory content, which is symbolized by the open position of the switch 11 . The predefinable threshold, on which the opening and closing of the switch 11 by the function block 8 depends, is preferably between 0 and 10 dB.

The buffer store 9 supplies the gradient estimated values grad (i) stored in it to the functional unit 10, to which sample values of the speech signal x1 (i) are also fed and which both serve to supply the speech signal estimated values x1 _int (i) and also for setting the delay element 4.

zu geglätteten ("smoothed") Gradientenschätzwerten sgrad(i) weiterverarbeitet. α ist eine Konstante, die im Ausführungsbeispiel den Wert 0,95 besitzt. Die Werte sgrad(i) werden von einem Funktionsblock 13 zur Adaption von Verzögerungsschätzwerten T1'(i) nach

$\text{T1'(i+1) = T1'(i) - µ * sgrad(i) (6)}$

verwendet. Die Bestimmung von Verzögerungsschätzwerten T1'(i) erfolgt damit rekursiv. µ ist ein konstanter Faktor bzw. Konvergenzparameter und liegt im Bereich

R_x2x2 bezeichnet eine Autokorrelationsfunktion des Sprachsignals x2(i) an der Stelle Null. Ein besonders vorteilhafter Wertebereich von µ ist im vorliegenden Ausführungsbeispiel 1,5 < µ < 3.The gradient estimated values grad (i) are followed by a function block 12

$\text{degree (i) = α * degree (i-1) + (1-α) * degree (i) (5)}$

processed to smoothed gradient estimates sgrad (i). α is a constant that has the value 0.95 in the exemplary embodiment. The values sgrad (i) are followed by a function block 13 for adapting delay estimated values T1 '(i)

$\text{T1 '(i + 1) = T1' (i) - µ * sgrad (i) (6)}$

used. Delay estimated values T1 '(i) are thus determined recursively. µ is a constant factor or convergence parameter and is in the range

R _x2x2 denotes an autocorrelation function of the speech signal x2 (i) at the zero position. A particularly advantageous value range of μ in the present exemplary embodiment is 1.5 <μ <3.

The delay estimated values T1 '(i) can also be non-integer values, ie non-integer multiples of a sampling interval. A function block 14 rounds the delay estimated values T1 '(i) to integer delay values T1 (i) with which the delay device 4 is set. The rounding operation by function block 14 is necessary because of the values of the delay element 4 speech signal x1 (i) to be delayed is only available at the corresponding sampling times.

durch Interpolation dreier benachbarter Abtastwerte x1(i+T1(i)-1), x1(i+T1(i)) und x1(i+T1(i)+1) des Sprachsignals x1 bildet. Der Funktionsblock 15 ist somit in der Lage, durch den Sprachsignalschätzwert x1_int(i) im Programmzyklus i einen Wert des Sprachsignals x1 zum Zeitpunkt i+T1(i), d.h. zu einem Zeitpunkt zwischen zwei Abtastzeitpunkten, zu bilden bzw. zu interpolieren. Die beschriebene Interpolation durch Funktionsblock 15 kann dadurch ersetzt werden, daß Funktionsblock 15 eine Tiefpaßfilterung der Abtastwerte x1(i) zur Interpolation von Werten zwischen den Abtastzeitpunkten durchführt.The functional unit 10 furthermore has a functional block 15 which tracks the speech signal estimated values x1 _int (i)

${\text{x1}}_{\text{int}}\text{(i) = x1 (i + T1 (i)) + 0.5 * [T1 '(i) - T1 (i)] * [x1 (i + T1 (i) +1)) - x1 (i + T1 (i) -1)] (8)}$

by interpolating three adjacent samples x1 (i + T1 (i) -1), x1 (i + T1 (i)) and x1 (i + T1 (i) +1) of the speech signal x1. Function block 15 is thus able to use the speech signal estimate x1 _int (i) in program cycle i to form or interpolate a value of speech signal x1 at time i + T1 (i), ie at a time between two sampling times. The described interpolation by function block 15 can be replaced by function block 15 performing low-pass filtering of the sample values x1 (i) for the interpolation of values between the sample times.

