EP2518724A1 - Microphone/headphone audio headset comprising a means for suppressing noise in a speech signal, in particular for a hands-free telephone system - Google Patents

Abstract

The headset has a physiological sensor (18) provided in an insulating cushion of a closed shell of one of earpieces. A microphone set has front and rear microphones (20, 22) placed on the shell and aligned to form a linear array in a main direction toward a mouth of a wearer. A combiner-and-phase shifter (56) reduces non-frequency noise of speech signal and includes a combiner for applying a delay to a signal delivered by the microphone and subtracting a signal delivered by another microphone from the delay signal so as to remove noise from a near speech signal uttered by the wearer.

Description

The invention relates to a headset type microphone / headset combined.

Such a headset can in particular be used for communication functions such as "hands-free" telephony functions, in addition to listening to an audio source (music for example) coming from a device on which the headphones are connected. .

In the communication functions, one of the difficulties is to ensure sufficient intelligibility of the signal picked up by the microphone ("microphone"), that is to say the speech signal of the close speaker (the helmet wearer). . The helmet can indeed be used in a noisy environment (metro, busy street, train, etc.), so that the microphone will not only capture the speech of the wearer of the helmet, but also the surrounding noise.

The wearer can be protected from these noises by the helmet, especially if it is a model with closed headphones isolating the ear from the outside, and even more if the headset is provided with an "active control of noise". On the other hand, the distant speaker (the one at the other end of the communication channel) will suffer from the noise picked up by the microphone, coming to overlap and interfere with the speech signal of the nearby speaker (the helmet wearer).

In particular, some speech formers essential to the understanding of the voice are often embedded in noise components commonly encountered in the usual environments, components which are mainly concentrated in the low frequencies.

It has been proposed to collect certain vocal vibrations by means of a physiological sensor applied against the cheek or the temple of the helmet wearer. Indeed, when a person makes a voiced sound (that is, a speech component whose production is accompanied by a vibration of the vocal cords), a vibration propagates from the vocal cords to the pharynx and to the bucco-nasal cavity, where it is modulated, amplified and articulated. The mouth, the soft palate, the pharynx, the sinuses and the nasal fossae serve as a sounding board for this voiced sound and, their walls being elastic, they vibrate in turn, and these vibrations are transmitted by internal bone conduction and are perceptible at level of the cheek and temple.

These vocal vibrations at the level of the cheek and the temple have the characteristic of being, by nature, very little corrupted by the surrounding noise: indeed, in the presence of external noises, the fabrics of the cheek and the temple do not vibrate. almost no, whatever the spectral composition of the external noise.

Moreover, because of the filtering generated by the propagation of vibrations to the temple, the signal collected by the physiological sensor is usable only in the low frequencies. But as the noises generally encountered in a usual environment (street, metro, train ...) are mainly concentrated in the low frequencies, the physiological sensor delivers a signal naturally devoid of noise component noise (which is not possible with a classic microphone).

• deux écouteurs comportant chacun un transducteur de reproduction sonore d'un signal audio ;
• un capteur physiologique, apte à venir en contact avec la joue ou la tempe du porteur du casque pour y être couplé et capter les vibrations vocales non acoustiques transmises par conduction osseuse interne, ce capteur physiologique délivrant un premier signal de parole ;
• un ensemble microphonique, comprenant au moins un microphone apte à capter les vibrations vocales acoustiques transmises par voie aérienne depuis la bouche du porteur du casque, cet ensemble microphonique délivrant un second signal de parole ; et
• des moyens de mixage, pour combiner le premier signal de parole et le second signal de parole, et donner en sortie un troisième signal de parole représentatif de la parole émise par le porteur du casque.

The JP 2000-261534 A describes such a headset handset comprising:

• two earphones each having a sound reproduction transducer of an audio signal;
• a physiological sensor, adapted to come into contact with the cheek or the temple of the helmet wearer to be coupled thereto and to capture the non-acoustic vocal vibrations transmitted by internal bone conduction, this physiological sensor delivering a first speech signal;
• a microphone assembly, comprising at least one microphone capable of capturing the acoustic vocal vibrations transmitted by air from the mouth of the helmet wearer, this microphone assembly delivering a second speech signal; and
• mixing means, for combining the first speech signal and the second speech signal, and outputting a third speech representative signal of the speech transmitted by the helmet wearer.

