METHOD FOR CHARACTERIZING A FILTER FOR PROCESSING AN INDIVIDUAL'S VOICE, COMMUNICATION DEVICE
Patent Information
- Application Number
- FR2022012450
- Authority / Receiving Office
- FR · FR
- Patent Type
- Patents
- Current Assignee / Owner
- Filing Date
- 2022-11-28
- Publication Date
- 2024-11-29
- Estimated Expiration
- 2042-11-28
AI Technical Summary
Existing audio processing systems for individuals using underwater breathing aids face challenges in intelligibility due to mouthpiece distortion and require significant calculation times, making real-time communication difficult.
A method for characterizing a filter using a computer-implemented process that acquires and processes audio signals with and without an underwater breathing aid, employing differential equations and neural networks to minimize errors, allowing for real-time intelligibility reconstruction.
The method enables real-time processing and reconstruction of intelligible audio streams with reduced computational requirements, adapting to individual voices and environments, and requires minimal code and memory, facilitating quick filter training and deployment on embedded systems.
Abstract
Description
Description Title of the invention: METHOD FOR CHARACTERIZING A FILTER FOR PROCESSING A VOICE OF A INDIVIDUAL, COMMUNICATION DEVICE Field of invention
[0001] = The invention relates to a method for processing and correcting an altered audio stream coming from an individual having, for example, a mouthpiece in his mouth allowing him to breathe underwater and comprising a microphone. The field of the invention relates to the processes implemented by a computer embedded in electrical equipment- electronics to be executed in real time, without external connection and configured for make an audio stream distorted by the presence of a maintained element more intelligible in the mouth. State of the art
[0002] = We know the document US2012 / 0213034 describing a regulator mouthpiece to be placed in the mouth and including a microphone and an earphone bone conduction vibrating an acoustic conduction plate to make pass the signal through the user's teeth.
[0003] — A defect is that the captured signal is not very intelligible due to the labial part of the mouthpiece preventing the user from moving their lips sufficiently. It there is a need to reconstruct an intelligible sound from a processing method of the voice.
[0004] — A first problem is that the audio stream depends a lot on the individual who pronounces sentences that are distorted by the presence of the mouthpiece. A second The problem is that some algorithms require significant calculation times which do not allow for real-time processing of an individual's voice, on a embedded system, thus making communication between two individuals difficult using this type of equipment. At most it is possible to recover the voice of a swimmer and process it later.
[0005] — The invention provides a computer-implemented method, in particular an or- on-board computer, such as an electronic control unit, called MCU, a mi- microprocessor, a microcontroller or an FPGA. The method of the invention allows the reconstruction and / or correction of certain syllables by real-time processing of the audio stream acquired by a microphone located near a diver's mouth. The method of the invention makes it possible in particular to transmit or process on reception the audio stream so that another diver can listen to the stream intelligibly audio emitted by a first diver. Summary of the invention According to a first aspect, the invention relates to a method for characterizing a filter for processing a voice of a given subject, said voice being acquired by a microphone arranged in an underwater breathing aid device intended to be worn near the mouth of a subject, said method comprising: Acquisition of at least one first signal corresponding to the pronunciation of a first set of phonemes by said given subject; said first signal corresponding to a subject's voice acquired by the device's microphone underwater breathing aid when the latter is worn by said subject; Acquisition of at least a second signal corresponding to the pronunciation of the first set of phonemes by said subject; said second signal cor- responding to a subject's voice acquired by a microphone without the subject wearing an underwater breathing aid device; For the first signal, perform the following steps: Sampling the first signal; Estimation of a first set of average values of the amplitude of the first signal over a plurality of first windows predefined, each first window comprising a plurality of samples of the first signal; Estimation of a second set of signal speed values at from a finite-time homogeneous differentiator method on the plurality of first predefined windows; Calculation of a first set of coefficients of a system of differential equations representing a first model whose solutions are the first and second set of values; Calculation of an error between at least one set of samples of the second signal and the samples obtained at the output of the first model executed with values of the first signal and the estimated coefficients, Iteration(s) of the calculation of the coefficients of the first model by minimizing to each iteration said error calculated, the number of iterations being configured to minimize the error below a predefined threshold so as