EP0874352A2 - Voice activity detection - Google Patents

Abstract

The speech activity identification method involves using segmentation of a speech signal with a wavelet transformation calculated for each frame, from which a set of parameters are extracted. A set of decision variables is provided for controlling a decision logic, providing a signal indicating whether or not speech is present. The speech activity identification method is employed by a speech activity identification module (5) controlling a speech coder (7) and a speech decoder (22) and a background noise coder (10) and background noise decoder (23).

Description

The invention relates to a method and Circuit arrangement for automatic Voice activity recognition according to the generic term of Claims 1 and 5 respectively.

For digital mobile radio or voice storage systems and a large number of other applications it is beneficial a discontinuous transfer of the Make speech coding parameters. This allows during of speech pauses or time intervals that are essentially be dominated by background noise, the bit rate be significantly reduced. Advantages arise from this among other things through low energy consumption in mobile End devices, due to a higher average bit rate for simultaneous services, such as data transmission or through a higher memory capacity on memory chips. The extent of Benefits depends on the percentage of breaks in the speech signal as well on the quality of the automatic voice activity recognizer that is necessary for the detection of such intervals. To the a low voice activity rate is beneficial, on the other hand, active language should not be cut off in order not to impair the speech quality. This problem is the fundamental problem of automatic voice activity detectors, especially at the Presence of high level background noise.

Known methods for automatic Speech activity detection usually use Decision parameters based on time averages Based windows of constant length. As an example of this the autocorrelation coefficients, Called zero crossing rate or basic speech period, where these parameters have limited flexibility in the Have selection of time / frequency domain resolution, which is usually determined by the frame length of the associated Speech encoder / decoder is fixed. in the In contrast, the well-known wavelet transformation calculates a breakdown into the time / frequency domain too low frequency but high time domain resolution at high frequencies and too low time - but high Frequency domain resolution at low frequencies leads. These are properties for the analysis of speech signals well suited and therefore for the classification of active Speech in the classes voiced, unvoiced and transitions have already been used, as in the published patent application DE 195 38 852 A1 "Method and arrangement for Classification of Speech Signals ", 1997.

The invention is therefore based on the object Method and a circuit arrangement for Voice activity detection based on the wavelet transformation to create, it should be decided whether language for the time segment to be considered at all or speech sounds.

The solution of the method according to the invention is in Characteristic of claim 1 characterized.

Further refinements of the method according to the invention result from the patent claims 2 to 4.

The solution for the circuit arrangement according to the invention is characterized in the characterizing part of claim 5.

Further features or configurations of the Circuit arrangements are in claim 6 characterized.

The present procedure for automatic Speech activity detection for speech encoders / decoders for source-controlled reduction of the middle one Transmission rate is characterized in that after the Segmentation of the speech signal for each frame one Wavelet transformation is calculated from the one sentence Parameters are determined from which using fixed Thresholds are calculated using a set of binary decision variables that control a decision logic, the result After smoothing the time, a statement for each frame "Language available / no language available" returns. In that it is determined whether for the contemplating time segment there is language at all a source-controlled reduction of the middle one Transfer rate reached.

Further advantages of the method for voice activity detection and the corresponding circuit arrangement are based on exemplary embodiments, the following are described in more detail.

In the description, in the claims, in the Summary and in the drawing are those in the list of reference numerals given below Terms and reference symbols used.

The invention will now be described on the basis of exemplary embodiments, shown in the drawings.

Fig. 1

ein Blockschaltbild für die Sprachaktivitätserkennung als Vorstufe für einen variabelratigen Sprachcodierer-/decodierer und

Fig. 2

ein Blockschaltbild eines automatischen Sprachaktivitätserkenners.

In the drawing:

Fig. 1

a block diagram for voice activity detection as a preliminary stage for a variable-rate speech encoder / decoder and

Fig. 2

a block diagram of an automatic voice activity recognizer.