The EP 0 683 621 A2 for its part, it describes more precisely how to integrate the physiological sensor and the external microphone into one and the same ear canal.

Of course, the signal collected by the physiological sensor is not strictly speaking speech since speech is not only formed of voiced sounds, it contains components that are not born at the level of the vocal cords: the frequency content is for example much richer with the sound coming from the throat and emitted through the mouth. Moreover, the internal bone conduction and the crossing of the skin has the effect of filtering certain vocal components, which makes that the signal delivered by the physiological sensor is exploitable only in the lowest part of the spectrum. That is why this signal is supplemented by another signal, delivered by a conventional microphonic sensor, to which it is combined.

The general problem of the invention is, in such a context, to deliver to the remote speaker a voice signal representative of the speech transmitted by the near speaker, a signal which is freed from the parasitic components of external noise present in the environment of the close speaker .

An important aspect of this problem is the need to restore a natural and intelligible speech signal, that is to say, undistorted and whose range of useful frequencies is not amputated by the combination of signal processing from operating sensors vibrations that are different in nature and transmitted by different paths.

Another aspect of the invention resides in the ability to efficiently use the signal from the physiological sensor to control various signal processing functions. This signal makes it possible to access new information concerning the content of the speech, which will then be used for the denoising as well as for various auxiliary functions that will be explained below, in particular the calculation of a cutoff frequency of a dynamic filter.

• le capteur physiologique est incorporé à un coussinet circumaural d'une coque de l'un des écouteurs ;
• l'ensemble microphonique comprend deux microphones placés sur la coque de l'un des écouteurs ;
• les deux microphones sont alignés en un réseau linéaire suivant une direction principale dirigée vers la bouche du porteur du casque ; et
• il est prévu des moyens de réduction de bruit non fréquentielle du second signal de parole, comprenant un combineur apte à appliquer un retard au signal délivré par l'un des microphones et à soustraire ce signal retardé du signal délivré par l'autre microphone, de manière à opérer un débruitage du signal de parole proche émis par le porteur du casque.

To solve these problems, the invention proposes a microphone / headset of the type described above as taught by the JP 2000-261534 A and wherein, in a characteristic manner of the invention:

• the physiological sensor is incorporated in a circumaural pad of a shell of one of the earphones;
• the microphone set includes two microphones placed on the shell of one of the earphones;
• the two microphones are aligned in a linear array in a principal direction directed towards the mouth of the wearer of the helmet; and
• there is provided non-frequency noise reduction means of the second speech signal, comprising a combiner able to apply a delay to the signal delivered by one of the microphones and to subtract this delayed signal from the signal delivered by the other microphone, way to operate a denoising of the near speech signal emitted by the wearer of the helmet.

Advantageously, the microphone / headset combination comprises low-pass filtering means of the first speech signal before combination by the mixing means, and / or high-pass filtering means of the second speech signal before denoising and combination by the means. mixing. These low-pass and / or high-pass filtering means comprise an adjustable cutoff frequency filter, and the headset comprises means for calculating the cutoff frequency, operating as a function of the signal delivered by the physiological sensor. The means for calculating the cutoff frequency may in particular comprise means for analyzing the spectral content of the signal delivered by the physiological sensor, able to determine the cutoff frequency as a function of the relative levels of the signal / noise ratio evaluated in a plurality of distinct frequency bands of the signal delivered by the physiological sensor.

Preferably, the denoising means of the second speech signal are non-frequency noise reduction means with, in a particular embodiment of the invention, the microphone assembly which comprises two microphones, and the noise reduction means. non-frequency which comprise a combiner able to apply a delay to the signal delivered by one of the microphones and to subtract this delayed signal from the signal delivered by the other microphone.

In particular, the two microphones can be aligned in a linear array in a main direction directed towards the mouth of the wearer of the helmet.

