EP0714088A1 - Voice activity detection - Google Patents

Abstract

The detector calculates autocorrelation coefficients, R(k) for a signal. A first vector (RO) is composed from a first series, K = 0 ... (N-q), and a second vector (Rq) is formed from the components, k = q ...N. which are shifted by q relative to the first. The first vector is subtracted from the second to yield a difference vector, delta, R, from which a first standard vector is obtained. A reduced standard is obtained by dividing the first standard vector by a reduction value to give a second indicator. The reduction value is calculated from the energy of the audio signal or the sum of the audio energy and a bottom value, C. The linear combination of the present and previous value give a third indicator and these are measured against a threshold.

Description

The field of the invention is that of the detection of voice activity in an audio signal.

In the presence of an audio signal which often comes from a microphone, it is sometimes necessary to know whether this signal contains speech or whether it only contains noise.

Indeed, the detection of voice activity will often condition certain processing that the audio signal is likely to undergo. Among the typical applications that should be activated in the presence of a speech signal, one can identify speech recognition, echo cancellation or the recording function.

On the contrary, if we consider a telephony signal where only speech represents useful information, it is now common in the radiocommunication field not to transmit this signal if it only includes noise, this is commonly known as discontinuous transmission.

Thus, solutions have already been proposed in an attempt to detect voice activity in an audio signal.

A first solution consists in monitoring the evolution of the signal energy. If it increases rapidly, it can correspond to the appearance of a vocal activity but it can also correspond to a variation of the ambient noise. It follows that this method, although very simple to implement does not appear to be very reliable in relatively noisy environments as is the case for example in a motor vehicle.

Many other solutions are also known which have been developed to overcome the lack of reliability of the previous one. This is particularly the case for those which implement a Fourier transform of the audio signal to measure the spectral distance separating it from an averaged noise signal which is updated in the absence of any vocal activity. This is also the case for methods using an analysis of the signal in sub-bands, methods which are close to those using a Fourier transform. This is also the case for methods using cepstral analysis.

These are much more complex techniques which, although they bring a gain in terms of reliability, do not however give complete satisfaction on this point.

Solutions are also known which take advantage of a certain periodicity of speech, among which is that described in patent application EP 0 123 349. Indeed, voiced sounds all have a determined periodicity while the noise is normally aperiodic or well presents a periodicity distinct from that of speech.

We can therefore search for the value of this determined periodicity (or "pitch" in English) to recognize the presence of voiced sounds.

To do this, the autocorrelation coefficients of the audio signal are generally calculated to find the second maximum of these coefficients, the first maximum representing the energy. This is again a relatively complex technique which does not give complete satisfaction in terms of reliability.

The present invention therefore provides a solution for detecting voice activity which provides acceptable reliability for reduced complexity.

• des moyens pour calculer les coefficients d'autocorrélation de ce signal,
• des moyens pour identifier un premier vecteur d'autocorrélation ayant pour composantes une première série de coefficients d'autocorrélation,
• des moyens pour identifier un second vecteur d'autocorrélation ayant pour composantes une deuxième série de coefficients d'autocorrélation décalée par rapport à la première série d'une valeur de décalage prédéterminée,
• des moyens pour soustraire le premier vecteur d'autocorrélation du second vecteur d'autocorrélation afin d'obtenir un vecteur de différentiation,
• des moyens pour calculer une norme de ce vecteur de différentiation, cette norme représentant un premier indicateur d'activité vocale.

According to the invention, a device for detecting voice activity in an audio signal comprises:

• means for calculating the autocorrelation coefficients of this signal,
• means for identifying a first autocorrelation vector having as components a first series of autocorrelation coefficients,
• means for identifying a second autocorrelation vector having as components a second series autocorrelation coefficients offset from the first series by a predetermined offset value,
• means for subtracting the first autocorrelation vector from the second autocorrelation vector in order to obtain a differentiation vector,
• means for calculating a norm of this differentiation vector, this norm representing a first indicator of vocal activity.

In addition, the device comprises reduction means for establishing a reduced standard by dividing the standard of the differentiation vector by a reduction value, this reduced standard representing a second voice activity indicator.

For example, the reduction value is equal to the energy of the signal or it is equal to the sum of the energy of the signal and a compression constant.

According to an additional characteristic of the device, it comprises means for smoothing one of these voice activity indicators to produce a linear combination of the present value of this indicator and its previous value, this linear combination representing a third voice activity indicator.

Furthermore, the device comprises decision means for producing a voice activity signal if one of these indicators exceeds a detection threshold.

It may be advantageous to establish this detection threshold from the energy of the audio signal in the absence of a voice activity signal.

In addition, an advantageous solution consists in choosing the sum of the absolute values of the components of the differentiation vector as the norm of this vector.