The procedure decides whether to do so contemplating time segment there is language at all. This allows it to function control or in general as Preliminary stage for a variable-rate speech coder / decoder Bit rate can be used. 1 is a block diagram for voice activity recognition as a preliminary stage for one variable rate speech encoder / decoder shown on its input 1 receives the respective input language. The input language is both on lines 2 and 3 a switch 4 as well as the input of a Voice activity detection circuit or module 5 given. The switch 4 directs the input language depending on the output signal of the voice activity detection circuit 5, which is used to control the switch 4 is connected to it via a feedback line 6, either on line 7 or on line 8. Die Line 7 leads to a speech encoder 9 and the line 8 to a background noise encoder 10. The bit stream of the speech encoder 9 is connected via a line 11 to the given an input of a switch 13 and the bit stream the background noise encoder 10 via a line 12 to the other input of the switch 13. The switch 13 also receives the output signals via a line 14 the voice activity detection circuit 5, whereby the Switch 13 is controlled. The output of the switch 13 is connected to a transmission channel 16 via a line 15 connected, also on the input side to line 14 for the output signals of the voice activity detection module 5 is connected. The output of the transmission channel 16 reaches the entrance once via a line 17 a further switch 19 and via a line 18 the control input of the switch 19 and the control input a switch 26. The switch 19 is over Output lines 20 and 21 with a speech decoder 22 and with a background noise decoder 23 connected, the outputs on lines 24 and 25 on the Get input of the switch 26 already mentioned, the depending on the control signals on line 18 at the output 27 either signals for the decoded speech or the provides decoded background noise.

2 is a block diagram of an automatic Voice activity recognizer represented, which in turn on its input 1 receives the input language and to one Segmentation circuit 28 passes on. The exit of this Segmentation circuit 28 is connected via a line 29 transmit a wavelet transform circuit 30 which again via a line 31 with the input of a Processor 32 connected to calculate the energy quantities is. The output of processor 32 or one Computing circuit is connected in parallel with a via a line 33 Pause detector 34, with a circuit for calculating a Stationarity measure 35, with a first Background detector 36 and with a second Background detector 37 connected. The outputs of the mentioned circuits 34 to 37 are about corresponding Lines 38 to 41 with decision logic 42 connected, the output via a line 43 with a Smoothing circuit 44 connected for temporal smoothing whose output 45 is also the output of the speech activity recognizer is.

1. Im ersten Verfahren wird die Filterbank direkt auf den Eingangs-Sprachrahmen s=(s(O),...,s(N-1))^T angewandt und die beiden Filterausgänge um den Faktor zwei unterabgetastet. Dadurch entsteht am Tiefpaßfilterausgang ein Satz von Approximationskoeffizienten A ₁=(A₁(O),...A₁(N/2-1))^T sowie am Hochpaßfilterausgang ein Satz von Detailkoeffizienten D ₁=(D₁(O),...D₁(N/2-1))^T. Dieses Verfahren wird nun rekursiv immer auf die Approximationskoeffizienten der vorherigen Stufe angewandt, wodurch sich in der letzten Stufe L als Ergebnis der Transformation der Vektor DWT(s)=(D ₁ ^T,D₂ ^T...,D _L ^T,A _L ^T)^T mit insgesamt N Koeffizienten ergibt.

2. Die zweite Variante zur Berechnung der Transformation beruht wie die erste auf einer Filterbankzerlegung. Der Unterschied besteht aber darin, daß die Filterausgänge nicht mehr unterabgetastet werden. Dadurch entstehen nach jeder Stufe Vektoren der Länge N und nach der letzten Stufe ein Ausgangsvektor mit insgesamt (L+1)N Koeffizienten. Um die Auflösungseigenschaften der Wavelet-Transformation zu erhalten, müssen die Filterimpulsantworten für jede Stufe durch Überabtastung um den Faktor zwei aus der vorherigen Stufe gewonnen werden. In der ersten Stufe werden die gleichen Filter benutzt wie für Variante 1. Durch den insgesamt höheren Aufwand - höhere Redundanz in der Darstellung im Bildbereich - gegenüber Variante 1 kann eine Verbesserung der Leistungsfähigkeit des Verfahrens erreicht werden.

The method for automatic speech activity detection will now be described in more detail with reference to the block diagram of the automatic speech activity detector according to FIG. 2. After segmenting the input signal in the segmentation circuit 28, the wavelet transformation in the wavelet transformation circuit 30 is calculated for each segment. Subsequently, a set of energy parameters is determined in the processor 32 from the transformation coefficients and compared with fixed threshold values. This creates binary decision variables that control decision logic 42 that outputs a preliminary result for each frame. At the end, this provisional decision is further processed by means of temporal smoothing in the circuit 44, as a result of which the final result "language or no language" is available at the output 45 for the current frame. The individual processing circuits or blocks of the block diagram according to FIG. 2 will now be described in more detail below. The wavelet transform circuit 30 carries out the following operations: The input speech is divided into frames of length N samples, N being adaptable to a specific speech coding method. The discrete wavelet transformation is calculated for each frame. For many basic functions of the wavelet transformation, a filter bank with a high-pass filter or a low-pass filter can be derived, with which the transformation can be carried out recursively. Attention is drawn to the classes of the Daubechies Wavelets and the Spline Wavelets, which lead to a particularly efficient implementation of the transformation by means of filters of short length. Two methods are described below that are suitable for calculating the transformation.