Also preferably, there is provided denoising means of the third speech signal delivered by the mixing means, including frequency noise reduction means.

For this purpose, and according to an original aspect of the invention, there is provided input receiving means, and operating an intercorrelation between the first and the third speech signal, and outputting a speech presence probability signal. function of the result of the intercorrelation. The denoising means of the third speech signal receive as input this speech presence probability signal for selectively: i) making a noise reduction differentiated according to the frequency bands as a function of the value of the speech presence probability signal, and ii) perform a maximum noise reduction on all the frequency bands in the absence of speech.

In addition, there may be provided post-processing means capable of selectively frequency band equalizing in the part of the spectrum corresponding to the signal collected by the physiological sensor. These means determine an equalization gain for each of the frequency bands, this gain being calculated from the respective frequency coefficients of the signals delivered by the microphone (s) and signals delivered by the physiological sensor, considered in the frequency domain. They also operate smoothing on a plurality of successive signal frames of the calculated equalization gain.

• La Figure 1 illustre de façon générale le casque de l'invention, posé sur la tête d'un utilisateur.
• La Figure 2 est un schéma d'ensemble, sous forme de blocs fonctionnels, expliquant la manière dont est réalisé le traitement du signal permettant de délivrer en sortie un signal débruité représentatif de la parole émise par le porteur du casque.
• La Figure 3 est une représentation spectrale amplitude/fréquence illustrant le calcul d'intercorrélation servant à évaluer une probabilité de présence de parole.
• La Figure 4 est une représentation spectrale amplitude/fréquence illustrant le traitement final d'égalisation automatique opéré après la réduction de bruit.

An embodiment of the device of the invention will now be described with reference to the appended drawings in which the same reference numerals designate identical or functionally similar elements from one figure to another.

• The Figure 1 generally illustrates the headset of the invention, placed on the head of a user.
• The Figure 2 is a block diagram, in the form of functional blocks, explaining the manner in which the signal processing is carried out making it possible to output a speechless signal representative of the speech transmitted by the helmet wearer.
• The Figure 3 is an amplitude / frequency spectral representation illustrating the intercorrelation calculation used to evaluate a probability of presence of speech.
• The Figure 4 is an amplitude / frequency spectral representation illustrating the final automatic equalization processing performed after the noise reduction.

On the Figure 1 , the reference 10 generally designates the helmet according to the invention, which comprises two atria 12 joined by a hoop. Each of the atria is preferably constituted by a closed shell 12, housing a sound reproduction transducer, applied around the ear of the user with the interposition of an insulating pad 16 isolating the ear from the outside.

This helmet is provided with a physiological sensor 18 for collecting the vibrations produced by a voiced signal emitted by the wearer of the helmet, and which can be picked up at the level of the cheek or the temple. The sensor 18 is preferably an accelerometer integrated in the pad 16 so as to be applied against the cheek or the temple of the user with the closest possible coupling. The physiological sensor may in particular be placed on the inside of the skin of the pad so that, once the helmet is in place, the physiological sensor is applied against the cheek or the temple of the user under the effect of a slight pressure resulting from the crushing of the material of the pad, with only the interposition of the skin of the pad.

The headset also comprises a microphone array or antenna, for example two omnidirectional microphones 20, 22, placed on the shell of the earpiece 12. These two front and rear mics 22 and 20 are omnidirectional microphones arranged relative to each other. other so that their alignment direction 24 is approximately directed towards the mouth 26 of the helmet wearer.

The Figure 2 is a block diagram showing the different blocks and functions implemented by the method of the invention as well as their interactions.

The method of the invention is implemented by software means, which can be broken down and schematized by a number of illustrated blocks 30 to 64 Figure 2 . These processes are implemented in the form of appropriate algorithms executed by a microcontroller or a digital signal processor. Although, for the sake of clarity, these various treatments are presented in the form of separate blocks, they implement common elements and correspond in practice to a plurality of functions globally executed by the same software.