to define operating coefficients, said operating coefficients defining the filter coefficients. An advantage is that it allows for characterization of an audio filter in order to then reconstruct an intelligible sound when a voice is distorted by a mouthpiece. An advantage of the method of the invention is that it provides an algorithm that requires little computing power and memory to be implemented on an embedded computer. Another advantage is that it allows for simple encoding with minimal lines of code. Another advantage is that it allows for an algorithm that adapts to the user and their environment. According to one embodiment, the characterization method comprises a step of filtering the first signal carried out prior to sampling. One advantage is that it allows for the characterization of an audio filter that is free from noise, i.e., a characterization of the induced distortion of the voice due to the equipment present in the mouth. According to one embodiment, the sampling of the first signal comprises the following steps: Analysis of the signal spectral density in order to determine characteristic thresholds amplitude or power characteristics; Selection of a set of characteristic frequencies for which the thresholds are exceeded; Sampling segments of the first signal around each frequency characteristic ; Generating a set of samples for processing the first signal. One advantage is to focus the calculation times on the areas of interest of the voice, that is to say the characteristic frequencies of an individual's voice, in particular those which have a maximum useful signal. According to one embodiment, the sampling produces between 400 and 96,000 samples during 1 second. One advantage is that it allows for different operating modes depending on whether one wishes to prioritize calculation time or signal quality. Sampling at 4000 samples per second allows in particular to keep a voice spectrum between 300 Hz and 1300 Hz and reduce computing power. According to one embodiment, the method is repeated for a plurality of acquisitions of first signals, each first acquired signal corresponding to the pronunciation of a new set of phonemes by the same subject, the calculation of the error and its minimization being carried out for each new first acquired signal. An advantage is that the method of the invention can be carried out for a few phonemes, i.e. a few sentences, to allow a filter to be characterized quickly. Another advantage is that it allows rapid training of the algorithm from 3 or 4 sentences spoken by an individual. Thus, the equipment can be quickly ready, configured and trained just before use, for example just before carrying out a dive. According to one embodiment, the system of differential equations is of the first order. One advantage is the simplification of calculations and therefore the saving in calculation time. According to one embodiment, the characterization method comprises: Estimation of a third set of signal acceleration values at starting from the method of the homogeneous differentiator in finite time on the plurality of first predefined windows; Calculation of a first set of coefficients of a system of differential equations first-order annuities representing a first model whose solutions are the first and second set of values. One advantage is the improved characterization of the filter allowing better reconstruction of an individual's distorted voice. According to one embodiment, the characterization method comprises: Estimation of a N°th set of values of an Nth order derivative of the signal at starting from the method of the homogeneous differentiator in finite time on the plurality of first predefined windows; Calculation of a first set of coefficients of a system of differential equations N'*"th order annuities representing a first model whose solutions are the first and second set of values. One advantage is the improved characterization of the filter allowing better reconstruction of an individual's distorted voice. According to one embodiment, the first model is a linear model and is written: dx / dt{t) = Gx{(t}) + FS / {t), y(t)= Hx(t) where G, F and H are matrices whose coefficients are to be determined, the dimension of the matrix being determined by the order of the linear system of differential equations, y(t) is the signal that we seek to identify approaching the second signal, and x(t) is a state variable internal to the first model. According to one embodiment, the first model is a non-linear model and is written: dx / dt{t)} = Gx{t) + FS{(t) v{t) = Hx(t) + NN(x(t), S,)) where G, F and H are matrices whose coefficients are to be determined, the dimension of the matrix being determined by the order of the system of differential equations, NN designates a non-linear filter, y(t) is the signal that we seek to identify approaching the second signal, and x(t) is a state variable internal to the first model, the method further comprising a calculation of a set of coefficients of a neural network from a regression configured to minimize the error. According to one embodiment, the calculation of the error minimization step is carried out from the implementation of a neural network whose coefficients are calculated by a regression to minimize the error. One advantage is the simplicity and speed of learning. According to one embodiment, at each iteration a first condition