die Verzögerungswerte T1(i), mit denen das Verzögerungsglied 4 eingestellt wird, nicht mehr konvergieren. Es ergäben sich starke Oszillationen der gerundeten Verzögerungswerte T1(i). Diese würden zwischen zwei Verzögerungswerten mit dem Abstand eines Abtastintervalls schwanken. Die entsprechende wahre Zeitverzögerung zwischen den Sprachsignalanteilen, die durch die unterschiedlichen Wegstrecken vom Sprecher zu den Mikrophonen M1 und M2 bestimmt ist, würde dabei zwischen diesen zwei Verzögerungswerten liegen. Im vorliegenden Ausführungsbeispiel werden solche Oszillationen dadurch vermieden, daß bei der Bildung der Fehlerwerte Sprachsignalschätzwerte x1_int(i) verwendet werden, durch die Werte des Sprachsignals x1(i) auch für Verzögerungen um nicht ganzzahlige Vielfache eines Abtastintervalls verfügbar sind, also auch an Zeitpunkten ungleich der Abtastzeitpunkte i des Sprachsignals x1(i).Would be used to determine the error values e₁₂ (i) instead of the speech signal estimates x1 _int (i) the delayed samples of the speech signal x1 (i) present at the output of the delay element 4, as can be seen from "IEEE Transactions on Acoustics, Speech, and Signal Processing, VOL ASSP-29, No. 3, June 1981, pp. 582-587 "would be known when error values were reached $\text{e₁₂ (i) = 0}$

the delay values T1 (i) with which the delay element 4 is set no longer converge. There would be strong oscillations of the rounded delay values T1 (i). These would fluctuate between two delay values with the interval of a sampling interval. The corresponding true time delay between the speech signal components caused by the different distances from Speaker to the microphones M1 and M2 is determined, would lie between these two delay values. In the present exemplary embodiment, such oscillations are avoided by using speech signal estimates x1 _int (i) in the formation of the error values, by means of which the values of the speech signal x1 (i) are also available for delays by non-integer multiples of a sampling interval, that is to say also unequal at times the sampling times i of the speech signal x1 (i).

The function block 12 used to smooth the gradient estimated values grad (i) brings about an improved determination of the delay estimated values T1 '(i).

The control device 3 adapts the delay estimates T1 '(i) or the delay values T1 (i) so that the square or the power of the error values e 1 (i) is reduced from one program cycle to the next. The convergence of T1 '(i) or T1 (i) is thus ensured.

FIG. 3 shows a speech processing device which works in principle like the speech processing device from FIG. 1 and now has three microphones M1, M2 and M3 for the delivery of microphone or speech signals. The microphone signals are fed to analog-to-digital converters 20, 21 and 22, which deliver digitized and thus sampled speech signals x1 (i), x2 (i) and x3 (i), which consist of speech and noise signal components. The speech signals x1 (i) and x3 (i) are supplied to adjustable delay elements 23 and 24. Analogously to FIG. 1, the speech signal x2 (i) is fed to a delay element 27 with a fixed delay time T _max. The output values of the delay elements 23, 24 and 27 are added to the sum signal X (i) by an adding device 25. A control device 26 evaluates the samples of the speech signals x1 (i), x2 (i) and x3 (i) and derives rounded integer delay values T1 (i) and T3 (i) from these samples, analogous to the mode of operation of the control device 3 from FIGS. 1 and 2, the integer values Correspond to multiples of a sampling interval of the sampled speech signals x1 (i), x2 (i) and x3 (i) and with which the delay elements 23 and 24 are set, so that an expansion from two to three microphone or speech signals to be processed is made possible.

FIG. 4 shows a first embodiment of the control device 26 from FIG. 3. Two functional units 10 are provided, the structure of which is identical to the structure of the functional unit 10 from FIG. 2 and which are used to set the delay elements 23 and 24 with the rounded time delay values T1 (i) and T3 (i).

The upper functional unit 10 provides speech signal estimates x1 _int (i). The lower functional unit 10 supplies speech signal estimates x3 _int (i). From a difference x1 _int (i) - x2 (i) and from a difference x3 _int (i) - x2 (i), error values e₁₂ (i) and e₃₂ (i) are formed.

