EP0612059A2 - Method for estimating the propagation time in noisy speech channels - Google Patents

Abstract

The invention relates to a method for noise reduction in a speech recognition system. The phases of at least two disturbed signals are estimated. The phase estimation and the phase compensation required for noise reduction is carried out in the frequency domain. The background disturbance and the transient response of the room are constantly estimated. <IMAGE>

Description

The invention relates to a method according to the preamble of patent claim 1.

Such a method is used in automatic speech recognition systems or for hands-free systems, e.g. in offices, motor vehicles etc.

Disrupted speech is easier to grasp if it is recorded with two or more channels. Man uses two channels, his two ears. Through psychoacoustic postprocessing, the direction of the speaker is determined and the Background noise is hidden. In technical devices, two or more channels can be used for recording. These signals can then be processed using digital signal processing.

An important aspect of multi-channel processing is the estimation of the runtime difference between the individual channels. If the time difference is known, the direction of the sound event (speaker) can be determined. The signals of the individual channels can be corrected accordingly and processed further. E.g. If uncorrected signals are combined to form a sum signal, individual spectral components of the signal can be amplified, attenuated or canceled by interference.

A method for automatically determining the runtime differences between two microphones is known from a publication by M. Schlang, ITG Conference 1988, Bad Nauheim pp. 69-73. It works in the time domain. However, this method cannot be used in the case of severe disturbances.

The invention is therefore based on the object of specifying a method for estimating the runtime for a speech recognition system which can also be used in the case of strong background noise, is suitable for a multichannel transmission system and saves time and costs.

The object is achieved by the features specified in the characterizing part of patent claim 1. Advantageous refinements and / or further developments can be found in the subclaims.

The invention is described using an exemplary embodiment with reference to schematic drawings.

In FIG. 1, the phase estimation is explained using a block diagram.

FIG. 2 shows a representation of the quantities S _B , S _I , S _N and g as a function of time for a driving noise of 140 km / h.

In the present invention, a 2-channel runtime compensation is presented. The expansion to several channels is easily possible with the corresponding additional effort. The runtime compensation is part of the signal preprocessing of a multi-channel noise reduction, which e.g. can be used for a speech recognizer in the vehicle.

The runtime is determined in the frequency domain. This enables a simple runtime correction by multiplying the spectrum by the new phase and leads to a low computing effort.

The speech and sound recordings for the development and evaluation of the present method were carried out in a vehicle with two microphones. The disturbance is the driving noise in different driving situations.

With the method according to the invention, the phases are determined in the frequency domain at a number of maxima of the cross-correlation. The background disturbance and the transient response of the room are constantly valued. The individual phase values are only processed at the start of a transient process and when the background noise is exceeded by a certain factor. In the further processing of the phase values, a linear phase relationship is assumed and the variance of the estimate is taken into account when smoothing the values. The consideration of the settling process of the room leads to the fact that a phase estimation takes place only with strong energy increases of the speech. A new phase estimate is available immediately at the beginning of the word. The influence of reflections is reduced. By taking the background noise into account, the method is well suited for practical use, for example in a vehicle. Using a block diagram in FIG. 1, the procedure of the phase estimation is explained in more detail.

The microphone signals x and y are transformed into the frequency range (FFT, Fast Fourier Transformation). The transformation length is chosen to be N = 256. The transformed segments X _l (i) and Y _l (i) resulted. l denotes the block index of the segments, i the discrete frequency (i = 0,1,2, ..., N-1). The segments are half overlapped and weighted with a hanning window. (The sampling rate of the signals x and y is 12 kHz.)

In the frequency domain, the long-term mean of the magnitude spectrum is subtracted (SPS, spectral subtraction). The phase of the signals is not changed. The noise is reduced. The estimated values X̂ and Ŷ result. The PLC is a standard procedure and a simple version can be used here. Are only minor Faults are present, the PLC can be dispensed with entirely.

Für den zweiten Kanal Y gelten die entsprechenden Gleichungen.The interference spectrum S _nn (i) is estimated with the smoothing constant β. The interference spectrum is normalized and subtracted. l denotes the block index, i the discrete frequency. For example, β _l = 0.03 is used as the smoothing constant.