In this figure, we find the physiological sensor 18 and the two omnidirectional microphones 20 and 22 behind. The reference 28 also designates the sound reproduction transducer placed inside the hull. of the earpiece. These various elements deliver signals that are processed by the block referenced 30, which can be coupled to an interface 32 to the communication circuits (telephone circuits) and receives at the input E the sound intended to be reproduced by the transducer 28 (speech of the remote speaker during a telephone call, music source out periods of telephone communication), and delivers on the output S a signal representative of the speech of the next speaker, that is to say, the wearer of the headset.

The signal to be reproduced applied to the input E is a digital signal converted into analog by the converter 34, then amplified by the amplifier 36 for reproduction by the transducer 28.

The manner in which the speech signal representative of the speech of the near speaker is produced from the respective signals collected by the physiological sensor 18 and the microphones 20 and 22 will now be described.

The signal collected by the physiological sensor 18 is a signal mainly comprising components in the lower region of the sound spectrum (typically 0-1500 Hz). As explained above, this signal is naturally non-noisy.

The signals collected by the microphones 20, 22 will be used mainly for the high spectrum (above 1500 Hz), but these signals are strongly noisy and it will be essential to carry out a strong denoising treatment to eliminate the components of parasitic noise, the level of which may be such, in certain environments, that they completely obscure the speech signal picked up by these microphones 20, 22.

The first stage of the treatment is an anti-echo treatment, applied to the signals of the physiological sensor and the microphones.

Indeed, the sound reproduced by the transducer 28 is captured by the physiological sensor 18 and the microphones 20, 22, generating an echo that disrupts the operation of the system, and must be eliminated upstream.

This anti-echo treatment is implemented by the blocks 38, 40 and 42, each of these blocks receiving on a first input the signal emitted by the sensor 18, 20 or else 22 and on a second input the signal reproduced by the transducer. 28 (echo generator signal), and outputs, for further processing, a signal whose echo has been eliminated.

The anti-echo treatment is for example carried out by an adaptive algorithm treatment such as that described in FIG. FR 2 792 146 A1 (Parrot SA), which can be referred to for more details. This is an echo cancellation or AEC technique consisting in dynamically defining a compensation filter modeling the acoustic coupling between the transducer 28 and the physiological sensor 18 (or the microphone 20, or the microphone 22, respectively) by a linear transformation between the signal reproduced by the transducer 28 (that is to say the signal E applied at the input of the blocks 38, 40 or 42) and the echo picked up by the physiological sensor 18 (or the microphone 20 or 22). This transformation defines an adaptive filter which is applied to the reproduced incident signal, and the result of this filtering is subtracted from the signal collected by the physiological sensor 18 (or the microphone 20 or 22), which has the effect of canceling the major part acoustic echo.

This modeling is based on the search for a correlation between the signal reproduced by the transducer 28 and the signal collected by the physiological sensor 18 (or the microphone 20 or 22), that is to say on an estimate of the impulse response. the coupling constituted by the body of the earphone 12 supporting these various elements.

The processing is performed in particular by an adaptive APA ( Affine Projection Algorithm ) algorithm , which provides fast convergence, well suited to hands-free applications with intermittent speech rate and a level that can quickly vary.

Advantageously, the iterative algorithm is executed with a variable pitch, as described in FIG. FR 2 792 146 A1 supra. With this technique, the pitch μ varies continuously according to the energy levels of the signal picked up by the microphone, before and after filtering. This step is increased when the energy of the sensed signal is dominated by the energy of the echo, and, conversely, reduced when the energy of the signal picked up is dominated by the energy of the background noise and / or the speech from the remote speaker.

The signal collected by the physiological sensor 18 after the anti-echo processing by the block 38 will be used as the input signal of a block 44 for calculating a cutoff frequency FC.

The next step consists in filtering the signals, with a low-pass filter 48 for the signal of the physiological sensor 18 and with a filter high pass 50, 52 for the signals collected by the microphones 20, 22, respectively.

These filters 48, 50 and 52 are preferably infinite impulse response type IIR (recursive filter) type digital filters, which have a relatively steep transition between the bandwidth and the rejected band.