Here too, a digital filter 6 is provided, which has already been described in more detail in the explanations relating to FIG. 2, and which serves to receive the sample values x2 (i) and to supply values x2 _H (i) which are obtained by a Hilbert transformation of the Samples x2 (i) are generated. The values x2 _H (i) are multiplied on the one hand by the error values e₁₂ (i) and on the other hand by the error values e₃₂ (i). The first product x2 _H (i) * e₁₂ (i) is the upper, the second product x2 _H (i) * e₃₂ (i) is fed to the lower functional unit 10. The arrangement of the function blocks 7 and 8, the buffer 9 and the switch 11 becomes analogous to Fig. 2 performed and is not shown in Fig. 4 for reasons of clarity.

FIG. 5 shows a version of the control device 26 that is expanded compared to FIG. 4. In contrast to FIG. 4, three digital filters 6 are now arranged instead of just one digital filter 6. These form the values x1 _H (i), x2 _H (i) and x3 _H (i) from the speech signal samples x1 (i), x2 (i) and x3 (i) by Hilbert transformation.

eingehen. Ein zweites Produkt ergibt sich aus ${\text{0,7*e₁₂(i)*x2}}_{\text{h}}\text{(i)}$

. Die beiden Produkte entsprechen gewichteten Gradientschätzwerten der Quadrate der Fehlerwerte e₁₃(i) und e₁₂(i). Die Summe aus erstem und zweitem Produkt und damit eine Linearkombination der gewichteten Gradientschätzwerten wird der oberen Funktionseinheit 10 zugeführt.In the upper half of the block diagram shown in Fig. 5, error values e₁₃ (i) from the difference x1 _int (i) -x2 (i) are formed, which in a first product ${\text{0.3 * e₁₃ (i) * x3}}_{\text{H}}\text{(i)}$

come in. A second product results from ${\text{0.7 * e₁₂ (i) * x2}}_{\text{H}}\text{(i)}$

. The two products correspond to weighted gradient estimates of the squares of the error values e₁₃ (i) and e₁₂ (i). The sum of the first and second product and thus a linear combination of the weighted gradient estimated values is fed to the upper functional unit 10.

und ein viertes Produkt ${\text{0,7*e₃₂(i)*x2}}_{\text{H}}\text{(i)}$

werden aufaddiert und die sich ergebende Summe wird der unteren Funktionseinheit 10 zugeführt.Similarly, error values e₃₁ (i) and e₃₂ (i) are formed in the lower half of the block diagram shown in FIG. 5. The error values e₃₁ (i) result from the difference x3 _int (i) -x1 (i). The error values e₃₂ (i) are formed by the difference x3 _int (i) -x2 (i). A third product ${\text{0.3 * e₃₁ (i) * x1}}_{\text{H}}\text{(i)}$

and a fourth product ${\text{0.7 * e₃₂ (i) * x2}}_{\text{H}}\text{(i)}$

are added up and the resulting sum is fed to the lower functional unit 10.

3, which contains a control device according to FIG. 4 or 5, a sum signal X (i) which is improved compared to the speech processing device with two microphones according to FIG. 1 can be generated. The signal / noise ratio and 3, the speech quality of the sum signal X (i) of the speech processing device according to FIG. 3 is further increased compared to the sum signal X (i) generated by the speech processing device according to FIG. 1. The control device according to FIG. 5 has an increased stability compared to the control device according to FIG. 4 when used in the speech processing device according to FIG. 3.

Both in FIG. 4 and in FIG. 5, for the sake of clarity, a representation of means (see function blocks 7 and 8, buffer store 9 and switch 11 in FIG. 2) has been dispensed with, which means that the speech processing is dependent on estimated values SNR ( i) for one of the microphone signals x1 (i), x2 (i) or x3 (i). Also for reasons of clarity, the normalization of products from error values and the output values of the digital filters 6 performing the Hilbert transformation to the power of an associated microphone signal (see 1 / P _x2 (i) in FIG. 2) is not shown. The expansion of the control devices 26 according to FIGS. 4 and 5 by these two technical features results from their implementation in the control device 3 according to FIG. 2.