is verified, said first condition being the verification that the error is less than a predefined threshold and / or that a second condition is verified, said second condition being the verification that a predefined number of iterations has been carried out, the method comprising a step of generating the coefficients of the filter as soon as one of the two conditions is verified. According to another aspect, the invention relates to a computer program product comprising a memory and a calculator for executing code instructions implementing the method of the invention. One advantage is that it allows filter characterization on any type of equipment, for example a smartphone, or any other mobile electronic equipment from which the characterization filter can be developed and then transferred to diving equipment. According to another aspect, the invention relates to a method of processing a voice of a given subject comprising: Acquisition of a signal corresponding to a voice acquired by an arranged microphone in an underwater breathing aid device intended to be worn in the mouth of a subject, the acquired voice being defined by an audio signal acquired in real time corresponding to a subject's voice distorted by the presence of the mouthpiece in the mouth; Application of a first model trained according to the characterization method of the filter of the invention, said application aiming to obtain a second signal of voice transformed from the first signal of the voice distorted by a underwater breathing aid device. One advantage is to equip equipment in advance with a pre-configured filter and adapt the training to the voice of a given individual. According to another aspect, the invention relates to a communication device comprising a mouthpiece for wearing in the mouth of a subject comprising: a microphone to acquire an acoustic signal from the subject's voice; an electronic card comprising an audio filter, an analog-to-digital converter, a calculator for sampling the digitized signal, a memory for recording the coefficients of a filter to be applied to the sampled signal and a calculator for generating an output signal, said recorded coefficients being calculated from a method of the invention. Description of the invention The invention relates to a first computer-implemented method for configuring a filter for processing and correcting an audio stream. This method is called a method for characterizing a filter for processing a voice. The invention also relates to a second computer-implemented method intended to be embedded and executed in electronic equipment for communicating, in particular underwater. This second method implements a filter characterized by the first method. These two processes can be performed one after the other, successively or spaced out in time. However, it should be noted that the first process can be performed before the second process or an alternative to the second process. Nevertheless, it achieves the same ultimate objective of processing the audio stream of a diver with a mouthpiece in his mouth while speaking to make it intelligible. The invention also relates to the communication device forming an element of an underwater walkie-talkie system. It is a voice communication device between divers. This device generally comprises a mouthpiece containing the electronics necessary to record and reproduce the sound. The mouthpiece comprises an electronic interface such as a cable allowing it to be connected to a housing. An electronic card arranged in the housing comprises components for recording the coefficients of the filter defining its characterization and for carrying out the data processing and transmission and reception of the audio signal. An advantage of the invention is to create a voice communication device for scuba diving with little or no modification to the diver's equipment. This constraint involves using a regulator mouthpiece as a means of receiving and transmitting the audible sound signal. Since the voice is modified by the mouthpiece and its intelligibility is thus reduced, the invention makes it possible to recover part of the degraded intelligibility thanks to the filter which was characterized by the first method. Microphone According to one embodiment, the microphone can be attached to a mechanical part conducting the air flow created by the diver's voice, this mechanical part comprises at least one channel or a lumen extending between the mouthpiece and the regulator. Thus, according to different embodiments, the microphone can be attached between the mouthpiece and the regulator so as to capture the audio flow spoken by a diver. The invention relates to all these embodiments. In another example, the microphone can be integrated into an elastic covering can be protected by a protective wall arranged in front of the microphone sensor. The protective wall advantageously reduces the difficulties of integrating the microphone into the covering. The mouthpiece 1 preferably comprises a labial portion 42. The labial portion 42 is intended to be covered by the user's lips during use. Said labial portion 42 comprises a lumen or lumens allowing the passage of air on either side of the labial portion. Advantageously, this lumen or these lumens allow(s) the passage of air from the channel of the regulator to the subject's mouth during use. The labial portion 42 makes it