Advantageously, these filters are adaptive filters whose cutoff frequency is variable and determined dynamically by the block 44.

This makes it possible to adapt the filtering to the particular conditions of use of the headset: the voice of the wearer when he speaks, more or less close coupling between the physiological sensor 18 and the cheek or the temple of the wearer, etc. The cut-off frequency FC, which is preferably the same for the low-pass filter 48 and the high-pass filters 50 and 52, is determined from the signal of the physiological sensor 18 after the anti-echo treatment 38. For this purpose, an algorithm calculates the signal-to-noise ratio for a plurality of frequency bands in a range between, for example, 0 and 2500 Hz (the noise level being given by a calculation of the energy in a higher frequency band, for example between 3000 and 4000 Hz, because it is known that in this zone the signal can only be noise, because of the properties of the component constituting the physiological sensor 18). The cutoff frequency chosen will correspond to the maximum frequency for which the signal / noise ratio exceeds a predetermined threshold, for example 10 dB.

The following step consists in operating, by means of block 54, a mix to reconstruct the complete spectrum with, on the one hand, the lower region of the spectrum given by the filtered signal of the physiological sensor 18 and, on the other hand, the top of the spectrum given by the filtered signal of the microphones 20 and 22 after passing through a combiner-phase shifter 56 for operating a denoising in this part of the spectrum. This reconstruction is performed by summing the two signals, which are applied synchronously to the mixing block 54 so as to avoid any deformation.

We shall now describe more precisely the manner in which the noise reduction is performed by the phase-shifter combiner 56.

The signal that we want to denoise (that is, the signal from the near speaker located in the upper part of the spectrum, typically the components of frequency greater than 1500 Hz) is derived from the two microphones 20, 22 disposed a few centimeters from each other on the shell 14 of one of the earphones of the helmet. As indicated, these two microphones are arranged relative to each other so that the direction 24 they define is approximately oriented in the direction of the mouth 26 of the helmet wearer. As a result, a speech signal emitted from the mouth will reach the microphone before 20 and then the rear microphone 22 with a delay, and therefore a substantially constant phase shift, while the ambient noise will be picked up without phase shift by the two microphones 20 and 22. (which are omnidirectional microphones), given the distance of sources of parasitic noise compared to the two microphones 20 and 22.

The noise reduction on the signals picked up by the microphones 20 and 22 is not operated in the frequency domain (as is often the case), but in the time domain, by means of the phase shifter combiner 56 which comprises a phase-shifter 58 applying a delay τ to the signal of the rear microphone 22 and a combiner 60 for subtracting this delayed signal from the signal from the microphone before 20.

Thus, a differential network of first-order microphones equivalent to a single virtual microphone whose directivity can be adjusted as a function of the value of τ, with 0 ≤ τ ≤ τ _A (τ _A being the value corresponding to the natural phase difference between the two microphones 20 and 22, equal to the distance between the two microphones divided by the speed of sound, a delay of about 30 microseconds for a spacing of 1 cm). A value τ = τ _A will give a cardioid directivity diagram, a value τ = τ _A / 3 a hypercardioid diagram, and a value τ = 0 a dipole diagram. An appropriate choice of this parameter can be achieved by attenuating about 6 dB on surrounding diffuse noises. For more details on this technique, we can for example refer to:

[1] M. Buck and M. Rößler, First Order Differential Microphones Arrays for Automotive Applications, Proceedings of the 7th International Workshop on Acoustic echo and Noise control (IWAENC), Darmstadt, 10-13 Sept 2001 .
We will now describe the operations performed on the overall signal (top and bottom of the spectrum) delivered at the output of the mixing means 54.

This signal is subjected by the block 62 to a frequency noise reduction.

Preferably, this frequency noise reduction is operated differently in the presence or absence of speech, by evaluating a probability p of absence of speech from the signal collected by the physiological sensor 18.

Advantageously, this probability of absence of speech is derived from the information given by the physiological sensor.