1. In the first method, the filter bank is applied directly to the input speech frame s = (s (O), ..., s (N-1)) ^T and the two filter outputs are subsampled by a factor of two. This results in a set of approximation coefficients A ₁ = (A ₁ (O), ... A ₁ (N / 2-1)) ^T at the low-pass filter output and a set of detail coefficients D ₁ = (D ₁ (O), at the high-pass filter output. ..D ₁ (N / 2-1)) ^T. This method is now always applied recursively to the approximation coefficients of the previous stage, whereby in the last stage L the vector DWT ( s ) = ( D ₁ ^T , D ₂ ^T ..., D _L ^T , A _L ^T ) ^T with a total of N coefficients.

2. Like the first, the second variant for calculating the transformation is based on a filter bank decomposition. The difference is that the filter outputs are no longer subsampled. This results in vectors of length N after each stage and an output vector with a total of (L + 1) N coefficients after the last stage. In order to obtain the resolution properties of the wavelet transformation, the filter impulse responses for each stage must be obtained by oversampling by a factor of two from the previous stage. In the first stage, the same filters are used as for variant 1. Due to the overall higher effort - higher redundancy in the display in the image area - compared to variant 1, an improvement in the performance of the method can be achieved.

In order to avoid edge effects due to the filter length M, the M 2 ^L-2 past and the M 2 ^L-2 future samples of the speech frame are also taken into account and the filter impulse responses - as far as possible - centered around the temporal origin. This increases the algorithmic delay of the method in principle by M 2 ^L-2 samples. If this is to be avoided, the input frame can alternatively be continued periodically or symmetrically.

First, the frame energies E ₁ ..., E _{L of} the detail coefficients D ₁ , ..., D _L and the frame energy E _{L + 1 of} the approximation coefficients A _{L are} calculated by the processor 32.

The total energy of the frame E _tot can now be determined efficiently by summing all partial energies if the underlying wavelet basis is orthogonal. All energy values are represented in the logarithmic range.

For the pause detection in the circuit 34, the frame energy E _{tot is} compared with a fixed threshold T _{1 in} order to identify frames with very low energy. For this purpose, a binary decision _variable f _{sil is set} according to the following formula:

In order to obtain a measure for stationary or unsteady frames when stationary frames are detected, k is the difference measure for each frame Δ ( k ) = 1 L i = 1 L ( E i ( k ) - E i ( k -1) ) 2nd calculated, into which the frame energies of the detail coefficients of all levels flow. The binary decision variable f _stat is now set using the threshold T ₂ taking into account the last K frames:

berechnet, wobei die Zeitkonstante α durch 0<α<1 begrenzt ist. Danach werden die P Subrahmenenergien ∈(k,1) Q2 ,...,∈(k,P) Q2 aus den Detailkoeffizienten D ₂ bestimmt und mit Hilfe der festen Schwellen T₃ und T₄ jeweils eine binäre Entscheidungsvariable f_Q1 für die Stufe Q1 sowie f_Q2 für die Stufe Q2 gemäß den folgenden beiden Formeln bestimmt:

When detecting background noise in circuits 36 and 37, the goal is to obtain a decision criterion that is insensitive to the current level of background noise. The properties of the DWT or wavelet transformation circuit 30 are used efficiently for this by considering the detail coefficients D _Q1 in the coarse time interval N and the detail coefficients D _Q2 in the finer time interval N / P. P denotes the number of subframes, Q1 a level for coarse and Q2 a level for fine time resolution, whereby the relationships Q1, Q2∈ {1, L} and Q1> Q2 must apply. In advance, an estimate B _i , i∈ {Q1, Q2} for the current level of the background noise is carried out for both stages

calculated, the time constant α being limited by 0 <α <1. Then the P subframe energies ∈ ( k ,1) Q2 , ..., ∈ ( k , P ) Q2 determined from the detailed coefficients D ₂ and, using the fixed thresholds T ₃ and T _4, each determine a binary decision variable f _Q1 for stage Q1 and f _Q2 for stage Q2 according to the following two formulas:

In decision logic 42, using equations (1), (3), (5) and (6), the preliminary result vad ^{(pre) of} the automatic speech activity recognizer is determined by the logical combination vad (pre) =! (f sil | f Q1 & f Q2 & f stat )), won, where '!, |, &' do not denote the logical operators ', or, and'.