With the help of FIGS. 6 and 7, the scheme is explained, on the basis of which the function block 7 from a sampled speech signal x (i), which consists of noise and speech signal components, the associated estimated values SNR (i) of the signal / noise power ratio, that is Ratio of the power of the speech signal components to the power of the noise signal components, determined. In FIG. 2, the sample values x2 (i) correspond to the sample values x (i). The function block 7 is shown in FIG. 6 on the basis of a block diagram. A function block 30 serves to form power values P _x (i) of the sample values x (i) by squaring the sample values. Function block 30 also leads a smoothing of these power values P _x (i) by. The resulting smoothed power values P _{x, s} (i) are supplied to both function block 31 and function block 32. Function block 31 continuously determines estimated values P _n (i) for estimating the power of the noise signal component of the sampled values x (i), ie the power of the noise signal components of the sampled values x (i) is determined. From the smoothed power values P _{x, s} (i) and the estimated values P _n (i), the function block 32 continuously determines estimated values SNR (i) of the signal / noise power ratio of the sampled values x (i).

ein Kurzzeitleistungswert P_x(i) und dann durch

${\text{P}}_{\text{x,s}}{\text{(i) = α * P}}_{\text{x,s}}{\text{(i-1) + (1-α)*P}}_{\text{x}}\text{(i) (2)}$

ein neuer geglätteter Leistungswert gebildet wird. Mit Formel (1) wird ein Kurzzeitleistungswert P_x(i) einer Gruppe von N aufeinanderfolgenden Abtastwerten x(i) ermittelt. N ist hier beispielsweise gleich 128. Der Wert α aus Gleichung (2) liegt zwischen 0,95 und 0,98. Die Ermittlung von geglätteten Leistungswerten P_x,s(i) kann auch nur durch Gleichung (2) durchgeführt werden, wobei dann allerdings der Wert α ungefähr auf den Wert 0,99 zu erhöhen und P_x(i) durch x²(i) zu ersetzen ist.FIG. 7 shows a flow chart which explains the function of the function block 7 in more detail. The flow chart shows how estimated values SNR (i) of the corresponding signal / noise power ratio are formed from the sampled values x (i) of the speech signal x by a computer program. In an initialization block 33, a counter variable Z is set to 0 and a variable P _{Mmin is set} to a value P _max at the beginning of the program described by FIG. P _max is chosen so large that the smoothed power values P _{x, s} (i) are always smaller than P _max . P _max can, for example, be set to the maximum representable numerical value of a computer used to implement the program. A new sample value x (i) is read in in block 34. In block 35, a counter variable Z is increased by the value 1, after which a new smoothed power value P _{x, s} (i) is formed in block 36. It results from the fact that initially by

${\text{P}}_{\text{x}}{\text{(i) = P}}_{\text{x}}\text{(i-1) + x² (i) - x² (iN) (1)}$

a short-term power value P _x (i) and then through

${\text{P}}_{\text{x, p}}{\text{(i) = α * P}}_{\text{x, p}}{\text{(i-1) + (1-α) * P}}_{\text{x}}\text{(i) (2)}$

a new smoothed power value is formed. A short-term power value P _x (i) of a group of N successive sample values x (i) is determined using formula (1). N here is 128, for example. The value α from equation (2) is between 0.95 and 0.98. The determination of smoothed power values P _{x, s} (i) can also only be carried out using equation (2), in which case however the value α should be increased approximately to the value 0.99 and P _x (i) by x² (i) is replace.