possible to create with the user's lips a sealed closure between the external environment and the air passage lumen(s). The labial portion 42 extends between a breathing zone comprising the acoustic conduction plates 20 and a connection means 43. The connection means 43 advantageously allows the mouthpiece to cooperate with a diving regulator. According to one embodiment, the connection means 43 is tubular in shape, the walls of which are rigid or flexible. Characterization process / treatment process Figures 1 and 2 relate more particularly to the steps of a method for processing an audio stream PROC, implementing a filter characterized by a method for characterizing the filter according to the invention. The audio stream processed by the PROC processing method is denoted S,(t). This audio stream is acquired and filtered using the F filter, characterized by a first characterization process of the CARAC filter. The application of the F filter makes it possible to reconstruct part of the intelligibility lost by the presence of the EM tip in the mouth of an individual. The invention therefore relates to a method noted PROC, in [Fig.2] which consists of applying APP, the filter F, to an acquired input ACQ, of an audio stream corresponding to the voice of an individual, the filter F, being characterized by a first method noted CARAC 1 which is described below using figures 3 and 4. Figures 3 and 4 therefore relate more specifically to the steps of a CARAC characterization process, of a filter from a few sequences spoken by a user with a mouthpiece in their mouth and the same sequences spoken without a mouthpiece in their mouth. The filter is then used for the diver's equipment. The audio streams processed by the CARAC characterization process of a filter F are noted S,(t) and S»(t). The CARAC process involves learning a function characterizing the filter in order to be adapted to correct the deformation induced by the presence of a mouthpiece in the mouth and to take into account a type of voice, its nature, its tessitura, its timbre, etc. [Fig. 1] represents in particular the passage of the audio and sound stream in coding within different components. For this purpose, [Fig. 1] represents the mouthpiece EM,, also noted 1 in [Fig. 5], of an underwater communication device. It may comprise a portion to be held in the mouth, for example formed by two acoustic conduction plates 20. The mouthpiece 1 preferably comprises a labial portion 42 and a connection 43 with, for example, a regulator (not shown). Openings 61 and 62 are preferably provided for the passage of air between the mouth and the regulator. A microphone MIC (not shown in [Fig. 5]) makes it possible to acquire the sounds produced by an individual speaking in the mouthpiece 1 before transmitting the signal to an analog-digital converter CAN,. The invention relates to any other mechanical architecture of a mouthpiece 1 intended to be connected to a regulator.In particular, the device of the invention may include different variants, in particular on the number of channels envisaged, on the presence or absence of acoustic conduction plates, their arrangement and their material. In the example of [Fig.5], a vibrating element 3 can be used to transcribe the audio emitted by another individual and thus allow listening by bone conduction. Another system could however be used to transcribe the voice and its intelligibility which has been processed by the method of the invention. Phonemes for learning [Fig. 3] describes two branches of a method. The first branch comprises an ACQ step of acquiring an incoming audio stream S,(t) comprising a first set of phonemes deformed by the mouthpiece 1. The second branch comprises the acquisition of an incoming audio stream S:(t) comprising the same set of phonemes defining a second audio stream not deformed by the presence of a mouthpiece 1. Preferably, the phonemes are the same and are preferably acquired by the same user. Indeed, in order to allow learning and definition of the coefficients of the optimal filter, it is necessary to isolate as much as possible the influence of the deformation linked to the presence of the mouthpiece 1 in the mouth. Thus, the invention is particularly relevant in the case where the two sequences of phonemes are substantially the same. Learning will also be all the more effective when the same individual pronounces the two audio streams S;(t) and S-(t) The phonemes may, for example, correspond to a few sentences to be pronounced with and without a mouthpiece 1. The invention thus proposes an embodiment in which an individual pronounces two same groups of sentences with and without a mouthpiece. Phonemes can correspond to a grouping of letters, syllables or phrases. Both groups preferably include the same group of phonemes and the same ordering of these phonemes within the sequence pronounced by an individual, with and without a hearing aid. It is understood that learning an algorithm based on a few spoken sentences, for example between 1 and 10 sentences, allows rapid learning for the user. The recording of the audio stream made with the mouthpiece is processed in particular by a differentiator. The recording of the audio stream made without the mouthpiece is used as the desired output when creating the filter. A differentiator of the invention may comprise a set of components for performing the mathematical