${\hat{\text{S}}}_{\text{nn, l}}{\text{(i) = (l-β}}_{\text{l}}\text{)}{\hat{\text{S}}}_{\text{nn, l-1}}{\text{(i) + β}}_{\text{l}}{\text{| X}}_{\text{l}}\text{(i) | ² (1)}$

The corresponding equations apply to the second channel Y.

${\text{B}}_{\text{xy,l}}{\text{(i) = |S}}_{\text{xy,l}}\text{(i)| (5)}$

Als Glättungskonstante α wird z.B. α = l gewählt. Werte α « l sind nicht sinnvoll.The magnitude of the cross power density B _{XY, l is} calculated from the estimated values X̂ and X̂. The range (N _u , N _o ) is, for example, between 300 and 1500 Hz (N _u = 6, N _o = 31, at N = 256). The following applies

${\text{S}}_{\text{xy, l}}{\text{(i) = (l-α) S}}_{\text{xy, ll}}\text{(i) + α}{\hat{\text{X}}}_{\text{l}}\text{(i)}{\hat{\text{Y}}}_{\text{l}}{\text{*(in}}_{\text{u}}{\text{≦ i ≦ N}}_{\text{O}}\text{(4th}$

${\text{B}}_{\text{xy, l}}{\text{(i) = | S}}_{\text{xy, l}}\text{(i) | (5)}$

For example, α = 1 is selected as the smoothing constant α. Values α «l do not make sense.

With a pre-emphasis higher frequencies can be raised. This is advantageous if the speech signal and the interference signal have a lower power at higher frequencies. The values of the cross power B _xy (i) can, for example, be increased linearly by 10dB in the range 300 to 1500 Hz. However, the pre-emphasis can also be predetermined by the microphone characteristics.

bestimmt.M maxima are determined and summed from the values B _xy (i). For example, M = 8 can be used. It becomes a current estimate

certainly.

Außerdem wird über einen Geräuschmonitor eine adaptive Glättungskonstante h berechnet. Mit dieser Glättungskonstanten ergibt sich ein Schätzwert S_N für die Störung. Wurde zuvor eine spektrale Substraktion (SPS) durchgeführt, ist S_N ein Schätzwert für die Reststörung. Für die Glättungskonstante h_o gilt z.B. h_o = 0.03

${\text{S}}_{\text{N,L}}{\text{= (1 - h}}_{\text{l}}{\text{)S}}_{\text{N,l-1}}{\text{+ h}}_{\text{l}}{\text{S}}_{\text{B,l}}\text{ (9)}$

Die Phase der gestörten Signale wird aus den Real- und Imaginärteilen von S_xy berechnet. Die Phase wird nur an den M zuvor bestimmten Maxima berechnet.

und

Daraus ergibt sich der Phasenanstieg:

Mit der Länge der Fouriertransformation N und der max. zulässigen Verschiebung um n Taps ergibt sich (N = 256):

$${\text{|φ'|}}_{\text{max}}\text{= |n|}\frac{\text{2π}}{\text{N}}\text{ (13)}$$

Übersteigt der Phasenanstieg |φ'| an einem der Maxima |φ'|_max, so wird dieser Wert φ' nicht weiterverwendet. Es wird eine adaptive Glättungskonstante g berechnet:

${\text{g}}_{\text{l}}{\text{≦ g}}_{\text{maX}}\text{ (15)}$

${\text{g}}_{\text{max}}{\text{= 0,25; g}}_{\text{O}}\text{= 0,25 (16)}$

Der aktuelle Wert S_B muß um den Faktor c größer sein als die simulierte Impulsantwort S_I

${\text{S}}_{\text{B,l}}{\text{≧ cS}}_{\text{I,l}}\text{; c = 2 (17)}$

sonst gilt:

${\text{g}}_{\text{l}}\text{= 0 (18)}$

Der aktuelle Wert S_B muß um den Faktor d größer sein als das Restrauschen S_N

${\text{S}}_{\text{B,l}}{\text{≧ dS}}_{\text{N,l}}\text{; d = 3 (19)}$

sonst gilt ebenfalls

${\text{g}}_{\text{l}}\text{= 0 (20)}$

Ist Gl. (17) oder Gl. (19) nicht erfüllt, d.h. gilt $\text{g = O}$

, so kann die Phasenschätzung abgebrochen werden. Es gilt der alte Phasenschätzwert.A "simulated impulse response" S _{I is} calculated via an impulse monitor. The transient response of the surrounding space to sudden high-energy sound events (speech) is roughly simulated (e.g. γ = 0.l is selected). The smoothing of the phase value "from the beginning of the word into the word" can be adjusted with γ.