A branch 37 then queries whether the smoothed power _value P _{x, s} (i) that has just been determined is less than P _Mmin . If this question is answered in the affirmative, ie P _{x, s} (i) is less than P _Mmin , block 38 _sets P _Mmin to the value of P _{x, s} (i). If the question of branch 37 is answered in the negative, block 38 is skipped. This means that the minimum of M smoothed power _values P _{x, s} is in P _Mmin after M program _cycles . Then the branch 39 is used to query whether the counter variable Z has a value greater than or equal to a value M. In this way it is determined whether M smoothed power values have already been processed.

bestimmt. Diese Operation stellt sicher, daß der vorläufige Schätzwert P_n(i) nicht größer als der aktuelle geglättete Leistungswert P_x,s(i) sein kann. Danach wird mit Block 41 nach der Formel

${\text{SNR(i) = [P}}_{\text{x,s}}{\text{(i) - min{c*P}}_{\text{n}}{\text{(i), P}}_{\text{x,s}}{\text{(i)}] / [c*P}}_{\text{n}}\text{(i)] (4)}$

ein aktueller Schätzwert SNR(i) des Signal-/Rauschleistungsverhältnisses des Sprachsignals x(i) ermittelt. Im Normalfall dient das Produkt c*P_n(i) zur Abschätzung der aktuellen Leistung des Rauschsignalanteils, und die Differenz ${\text{P}}_{\text{x,s}}{\text{(i)-c*P}}_{\text{n}}\text{(i)}$

dient zur Abschätzung der aktuellen Leistung des Sprachsignalanteils des Sprachsignals x(i). Die aktuelle Leistung des Sprachsignals wird durch den geglätteten Leistungswert P_x,s(i) geschätzt. Die Gewichtung mit einem Skalierungsfaktor c verhindert, daß durch P_n(i) die Rauschsignalleistung mit einem zu kleinen Wert abgeschätzt wird. Der Skalierungsfaktor c liegt typisch im Bereich von 1,3 bis 2. Durch die Minimumbildung in Block 41 bzw. Gleichung (4) wird sichergestellt, daß das nicht logarithmierte Signal-/ Rauschleistungsverhältnis SNR(i) auch dann positiv ist, wenn im Ausnahmefall c*P_n(i) größer als P_x,s(i) ist. Dann wird die Leistung des Rauschsignalanteils des Sprachsignals gleich der durch P_x,s(i) geschätzten Leistung des Sprachsignals gesetzt. Die durch P_x,s(i)-P_x,s(i) geschätzte Leistung des Sprachsignalanteils des Sprachsignals ist dann wie auch das nicht logarithmische Signal-/ Rauschleistungsverhältnis gleich Null. Das Programm wird nach der Berechnung des Schätzwertes SNR(i) mit dem Einlesen eines neuen Sprachsignalabtastwertes x(i) durch Block 34 fortgesetzt.If the question of branch 39 is answered in the negative, ie M smoothed power values have not yet been processed, the program is continued with block 40. There, a preliminary estimate P _n (i) of the noise signal power of the speech signal x is obtained

${\text{P}}_{\text{n}}{\text{(i) = min {P}}_{\text{x, p}}{\text{(i), P}}_{\text{n}}\text{(i)} (3)}$

certainly. This operation ensures that the preliminary estimate P _n (i) is not greater than the current smoothed one Power value P _{x, s} (i) can be. Then with block 41 according to the formula

${\text{SNR (i) = [P}}_{\text{x, p}}{\text{(i) - min {c * P}}_{\text{n}}{\text{(i), P}}_{\text{x, p}}{\text{(i)}] / [c * P}}_{\text{n}}\text{(i)] (4)}$

a current estimate SNR (i) of the signal / noise power ratio of the speech signal x (i) is determined. In the normal case, the product c * P _n (i) is used to estimate the current power of the noise signal component, and the difference ${\text{P}}_{\text{x, p}}{\text{(i) -c * P}}_{\text{n}}\text{(i)}$

is used to estimate the current power of the speech signal component of the speech signal x (i). The current power of the speech signal is estimated by the smoothed power value P _{x, s} (i). The weighting with a scaling factor c prevents P _n (i) from estimating the noise signal power with a value that is too small. The scaling factor c is typically in the range from 1.3 to 2. The minimum formation in block 41 or equation (4) ensures that the non-logarithmic signal / noise power ratio SNR (i) is also positive if in exceptional cases c * P _n (i) is greater than P _{x, s} (i). Then the power of the noise signal component of the voice signal is set equal to the power of the voice signal estimated by P _{x, s} (i). The power of the speech signal component of the speech signal estimated by P _{x, s} (i) -P _{x, s} (i) is then equal to zero, as is the non-logarithmic signal / noise power ratio. After the calculation of the estimated value SNR (i), the program continues with the reading in of a new speech signal sample value x (i) by block 34.