Indeed, as indicated above, the signal delivered by this sensor has a very good signal / noise ratio up to the cutoff frequency FC determined by the block 44. But beyond this cutoff frequency the signal / noise ratio is still good, and often better than that of the microphones 20 and 22. The sensor information is exploited by calculating (block 64) the frequency intercorrelation between the combined signal delivered by the mixing block 54 and the signal unfiltered physiological sensor, before filtering by the low-pass filter 48.

Thus, for each frequency f, for example between FC and 4000 Hz, and for each frame n, the following calculation is performed by block 64: $$\mathit{int}\mathit{}\mathit{e}\mathit{r}\mathit{VS}\mathit{o}\mathit{r}\mathit{r}\mathit{e}\mathit{l}\mathit{at}\mathit{t}\mathit{i}\mathit{o}\mathit{not}\left(\mathit{not}\mathit{f}\right)\mathit{=}{\mathit{\alpha }}_{\mathit{int}\mathit{}\mathit{e}\mathit{r}\mathit{vs}\mathit{o}\mathit{r}\mathit{r}}\mathit{•}\mathit{int}\mathit{}\mathit{e}\mathit{r}\mathit{VS}\mathit{o}\mathit{r}\mathit{r}\mathit{e}\mathit{l}\mathit{at}\mathit{t}\mathit{i}\mathit{o}\mathit{not}\left(\mathit{not}\mathit{-}\mathrm{1}\mathit{,}\mathit{f}\right)\mathit{+}\left(\mathit{1}\mathit{-}{\mathit{\alpha }}_{\mathit{int}\mathit{}\mathit{e}\mathit{r}\mathit{vs}\mathit{o}\mathit{r}\mathit{r}}\right)\mathit{•}\stackrel{\mathit{~}}{\mathit{S}\mathit{m}\mathit{i}\mathit{x}\left(\mathit{f}\right)}\mathit{\cdot }\stackrel{\mathit{~}}{\mathit{S}\mathit{at}\mathit{vs}\mathit{vs}\left(\mathit{f}\right)}$$

Smix ( f ) and smix ( f ) being the frequency (complex) vector representations, for the n- frame, respectively of the combined signal delivered by the mixing block 54, and of the signal of the physiological sensor 18.

To evaluate a probability of absence of speech, the algorithm searches for frequencies for which there is only noise (situation of absence of speech): on the spectrogram of the signal delivered by the mixing block 54 certain harmonics are embedded in the noise, while they stand out more on the signal of the physiological sensor.

The intercorrelation calculation by the formula described above produces a result whose figure 3 shows an example, in the frequency domain.

The peaks P ₁ , P ₂ , P ₃ , P ₄ , ... of this intercorrelation calculation indicate a strong correlation between the combined signal delivered by the mixing block 54, and the signal of the physiological sensor 18, and the Emergence of these correlated frequencies indicates the likely presence of speech for these frequencies.

To obtain a probability of absence of speech (block 66), the complementary value is considered: $$\mathit{AbsProbo}\left(\mathrm{not}\mathit{f}\right)\mathit{=}1\mathit{-}i\mathit{nterCorrelation}\left(\mathrm{not}\mathit{f}\right)\mathit{/}\mathit{coefficient\_normalisation}$$

The value coefficient_normalization makes it possible to regulate the distribution of the probabilities according to the value of intercorrelation, and to obtain values between 0 and 1.

• en l'absence probable de parole, la réduction de bruit est appliquée sur toutes les bandes de fréquences, c'est-à-dire que le gain maximal de réduction est appliqué de la même façon sur toutes les composantes du signal (puisque dans ce cas celui-ci ne contient vraisemblablement pas de composante utile) ;
• en revanche, en présence probable de parole, la réduction de bruit est une réduction de bruit fréquentielle appliquée sélectivement selon les différentes bandes de fréquences en fonction de la valeur p de la probabilité de présence de parole, selon un schéma classique, par exemple comparable à celui décrit dans le WO 2007/099222 A1 (Parrot ).