• calcul des coefficients d'autocorrélation de ce signal,
• identification d'un premier vecteur d'autocorrélation ayant pour composantes une première série de coefficients d'autocorrélation,
• identification d'un second vecteur d'autocorrélation ayant pour composantes une deuxième série de coefficients d'autocorrélation décalée par rapport à la première série d'une valeur de décalage prédéterminée,
• soustraction du premier vecteur d'autocorrélation du second vecteur d'autocorrélation afin d'obtenir un vecteur de différentiation,
• calcul d'une norme du vecteur de différentiation, cette norme représentant un premier indicateur d'activité vocale.

The invention also relates to a method for detecting voice activity in an audio signal, comprising the following operations:

• calculation of the autocorrelation coefficients of this signal,
• identification of a first autocorrelation vector having as components a first series of autocorrelation coefficients,
• identification of a second autocorrelation vector having as components a second series of autocorrelation coefficients offset with respect to the first series by a predetermined offset value,
• subtraction of the first autocorrelation vector from the second autocorrelation vector in order to obtain a differentiation vector,
• calculation of a standard of the differentiation vector, this standard representing a first indicator of vocal activity.

The present invention will now appear more clearly in the context of an exemplary embodiment given by way of illustration with reference to the appended figure which represents the flow of operations carried out by the voice activity detection device.

We place ourselves in the case where an audio signal is digital in nature, that is to say that it is in the form of a series of samples which correspond to the value of the signal at successive instants which repeat at the rate of a sampling frequency.

When the signal to be analyzed is analog in nature, if it comes from a microphone for example, it is first subjected to an analog-digital converter which operates at the rate of this sampling frequency to produce the audio signal .

The audio signal being digital, it seems natural to realize the voice activity detection device by means of a digital signal processor. This processor can of course be used for other purposes.

It is therefore understood that this detection device will not be described in its structure because it implements elementary operations well known to those skilled in the art such as additions, multiplications, comparisons. It is therefore a functional description which has been retained, because it seems far preferable to explain the implementation of the invention with the greatest clarity.

With reference to the single figure, the device therefore receives the audio signal and we consider a series of samples S (i) where i varies from 0 to N.

The first operation performed by the device is the calculation of the autocorrelation coefficients R (k) of the signal for all the values of k between O and N:

From these autocorrelation coefficients R (k) we can define a first R₀ and a second R _q autocorrelation vectors by considering in addition an offset value q which is a strictly positive integer. The first autocorrelation vector R₀ has as components the (N-q + 1) first autocorrelation coefficients R (k): $${\text{<figure-callout id="0" label="R" filenames="imgf0001.png" state="{{state}}">R</figure-callout>}}_{\text{0}}\text{= (R (O), R (1), ..., R (Nq))}$$

The second autocorrelation vector R _q has as components the (N-q + 1) last autocorrelation coefficients R (k): $${\text{R}}_{\text{q}}\text{= (R (q), R (q + 1), ..., R (N))}$$

The detection device then calculates a differentiation vector ΔR by subtracting the first autocorrelation vector R₀ from the second autocorrelation vector R _q : $${\text{ΔR = R}}_{\text{q}}{\text{- <figure-callout id="0" label="R" filenames="imgf0001.png" state="{{state}}">R</figure-callout>}}_{\text{0}}$$

If we denote by ΔR (k) the (k + 1) th component of this differentiation vector, this then applies for all k between 0 and Nq: $$\text{ΔR (k) = R (k + q) - R (k)}$$

We can see that the first R₀ and second R _q autocorrelation vectors have no utility in themselves. They have been introduced for the simple purpose of clarifying the presentation. The important point is the calculation of the differentiation vector. Thus, this vector is defined by the value of these components as defined above.

Consequently, the detection device calculates a standard ∥ΔR∥ of the differentiation vector ΔR. Advantageously, this standard is equal to the sum of the absolute values of the components of the vector:

It goes without saying that the invention also applies if one chooses to choose another standard such as, in particular, the Euclidean standard or the maximum value of the absolute values of each of the components.

This standard, whatever it is, constitutes a first indicator of vocal activity.

A first option is to compare this indicator with a threshold to establish that there is presence of voice activity in the audio signal if the indicator is greater than the threshold.

According to a second option, the detection device calculates a reduced standard P by dividing the standard ∥ΔR∥ of the differentiation vector by a reduction value. For example, this reduction value can be chosen equal to the energy R (0) of the audio signal, which will tend to compress the dynamics of the ∥ΔR∥ standard. Another solution which provides its own advantages consists in assigning to this reduction value the sum of the energy R (0) of the audio signal and of a constant which will be called the floor value C.

This reduced standard P, in any event constitutes a second indicator of vocal activity which can also be compared to a threshold to establish the absence or the presence of vocal activity in this signal.