die Komponenten eines Vektors minvec der Dimension W aktualisiert. Danach wird durch Verzweigung 43 abgefragt, ob die Komponenten minvec₁ bis minvec_W mit ansteigendem Vektorindex ansteigen, d.h. ob gilt:

${\text{minvec}}_{\text{j+1}}{\text{> minvec}}_{\text{j}}\text{für 1 ≦ j ≦ W-1 (6)}$

Wird die Abfrage von Verzweigung 43 verneint, d.h. die zuletzt ermittelten in den Komponenten des Vektors minvec stehenden zuletzt ermittelten W Minima steigen nicht monoton an, wird durch Block 44 nach

${\text{P}}_{\text{n}}{\text{(i) = min{minvec}}_{\text{W}}{\text{, minvec}}_{\text{W-1}}\text{, ... , minvec₁} (7)}$

der vorläufige Schätzwert P_n(i) der Rauschsignalleistung aus den Minima der Komponenten des Vektors minvec, d.h aus dem Minimum der letzten $\text{L=W*M}$

aufeinanderfolgenden geglätteten Leistungswerte P_x,s(i), bestimmt. Bei einer Bejahung der durch Verzweigung 43 gestellten Frage, d.h. bei einem monotonen Ansteigen der zuletzt ermittelten in den Komponenten des Vektors minvec stehenden W Minima wird in Block 45 P_n(i) gleich P_Mmin gesetzt, so daß eine Anpassung der Abschätzung des Rauschsignalanteils beschleunigt erfolgt, da P_n(i) an dem Minimum des letzten (M<L) Werte bestimmt wird. Danach wird in Block 46 die Zählervariable Z wieder auf 0 gesetzt und P_Mmin erhält erneut den Wert P_max.If the query of branch 39 is answered in the affirmative, ie M smoothed sample values P _{x, s} (i) have been processed, in block 42 by

updated the components of a vector minvec of dimension W. Then it is queried by branch 43 whether the components minvec 1 to minvec _{W increase} with increasing vector index, ie whether:

${\text{minvec}}_{\text{j + 1}}{\text{> minvec}}_{\text{j}}\text{for 1 ≦ j ≦ W-1 (6)}$

If the query of branch 43 is negated, ie the last W minima determined in the components of the vector minvec do not rise monotonously, block 44 follows

${\text{P}}_{\text{n}}{\text{(i) = min {minvec}}_{\text{W}}{\text{, minvec}}_{\text{W-1}}\text{, ..., minvec₁} (7)}$

the preliminary estimate P _n (i) of the noise signal power from the minima of the components of the vector minvec, ie from the minimum of the last $\text{L = W * M}$

successive smoothed power values P _{x, s} (i). If the question posed by branch 43 is answered in the affirmative, ie if the W minima found last in the components of the vector minvec increases monotonously, P _n (i) is set equal to P _Mmin in block 45, so that an adaptation of the estimation of the noise signal component is accelerated takes place since P _n (i) is determined at the minimum of the last (M <L) values. Then in block 46 the counter variable Z is reset to 0 and P _Mmin again receives the value P _max .

aufeinanderfolgenden geglätteten Leistungswerten P_x,s(i) W aufeinanderfolgende Untergruppen zusammengefaßt. Die Gruppen mit jeweils L Werten folgen lückenlos aufeinander und überlappen sich jeweils mit L-M gelätteten Leistungen P_x,s(i).Through the described program, M successive smoothed P _{x, s} (i) samples x (i) of the speech signal x are combined into a subgroup. Within In such a subgroup, the minimum of the smoothed power values P _{x, s} (i) is determined by the operations carried out with branch 37 and block 38. The W minima determined last are stored in the components of the vector minvec. If the last W minima are not monotonically increasing (see branch 43), then a preliminary estimate P _n (i) of the power of the noise signal component is determined from the minimum of the minima of the last W subgroups, ie from the minimum of a group, according to block 44. They are each used to form a group $\text{L = W * M}$

successive smoothed power values P _{x, s} (i) W summarized successive sub-groups. The groups with L values follow each other without gaps and overlap with powers P _{x, s} (i) smoothed with LM.