The probability p of absence of speech thus obtained is applied to the block 62 which operates on the signal delivered by the mixing block 54 a frequency noise reduction selectively with respect to a given threshold of the probability of absence of speech:

• in the probable absence of speech, the noise reduction is applied to all the frequency bands, ie the maximum reduction gain is applied in the same way to all the signal components (since in this case this one probably does not contain a useful component);
• on the other hand, in the probable presence of speech, the noise reduction is a frequency noise reduction applied selectively according to the different frequency bands as a function of the value p of the probability of presence of speech, according to a conventional scheme, for example comparable to the one described in WO 2007/099222 A1 (Parrot ).

The system that has just been described makes it possible to obtain excellent overall performance, typically of the order of 30 to 40 dB of noise reduction on the speech signal of the nearby speaker. By eliminating all the noises, especially the most troublesome (train, metro, etc.) that are concentrated in the low frequencies, this gives the impression to the distant speaker (the one with which the wearer of the headset is in communication ) that his interlocutor (the helmet wearer) is in a quiet room.

Finally, it is advantageous to apply to the signal a final equalization (block 68), especially on the low end of the spectrum.

Indeed, the low frequency content collected at the cheek or temple by the physiological sensor 18 is different from the low frequency content of the sound emitted by the mouth of the user, as it would be captured by a microphone located a few centimeters from the mouth, or even by the ear of an interlocutor. The use of the physiological sensor and the filtering described above certainly makes it possible to obtain a very good signal in terms of signal-to-noise ratio, but which may present for the interlocutor who hears it a tone a little deaf and unnatural.

To overcome this difficulty, it is advantageous to operate an equalization of the output signal with gains adjusted selectively on different frequency bands in the region of the spectrum corresponding to the signal collected by the physiological sensor. The equalization can be performed automatically, from the signal delivered by the microphones 20, 22, before filtering.

The Figure 4 shows an example, in the frequency domain (thus after Fourier transform) of the ACC signal produced by the physiological sensor 18, with respect to a MIC microphone signal that would be captured a few centimeters from the mouth.

In order to optimize the rendering of the signal collected by the physiological sensor, differentiated gains G ₁ , G ₂ , G ₃ , G ₄ ,... Are applied to different frequency bands of the part of the spectrum located in the low frequencies.

These gains are evaluated by comparing the signals picked up, in a common frequency band, by both the physiological sensor 18 and the microphones 20 and / or 22.

Specifically, the algorithm calculates respective Fourier transforms of the two signals, providing a series of frequency coefficients (expressed in dB) NormPhysioFreq_dB (i) and NormMicFreq_dB (i) respectively corresponding to the standard of the ^¡th Fourier coefficient physiological sensor signal and the standard of the ¡ ^th Fourier coefficient of the microphonic signal.

est positive, le gain qui sera appliqué sera inférieur à l'unité (négatif en dB) ; réciproquement si la différence est négative le gain à appliquer sera supérieur à l'unité (positif en dB).For each frequency coefficient of rank i , if the difference: $$\mathit{DifferenceFreq\_dB}\left(\mathit{i}\right)\mathit{=}\mathit{NormPhysioFreq}\mathit{\_}\mathit{d}\mathit{B}\left(\mathit{i}\right)-\mathit{NormMicFreq}\mathit{\_db}\left(\mathit{i}\right)\mathit{.}$$

is positive, the gain that will be applied will be less than unity (negative in dB); Conversely, if the difference is negative, the gain to be applied will be greater than unity (positive in dB).

If the gain were applied as is, the differences are not exactly constant from one frame to another, especially when it is not about voiced sounds, there would be significant variations of equalization in the timbre. To avoid these variations, the algorithm operates a smoothing of the difference, which makes it possible to refine the equalization: $$\mathit{Gain\_dB}\left(\mathit{i}\right)\mathit{=}\mathit{\lambda }\mathit{.}\mathit{Gain\_dB}\left(\mathit{i}\right)\mathit{-}\left(\mathit{1}\mathit{-}\mathit{\lambda }\right)\mathit{DifferenceFreq\_dB}\left(\mathit{i}\right)\mathit{.}$$

Plus the λ coefficient will be close to 1, less the information of the current frame will be taken into account for the calculation of the gain of the ^ith coefficient. Conversely, the closer the coefficient λ is to 0, the more instantaneous information will be taken into account. In practice, for the smoothing to be effective, we will take a value λ close to 1, for example λ = 0.99. The gain applied to each frequency band of the signal from the physiologic sensor will give, for the i ^th frequency modified: $$\mathit{NormPhysioFreq\_dB\_corrigée}\left(\mathit{i}\right)\mathit{=}\mathit{NormPhysioFreq\_dB}\left(\mathit{i}\right)\mathit{+}\mathit{Gain\_dB}\left(\mathit{i}\right)$$

It is this standard that will be used by the equalization algorithm.