Further stages Q3, Q4, ..., etc. can also be defined, for which background noise detection can be carried out in the same way. Further binary decision _parameters f _Q3 , f _Q4 , ... are then to be defined, which are also to be taken into account in equation (7).

The temporal smoothing takes place in the circuit 44. In order to take into account the long-term stationarity of speech, the preliminary decision of the VAD is smoothed over time in a post-processing stage. If the number of the last frames marked contiguously as active exceeds the value C _B , as long as vad ^(pre) = 0, a maximum of C _H active frames are appended. The final decision vad∈ {0,1} of the voice activity recognizer is now made.

List of reference numbers

1

entrance

2.3

cables

4th

switch

5

Voice activity detection module or circuit

6

Feedback line

7.8

Lines or outputs of the switch 4

9

Speech coder

10th

Background noise encoder

11.12

cables

13

switch

14.15

cables

16

Transmission channel

17.18

cables

19th

switch

20.21

cables

22

Speech decoder

23

Background noise decoder

24.25

cables

26

switch

27

exit

28

Segmenter

29,31,33

cables

30th

Wavelet transform circuit 32 processor

34

Phase detector

35

Circuit for determination for the Stationarity measure

36.37

Background detector

38-41

cables

42

Decision logic

43

management

44

Smoothing circuit

45

exit

Claims (6)

Method for automatic speech activity recognition based on the wavelet transformation, characterized in that that a voice activity detection circuit or module (5) for controlling a speech encoder (7) and a speech decoder (22) and for controlling a background noise encoder (10) and a background noise decoder (23) is used for the source-controlled reduction of the average transmission rate, after segmentation of a speech signal, a wavelet transformation is calculated for each frame, from which a set of parameters is determined, from which a set of binary decision variables in a computing circuit or processor (32) that control decision logic (42) is calculated with the aid of fixed thresholds, the result of which, after smoothing the time, provides a statement "language available / no language" for each frame. Method for recognizing speech activity according to claim 1, characterized in that that after the wavelet transformation, a set of energy parameters is determined for each segment from the transformation coefficients and compared with fixed threshold values, which results in binary decision variables with which the decision logic (42) is controlled, which outputs a preliminary result for each frame at the output. Method for recognizing speech activity according to one of Claims 1 or 2, characterized in that that the preliminary result for each frame, which is determined by the decision logic, is post-processed by means of temporal smoothing, whereby the final result "language available or no language" is formed for the current frame. Method for recognizing speech activity according to one of Claims 1 to 3, characterized in that that background noise detectors (36 and 37) are controlled with signals for the detection of background noise and the detail coefficients (D) in the coarse time interval (N) and detail coefficients (D2) in the finer time interval (N / P) are analyzed, where P represents the number of subframes and the relationships Q1, Q2∈ {1, L} and Q1> Q2 apply. Circuit arrangement for carrying out the method for voice activity detection according to one of the claims 1 to 4, characterized in that that the signals of the input language reach the input (1) of a switch (4), that a voice activity detection circuit or module (5) is connected to the input (1), the output of which controls the said switch (4), a further switch (13) and is also connected to a transmission channel (16), that the output of the switch (4) is connected via lines (7 or 8) to a speech encoder (9) or to a background noise encoder (10), the outputs of which are connected via lines (11 or 12) to the inputs of the switch (13 ) are connected, the output of which is connected via a line (15) to the input of the transmission channel (16), which on the one hand is connected to a further changeover switch (19) and on the other hand via a line (18) to control the changeover switch (19) and to control a switch (26) arranged at the output (27) is connected, and that between the two switches (19 and 26) a speech decoder (22) and a decoder (23) for background noise is arranged. Circuit arrangement for performing the method according to one of the claims 1 to 4, characterized in that that the input (1) is connected to a segmentation circuit (28), the output of which is connected via a line (29) to a wavelet transformation circuit (30) which is connected to the input of a computing circuit or a processor (32) for calculating the Energy quantities is connected that the output of the processor (32) is connected via a line (33) in parallel to a pause detector (34), to a circuit for calculating a stationarity measure (35), to a first background detector (36) and to a second background detector (37) , that the outputs of said circuits (34 to 37) are connected to a decision logic (42), the output of which is connected to a smoothing circuit (44) for temporal smoothing, and that the output of the smoothing circuit (44) is also the output (45) of the voice activity detector.

Images