In the event that the minima of W successive subgroups increase monotonously (see branch 43), the minimum of the last subgroup with M smoothed power values P _x is determined by block 45 to estimate the current estimated value P _n (i) of the power of the noise signal component _{. s} (i) used. This shortens the time period with which monotonically increasing smoothed power values P _{x, s} (i) also cause a change in the estimated values SNR (i).

geglätteten Leistungswerten P_x,s(i) zusammengefaßt. Nach jeweils M geglätteten Leistungen P_x,s(i) wird aus dem Minimum der letzten W Untergruppenminima bzw. der letzten L geglätteten Leistungswerte P_x,s(i) der Wert P_n(i) bestimmt, der zur Abschätzung der Rauschsignalleistung dient. In Fig. 8 sind acht Gruppen mit jeweils L Abtastwerten x(i) dargestellt, die jeweils W = 4 Untergruppen mit M geglätteten Leistungswerten P_x,s(i) enthalten. Die acht Gruppen überlappen sich teilweise. So enthalten zwei aufeinanderfolgende Gruppen jeweils L-M gleiche geglättete Leistungswerte P_x,s(i). Auf diese Weise wird ein guter Kompromiß zwischen dem erforderlichen Rechenaufwand und der jeweiligen Verzögerungszeit erreicht, mit der eine Aktualisierung eines Schätzwertes P_n(i) der Rauschsignalleistung zur Aktualisierung eines Schätzwertes SNR(i) des Signal/ Rauschleistungsverhältnisses erfolgt. Eine Realisierung mit aneinandergrenzenden, d.h. sich nicht überlappenden Gruppen ist auch denkbar. Allerdings ist dann bei verringertem Rechenaufwand die Zeitspanne zwischen zwei Schätzwerten SNR(i) vergrößert, so daß die Reaktionszeit auf sich ändernde SNR des Sprachsignals x(i) vergrößert ist.8 illustrates how the smoothed power values P _{x, s are combined} in groups and subgroups. In each case, M smoothed power values P _{x, s} (i), which are present at sampling times i, are combined into a subgroup. The subgroups are contiguous. The minimum of the smoothed power values P _{x, s} (i) is determined for each subgroup. W subgroup minima are stored in the vector minvec. As a rule, ie with non-monotonically increasing W subgroups Minima, W subgroups become a group with $\text{L = W * M}$

smoothed power values P _{x, s} (i) summarized. After M smoothed powers P _{x, s} (i), the value P _n (i) is determined from the minimum of the last W subgroup minima or the last L smoothed power values P _{x, s} (i), which is used to estimate the noise signal power. 8 shows eight groups each with L samples x (i), each of which contains W = 4 subgroups with M smoothed power values P _{x, s} (i). The eight groups partially overlap. Two successive groups each contain the same smoothed power values P _{x, s} (i). In this way, a good compromise is achieved between the required computational effort and the respective delay time with which an update of an estimated value P _n (i) of the noise signal power takes place in order to update an estimated value SNR (i) of the signal / noise power ratio. Implementation with adjacent, ie non-overlapping groups is also conceivable. However, the time span between two estimated values SNR (i) is then increased with a reduced computing effort, so that the reaction time to changing SNR of the speech signal x (i) is increased.

The described speech processing device thus has an estimation device which is suitable for the continuous formation of estimated values SNR (i) of the signal / noise power ratio of noisy speech signals x (i). In particular, no speech pauses are required to estimate the noise signal power. The estimation device described uses the special time profile of smoothed power values of the speech signal x (i), which is characterized by peaks and intermediate areas with smaller smoothed power values P _{x, s} (i), their temporal expansion from the respective speech source, ie the respective speaker , depends. The areas between the peaks are used to estimate the power of the noise signal component. The groups with L smoothed power values P _{x, s} (i) must follow one another without gaps, ie they must either adjoin or overlap. Furthermore, it must be ensured that at least one value of an area lying between two peaks with smaller smoothed power values P _{x, s} (i) can be recorded by each group, ie each group must contain so many smoothed power values P _{x, s} (i) that at least all values belonging to any peak can be recorded. Since the most extended peaks can be estimated by the most extended phonemes of a speech signal, ie the vowels, the number L describing the group size can be derived from this. For a sampling rate of the speech signal of 8 kHz, a useful value of L is in the range between 3000 and 8000. An advantageous value for W is 4. With such a dimensioning, there is a good compromise between the computational effort and the speed of reaction of the function block 7.