${\text{S}}_{\text{I, 1}}{\text{= (l - γ) S}}_{\text{I, l-1}}{\text{+ γS}}_{\text{B, l}}\text{(7)}$

In addition, an adaptive smoothing constant h is calculated using a noise monitor. This smoothing constant results in an estimate S _N for the disturbance. If spectral subtraction (SPS) was carried out beforehand, S _{N is} an estimate of the residual interference. For the smoothing constant h _o , for example, h _o = 0.03 applies

${\text{S}}_{\text{N, L}}{\text{= (1 - h}}_{\text{l}}{\text{) P}}_{\text{N, l-1}}{\text{+ h}}_{\text{l}}{\text{S}}_{\text{B, l}}\text{(9)}$

The phase of the disturbed signals is calculated from the real and imaginary parts of S _xy . The phase is only calculated on the M predetermined maxima.

and

This results in the phase increase:

With the length of the Fourier transform N and the max. permissible shift by n taps results (N = 256):

$${\text{| φ '|}}_{\text{Max}}\text{= | n |}\frac{\text{2π}}{\text{N}}\text{(13)}$$

If the phase increase exceeds | φ '| at one of the maxima | φ '| _max , this value φ 'is no longer used. An adaptive smoothing constant g is calculated:

${\text{G}}_{\text{l}}{\text{≦ g}}_{\text{Max}}\text{(15)}$

${\text{G}}_{\text{Max}}{\text{= 0.25; G}}_{\text{O}}\text{= 0.25 (16)}$

The current value S _B must be greater by a factor c than the simulated impulse response S _I

${\text{S}}_{\text{B, l}}{\text{≧ cS}}_{\text{I, l}}\text{; c = 2 (17)}$

otherwise:

${\text{G}}_{\text{l}}\text{= 0 (18)}$

The current value S _B must be greater than the residual noise S _N by a factor of d

${\text{S}}_{\text{B, l}}{\text{≧ dS}}_{\text{N, l}}\text{; d = 3 (19)}$

otherwise also applies

${\text{G}}_{\text{l}}\text{= 0 (20)}$

Is Eq. (17) or Eq. (19) not fulfilled, ie applies $\text{g = O}$

, the phase estimation can be canceled. The old phase estimate applies.

gilt:

Von den ursprünglichen M Maxima werden wegen Gl. (21) nur M' für die Gl. (22, 23) verwendet. Ist die Anzahl M' der für die Summen gültigen Werte φ kleiner als M_min, gilt der geschätzte Phasenanstieg als zu unsicher oder außerhalb des Nutzbereichs (z.B. M_min = 6, bei M = 8). Die Phasenschätzung wird dann nicht aktualisiert und das Verfahren hier abgebrochen. Es gilt der alte Phasenschätzwert.For all

${\text{| φ '}}_{\text{l}}{\text{(i) | ≦ | φ}}_{\text{'}}\text{| max (21)}$

applies:

Because of Eq. (21) only M 'for Eq. (22, 23) used. If the number M 'of the values φ valid for the sums is less than M _min , the estimated phase increase is considered too uncertain or outside the useful range (eg M _min = 6, with M = 8). The phase estimate is then not updated and the process is terminated here. The old phase estimate applies.

Als maximale Varianz wird

${\text{σ²}}_{\text{max}}{\text{= | φ' |²}}_{\text{max}}\text{ (25)}$

verwendet.The variance of the estimate is calculated:

As the maximum variance

${\text{σ²}}_{\text{Max}}{\text{= | φ '| ²}}_{\text{Max}}\text{(25)}$

used.