operations of differentiation, i.e., providing an output proportional to the derivative of the input with respect to one or more variables. According to a first variant, the invention makes it possible to consider in the algorithm all the spoken sentences comprising a plurality of letters, syllables or sentences. In this case, they are all processed together. According to a second variant, the invention makes it possible to consider each sentence in the algorithm sequentially and independently of each other. In this case, each spoken sentence comprises a plurality of letters, syllables and each sentence is processed independently of the others. The sentences can be defined in such a way as to obtain a distribution of syllables most representative of the language that can be used in an underwater environment and in a given language. According to an exemplary embodiment, the sentences selected for the training and characterization of the filter F, comprise a representative sample of the different labials, in particular syllables comprising the following letters: B, P, M, F, V. According to one embodiment, the sentences can be displayed on a display of the diving equipment so that the latter reads them. According to another example, the sentences are displayed on a display of an electronic terminal such as that of a Smartphone by means of an application allowing a language to be selected. Thus, it is understood that the CARAC characterization method can be carried out according to the embodiments directly on the diving equipment or on a Smartphone. When the filter is thus characterized from a Smartphone, different possibilities can be implemented to transfer the characterized filter to a memory of an electronic card arranged on the diving equipment, such as the mouthpiece. According to one embodiment, the display of the sentences includes a cursor for indicating a reading speed. The display may include graphical indications for indicating to a user a reading sequence with or without a mouthpiece in the mouth in order to acquire the two sets of sentences. F1 filter computation resources In order to carry out this preliminary step, according to a first embodiment, a user can use an electronic terminal such as a Smartphone, i.e. a smartphone or any other device with a microphone. The two sequences of spoken sentences can be recorded using the microphone of the electronic terminal and transmitted to a remote server to carry out the method of the invention. According to this mode, the method is carried out remotely and can take advantage of significant computing resources. In particular, according to a preferred embodiment, the characteristics of the filter can be calculated by a remote server having significant computing resources without hardware constraints linked to a particular mobile device. For this purpose, the voice signal emitted by a user can be acquired by a mobile electronic device, such as a Smartphone, otherwise called a smart phone, and be transmitted to a remote server. The filter calculated by the remote server can subsequently be retransmitted to the Smartphone or the onboard electronic diving equipment. According to a second embodiment, a user can use diving equipment comprising an electronic module comprising a computer and a memory. In this case, the learning is carried out using fewer resources than a server. The algorithm is then optimized to be executed in a constrained environment. One advantage is not needing a network connection, this case can be interesting in a diving environment at sea, for example when located on a boat far from the coast. Another advantage is using the microphone of the diving equipment. Thus, the learning will be all the more effective as it will overcome the gaps linked to the differences in materials between the equipment used for learning and that used for operating the characterized filter. According to a third hybrid embodiment, the filter can be calculated on an electronic terminal of the Smartphone type from a downloaded application allowing the use of the computing resources of the electronic terminal. One advantage is to produce the filter with a little more resources than the case of an electronic card embedded in the diving equipment while allowing the operation to be carried out locally, that is to say without implementing a remote server. When the operation is carried out by an electronic terminal of the Smartphone type, a wireless connection may be provided allowing the loading of the filter F, calculated by a resource other than that of the diving equipment so that it is recorded in a memory or a resource of said diving equipment. Audio signal acquisition mode The signal from the microphone with the mouthpiece arranged in the mouth of an individual is called the first signal S,, the signal acquired by a microphone without the mouthpiece being arranged in the mouth of an individual is called the second signal S,. These signals S, and S; are acquired for the purpose of characterizing an audio filter. Preferably, a single microphone is used to acquire both signals S, and S,. This embodiment makes it possible to overcome the intrinsic characteristics of the microphone and the influence of the latter on the acquired signal. According to one embodiment, this microphone is the microphone of the diving equipment, for example the one arranged in the part of the