The application of differentiated gains makes it possible to make the speech signal more natural in the lower part of the spectrum. A subjective study has shown that in a quiet environment and when such equalization is applied, the difference between a reference microphonic signal and the signal produced by the physiological sensor in the low end of the spectrum is practically imperceptible.

Claims (9)

An audio headset (10) of the combined microphone / headphone type, comprising: - two earphones (12) each having a transducer (28) for reproducing sound of an audio signal; a physiological sensor (18) able to come into contact with the cheek or the temple of the helmet wearer to be coupled thereto and to pick up the non-acoustic vocal vibrations transmitted by internal bone conduction, this physiological sensor delivering a first speech signal; - A microphone assembly, comprising at least one microphone (20, 22) adapted to capture the acoustic vocal vibrations transmitted by air from the mouth of the wearer of the headset, the microphone assembly delivering a second speech signal; and mixing means (54), for combining the first speech signal and the second speech signal, and outputting a third speech signal representative of the speech transmitted by the headphone wearer, characterized in that : the physiological sensor (18) is incorporated in a circumaural pad (16) of a shell (14) of one of the earphones (12); the microphone assembly comprises two microphones (20, 22) placed on the shell (14) of one of the earphones (12); - the two microphones (20, 22) are aligned in a linear array in a main direction (24) directed towards the mouth (26) of the helmet wearer; and - means (56) for non-frequency noise reduction of the second speech signal are provided, comprising a combiner able to apply a delay to the signal delivered by one of the microphones and to subtract this delayed signal from the signal delivered by the other microphone,
in order to operate a denoising of the close speech signal emitted by the wearer of the helmet. The headset of claim 1, further comprising: means (48) for low-pass filtering of the first speech signal before combination by the mixing means, and / or means (50, 52) for high-pass filtering of the second speech signal before denoising and combining by the mixing means, said low-pass and / or high-pass filtering means (48, 50, 52) comprising an adjustable cutoff frequency filter; and - means (44) for calculating the cutoff frequency operating according to the signal delivered by the physiological sensor. The headset of claim 2, wherein the means (44) for calculating the cutoff frequency comprise means for analyzing the spectral content of the signal delivered by the physiological sensor, able to determine the cutoff frequency according to the levels. the signal-to-noise ratio evaluated in a plurality of separate frequency bands of the signal delivered by the physiological sensor. The headset of claim 1, further comprising: means (62) for denoising the third speech signal delivered by the mixing means, operating by frequency noise reduction. The headset of claim 4, further comprising input receiving means, performing cross correlation between said first and said third speech signal, and outputting a speech presence probability signal as a result of said cross correlation result. . The headset of claim 5, wherein the denoising means (62) of the third speech signal receives said speech presence probability signal and is adapted to selectively: i) performing a differentiated noise reduction according to the frequency bands as a function of the value of said speech presence probability signal, and ii) perform maximum noise reduction on all frequency bands in the absence of speech. The headset of claim 1, further comprising: - Post-processing means (64) capable of selectively frequency-band equalizing the part of the spectrum corresponding to the signal collected by the physiological sensor. The headset of claim 7, wherein the post-processing means is adapted to determine an equalization gain for each of said frequency bands, said gain being calculated from the respective frequency coefficients of the signals delivered by the microphones and signals delivered by the physiological sensor, considered in the frequency domain. The headset of claim 8, wherein the post-processing means is further operable to smooth out a plurality of successive signal frames of said calculated equalization gain.

Images