Bei einer mittleren Streuung gilt:

Bei sehr geringer Streuung gilt:

Entsprechend den Gl. 19 - 22 wird g in der Regel nur am Wortanfang größer Null sein. Dabei muß die Energie des Wortes größer sein als die Energie des Restgeräusches und der simulierten Impulsantwort. Mit der Variablen j wird die aufeinanderfolgende Anzahl für g > 0 gezählt. Entsprechend gilt für die Glättung:

Wird z.B. infolge einer Störung die Bedingung g > 0 nur einmal in Folge erfüllt, wird die Phasenschätzung nicht aktualisiert. Eine Aktualisierung der Phasenschätzung erfolgt nur dann, wenn g > 0 mindestens 2-mal in Folge erfüllt wird.The smoothing constant g is weighted according to the variance. In the case of a large spread:

The following applies to medium scatter:

With very little scatter:

According to Eq. 19-22, g will usually only be greater than zero at the beginning of the word. The energy of the word must be greater than the energy of the residual noise and the simulated impulse response. The variable j is used to count the successive number for g> 0. The following applies accordingly to smoothing:

For example, if the condition g> 0 is fulfilled only once in succession as a result of a fault, the phase estimate is not updated. The phase estimate is only updated if g> 0 is met at least twice in succession.

An example of the intermediate quantities S _B , S _I , S _N 'and g and the phase estimate derived therefrom is shown in FIG. 2. The word "station selection" is spoken and the driving noise at 140 km / h is added. The method is used as indicated above. The phase estimate is given in samples n. With the size S _I , the "speech impulse" is partially masked and thus an estimate is only allowed in the case of strong energy increases (S _B must exceed S _I by a factor of 2). The estimation of the residual disturbance S _N enables greater robustness against noise (S _B must exceed S _N by a factor of 3).

Claims (11)

Method for time-of- flight estimation in which the time-of- flight differences of noise- distorted signals from at least two voice channels are determined by means of a cross correlation , that the phase values of at least two signals over a certain number of maxima of the cross power density are determined and their phase shift is determined in the frequency range, and - That the required phase compensation is also carried out in the frequency domain. Method according to Claim 1, characterized in that background disturbances and the transient response of the Space in the determination of the phase values. A method according to claim 2, characterized in that the background noise is estimated by a noise monitor and that a new phase value is only determined if the estimated value of the background noise is exceeded by a certain factor. Method according to Claim 2, characterized in that the transient response of the surrounding space is estimated using a pulse monitor in such a way that a new phase estimate is determined only when there is a strong increase in energy in the signals. Method according to one of the preceding claims, characterized in that a linear delay of the signals is assumed. Method according to one of the preceding claims, characterized in that the phase value is smoothed from the beginning of the word into the spoken word, and that the variance of the estimate is taken into account when smoothing the phase values. Method according to one of the preceding claims, characterized in that that at least two microphone signals x, y are transformed into the frequency range by means of an FFT (Fast Fourier Transform), - that the estimated values X̂, Ŷ are determined by spectral subtraction from the transformed signals, the amount of the cross power density B _{xy is} determined from the estimated values X Werten, Ŷ, - that the maxima of the cross power density are determined, and that from a certain number of maxima of the cross power density B _xy, a current value S _B for the disturbed signals is determined that depends on the actual value S _B the phases φ of the disturbed signals are ascertained and thus the Phase increase φ 'is determined, - that the phase rise φ 'is smoothed out by coupling a simulated speech pulse S _I with the current value S _{B of} the disturbed signals via a pulse monitor, in such a way that a new phase estimation is only carried out if a strong energy rise of the microphone signal is registered, and - That an estimated value S _N for the background noise disturbance is determined with a noise monitor and is coupled with the current value S _{B of} the disturbed signals, such that a renewed phase estimation is only carried out when the background disturbance is clearly exceeded by the signal. A method according to claim 7, characterized in that a maximum phase increase | φ '| _{max can be specified} for the phase at the individual maxima and a new phase estimate is only carried out if the phase increase by at least M 'of the M maxima exceeds the maximum increase | φ' | _max does not exceed. Method according to Claim 7, characterized in that the variance of the phase increases at the individual maxima is taken into account when smoothing the phase increase. Method according to claims 7 to 9, characterized in that a renewed phase estimation is only carried out if the conditions for a valid phase increase occur several times in succession. Method according to one of the preceding claims, characterized in that the disturbed speech is recorded on more than two speech channels and that the runtime differences of the individual channels are estimated.

Images