distal mouthpiece located between the part intended to be positioned in the mouth and the regulator. According to another embodiment, the microphone is the microphone of an electronic terminal such as that of a mobile phone of the Smartphone type. In the latter case, an individual positions the mouthpiece 1 in the mouth and says the few sentences near the microphone of the electronic terminal to train the algorithm executed by the latter. According to one embodiment, a known corrective filter can be used to correct the signal acquired by the Smartphone. Other embodiments may be implemented, including modes involving the use of signals acquired by different microphones. According to one embodiment, the two signals S, and S, are digitally encoded after their conversion in an analog-to-digital converter, called CAN. The conversion and encoding advantageously comprise sampling of each signal. In the same way, during the acquisition of the signal in the processing method using a characterized filter, an analog-digital conversion step is implemented and sampling is carried out of the acquired audio signal. According to one embodiment, the first signal S, corresponding to the sentences spoken with the mouthpiece in the mouth, are sampled according to a first sampling ECH. The samplings of the signals implemented in the characterization method CARAC, of the filter F, and in the processing method PROC, may be identical, that is to say that the sampling frequency, the measurements carried out on the samples by grouping the latter by considering time windows are configured identically. However, the invention could implement sampling specific to the characterization of the filter F, and sampling specific to the method of processing the audio stream spoken by a diver. Typically, in this embodiment, the sampling frequencies could be different.One interest is to consume less information during the processing stages during a dive in which we wish to maintain the communication link as long as possible, whereas in the learning and characterization phase of the filter F,, finer sampling, i.e. allowing more samples to be obtained, makes it possible to obtain a more efficient filter F. According to an exemplary embodiment, sampling resulting in the recording of digital values is configured for a high value of 96,000 samples per second. However, a lower number of samples can also achieve very good results with a value between 4,000 and 8,000 samples. per second. During audio signal acquisition, a number of sliding samples is stored in memory during cycles of real-time calculations of the average values. A number of samples between 200 and 400 samples is stored in each cycle of real-time calculations. Additional bandpass or lowpass filter According to an exemplary embodiment, a preliminary filtering can be carried out before sampling. The latter filtering aims to eliminate noise or to restrict the analysis band of the signal. According to one example, the invention comprises the implementation of a low-pass filter whose cutoff frequency depends on the sampling frequency. According to one example embodiment, a suitable filter makes it possible to filter the effect produced by bubbles during breathing or when a sound is emitted from the mouth of a diver. The specific frequency can be filtered so as to improve intelligibility upstream of the construction of the characteristic filter F,. Sample selection According to an exemplary embodiment, in order to reduce the calculation time, a selection of samples is chosen. To this end, the selection of a set of samples amounts to choosing a set of time windows Tf, each comprising a subset of samples from which the average values will be calculated. In order to select the best samples or the best groups of samples to define a high-performance characteristic filter F, a step of calculating characteristic frequencies of the spectrum of the acquired signal can be carried out. One advantage is to select a few frequencies comprising either the greatest signal intensities or the frequencies having the high spectral density, for example by considering the spectral power of the acquired signal. These selected frequencies make it possible to select an interval around the frequency to retain characteristic samples of the audio stream. This method saves time and calculations and therefore makes it possible to embed the method of the invention in an electronic component on which learning can be carried out to define the characteristic filter F, quickly. According to an example between 5 and 20 characteristic frequencies can be configured to determine between 5 and 20 intervals around each frequency allowing to define between 5 and 20 sample time windows. Speed and acceleration of spoken sounds [Fig.6] represents an output S,(t) corresponding to the audio stream acquired with a tip and an output S,(t) corresponding to the audio stream acquired without a tip. We recall that the sentences consisting of phonemes in each signal S;(t) and S2(t) are identical in this example. We note on the graph of [Fig.6] that the duration of the audio stream Sy(t) is longer than the duration of the audio stream S(t) because the speed of speech is slowed down by the presence of the mouthpiece in the mouth. The method of the invention makes it possible to take into account in the model used the deformation of the sound induced by the