_i de cette norme réduite sera une combinaison linéaire de la valeur lissée $\overline{\text{P}}$

_i-1 de la norme réduite P_i-1 associée à la série précédente et de cette norme réduite P_i : $${\overline{\text{P}}}_{\text{i}}\text{= α}{\overline{\text{P}}}_{\text{i-1}}{\text{+ βP}}_{\text{i}}$$

According to a third option, the detection device smoothes this reduced standard. Thus, if we consider several successive series of N samples of the audio signal, a reduced standard P _i corresponds to the i th series. The smoothed value $\overline{\text{P}}$

_i of this reduced norm will be a linear combination of the smoothed value $\overline{\text{P}}$

_i-1 of the reduced standard P _i-1 associated with the previous series and of this reduced standard P _i : $${\overline{\text{P}}}_{\text{i}}\text{= α}{\overline{\text{P}}}_{\text{i-1}}{\text{+ βP}}_{\text{i}}$$

We can choose α and β so that their sum is equal to unity.

₀ à l'aide d'une constante quelconque, 0 par exemple.In addition, it is necessary to initialize $\overline{\text{P}}$

₀ using any constant, 0 for example.

_i constitue un troisième indicateur d'activité vocale que l'on peut aussi comparer à un seuil pour établir si le signal audio présente ou non une activité vocale.This smoothed value $\overline{\text{P}}$

_i constitutes a third voice activity indicator which can also be compared to a threshold to establish whether or not the audio signal has voice activity.

Whatever the voice activity indicator selected, the detection device therefore compares it to a detection threshold T. The simplest solution consists in assigning a constant value to this detection threshold.

However, an advantageous solution consists in adapting this threshold to the level of the reduced standard P when the audio signal is devoid of voice activity.

We can therefore calculate the average value of the reduced standard over several successive series of samples of the audio signal for which no voice activity has been detected and multiply this average value by a constant coefficient to obtain the detection threshold T. It s 'This is a technique similar to that of smoothing well known to those skilled in the art and it will therefore not be more detailed.

In addition to the actual detection device, the invention naturally relates to the method for detecting voice activity which is implemented by this device.

By way of digital application and to present a concrete case of use of the invention, we will take as illustration the pan-European digital cellular radiocommunication system called GSM system. In this system the analog signal to be processed is sampled at the frequency of 8 kHz. The samples thus obtained are grouped in series of 160 which therefore each correspond to 20 ms.

Thus, N, the number of samples, is worth 160 and it will be advantageous to choose to set the offset value q equal to unity.

The components of the differentiation vector are then written for all k between 1 and 160: $$\text{ΔR (k) = R (k + 1) - R (k)}$$

The norm of this vector can therefore be written:

Claims (9)

Voice activity detection device in an audio signal comprising: means for calculating the autocorrelation coefficients (R (k)) of this signal, means for identifying a first autocorrelation vector (R₀) having as components a first series (k = 0, ..., Nq) of autocorrelation coefficients (R (k)), means for identifying a second autocorrelation vector (R _q ) having as components a second series (k = q, ..., N) of autocorrelation coefficients (R (k)) offset with respect to said first series a predetermined offset value (q), means for subtracting said first autocorrelation vector (R₀) from said second autocorrelation vector (R _q ) in order to obtain a differentiation vector (ΔR), - Means for calculating a norm (∥ΔR∥) of said differentiation vector, this norm representing a first indicator of vocal activity. Device according to claim 1, characterized in that it further comprises reduction means for establishing a reduced standard by dividing said standard (∥ΔR∥) of the differentiation vector by a reduction value, this reduced standard representing a second indicator voice activity. Device according to claim 2 characterized in that said reduction value is equal to the energy of the audio signal. Device according to Claim 2, characterized in that the said reduction value is equal to the sum of the energy of the audio signal and of a floor value (C). Device according to any one of claims 1 to 4, characterized in that it comprises means for smoothing one of said activity indicators voice to produce a linear combination of the present value of this indicator and its previous value, said linear combination representing a third voice activity indicator. Device according to any one of claims 1 to 5, characterized in that it comprises decision means for producing a voice activity signal if one of said indicators exceeds a detection threshold. Device according to claim 6, characterized in that said detection threshold is established from the value of the reduced standard of said audio signal in the absence of said voice activity signal. Device according to any one of claims 1 to 7, characterized in that said standard (∥ΔR∥) of the differentiation vector is equal to the sum of the absolute values of the components of this vector. Method for detecting voice activity in an audio signal, comprising the following operations: - calculation of the autocorrelation coefficients (R (k)) of this signal, - identification of a first autocorrelation vector (R₀) having as components a first series (k = 0, ..., Nq) of autocorrelation coefficients (R (k)), - identification of a second autocorrelation vector (R _q ) having as components a second series (k = q, ..., N) of autocorrelation coefficients (R (k)) offset with respect to said first series d 'a predetermined offset value (q), subtraction of said first autocorrelation vector (R₀) from said second autocorrelation vector (R _q ) in order to obtain a differentiation vector (ΔR), - Calculation of a standard (∥ΔR∥) of said differentiation vector, this standard representing a first indicator of vocal activity.

Images