FIG. 9 shows a use of the voice processing device from FIG. 3 in a mobile radio terminal 50. The speech processing means 20 to 26 are combined in a function block 51 which forms the sum signal values X (i) from the microphone or speech signals generated by the microphones M1, M2 and M3. A function block 52 processing the sum signal values X (i) combines all the other means of the mobile radio terminal 52 for receiving, processing and transmitting signals which are used for communication with a base station (not shown), the transmission and reception of signals via a to the function block 52 coupled antenna 54 takes place. Furthermore, one with the function block 52 coupled speakers 53 are provided. A user (speaker, listener) communicates acoustically with the mobile radio terminal 50 via the microphones M1 to M3 and the loudspeaker 53, which are parts of a hands-free device integrated in the mobile radio terminal 50. The use of such a mobile radio terminal 50 is particularly advantageous in motor vehicles, since there the hands-free communication via the mobile radio terminal is particularly disturbed by engine or driving noise (noise).

Claims (6)

Mobile radio terminal (50) with a speech processing device for processing speech signals (x (i)) consisting of noise and speech signal components, with an estimation device (3) for the continuous formation of estimated values (SNR (i)) of the signal / noise power ratio of the speech signals (x ( i)) by means of - Determination of the power values (P _x (i)) of samples of the speech signals (x (i)) - smoothing the power values (P _x (i)) - Determination of the minimum of a group of L successive smoothed power values (P _{x, s} (i)), wherein the groups follow one another without gaps and contain at least as many smoothed power values (P _{x, s} (i)) that all each any phoneme the smoothed power values (P _{x, s} (i)) associated with the speech signal can be detected by a single group, - Formation of a current estimate (SNR (i)) of the signal / noise power ratio from the current smoothed power value (P _{x, s} (i)) and the last determined minimum. Mobile radio terminal according to claim 1,
characterized by
that means of forming contiguous sub-groups with each $\text{M = L / W}$

successive smoothed power values and for determining the minimum of the minima of each W successive subgroups for determining the minimum of the associated group are provided, where W represents a natural number and W subgroups form a group. Mobile radio terminal according to claim 2,
characterized by
that means are provided for using the last determined minimum of a subgroup instead of the last determined minimum of a group with a predeterminable number of monotonically increasing minima of subgroups in order to estimate a current value (SNR (i)) of the signal / noise power ratio. Mobile radio terminal according to one of claims 1 to 3,
characterized by
that means are provided for using the current smoothed power value instead of a recently determined minimum of a group or subgroup for estimating a current value of the signal / noise power ratio in the event that the current smoothed power value is less than the last determined minimum. Mobile radio terminal according to one of claims 1 to 4,
characterized by
that speech processing means are provided for processing the disturbed speech signals (x (i)) as a function of the estimated values of the signal / noise power ratio. Speech processing device (1) for processing speech signals (x (i)) consisting of noise and speech signal components with an estimation device (3) for continuously forming estimates (SNR (i)) of the signal / noise power ratio of the speech signals (x (i)) by means of - Determination of the power values (P _x (i)) of samples of the speech signals (x (i)) - smoothing the power values (P _x (i)) - Determination of the minimum of a group of L successive smoothed power values (P _{x, s} (i)), wherein the groups follow one another without gaps and contain at least as many smoothed power values (P _{x, s} (i)) that all each any phoneme the smoothed power values (P _{x, s} (i)) associated with the speech signal can be detected by a single group, - Formation of a current estimate (SNR (i)) of the signal / noise power ratio from the current smoothed power value (P _{x, s} (i)) and the last determined minimum.

Images