speed of speech and therefore of the sound of the audio stream which is recorded and filtered. An advantage of the invention is to take into consideration in the characterization of the filter the signal artifacts linked to the signal deformation induced by the change in speed and acceleration of the sound pronounced by a diver. Estimation of average sizes The method of the invention is based on the implementation of a method based on the theory of homogeneity by the definition of a homogeneous differentiator. This method makes it possible to carry out a finite-time estimation of noisy signals. An advantage of the method of the invention is that the algorithms implemented are simple, particularly fast and non-asymptotic in nature, unlike traditional methods. This method allows the estimation of signal derivatives according to the chosen order by considering all the sampled data. In order to smooth the estimations and reduce artifacts related to acquisition errors or noisy signals, average values are calculated to make the method of the invention more robust. The method makes it possible to estimate EST, a first set of values of data relating to the signal acquired over predefined time windows Tf,, each first window Tf, comprising a given number of samples of the signal. Thus, according to one example, the physical quantity relating to the acquired signal is the amplitude of the signal. According to one embodiment, the method makes it possible to estimate the average values of the amplitude of the signal taken over a set of samples from a given time window. Other physical quantities can be used or processed according to the method of the invention, however the embodiment if detailed below is relative to other processing of the average values of the amplitudes of the acquired signal. The method takes into consideration a value related to the dynamics of the signal: its speed. To this end, the method of the invention makes it possible to estimate EST; a second set of physical quantities relating to the dynamics of the signal. According to one example, the speed values of the first signal S, are estimated. The speed corresponds to the speed with which the sentences are pronounced. In order to calculate the speed of the first signal S,, a method called the homogeneous differentiator in finite time DHT is implemented. The differentiator is applied to a plurality of first predefined time windows each containing a set of samples. According to an improved example, the method takes into consideration another value linked to the dynamics of the signal: its acceleration. For this purpose, the method of the invention makes it possible to estimate EST; a third set of physical quantities relating to the dynamics of the signal. According to one example, the acceleration values of the first signal S, are estimated. The acceleration corresponds to the variations in the speed with which the sentences are pronounced. In order to calculate the acceleration of the first signal S,, an equivalent method called the homogeneous finite-time differentiator DHT is also implemented. The differentiator is applied to a plurality of first predefined time windows each comprising a set of samples. The samples are preferably the same as those considered to estimate the speeds. Similarly, during the PROC processing process, the values of the derivatives and accelerations of the input signal Sa({) which comes from a signal corresponding to an audio stream of a diver acquired with a mouthpiece in the mouth can be calculated. An interest is to generate the output stream Sp(t) from the filter F, characterized by the CARAC characterization process, obtained in the learning phase. Modeling of the dynamic system The method of the invention makes it possible to define a model based on the modeling of dynamic systems and to solve the problem posed by obtaining an approximate solution by optimizing the calculations by iteration while minimizing an error. The error can be defined by knowing the input of the system and the output of the system. A problem solved by the invention is to define a model which converges sufficiently quickly to define an operational and efficient filter F allowing the subsequent reconstruction of the audio stream distorted by the presence of the tip. The signal Sz(t) denotes the normal input voice that has not undergone any deformation, and can define the desired output of the model. The model of voice deformation due to the presence of the mouthpiece can be represented by this equality St) = W - St) - One advantage of modeling by a dynamic system is to introduce a differentiation operator with respect to the time variable t.- In such a modeling, we can write sit) = Wip) Syt) , where p is the differentiation operator with respect to the time variable, p = d / dt. The signal Sz(t) defines the output of the dynamic system in the model training phase. It is therefore appropriate to define the inverse problem of finding Sz(t) from S,(t). The inverse problem amounts to calculating the inverse function W-! which represents a learning model that we seek to optimize in order to calculate the signal Sp(t) represented on [Fig.2] corresponding to the corrected audio stream when the known signal S(t) will no longer be available. The first audio sequences S»(t) therefore make it possible to train the model in order to parameterize a filter F, trained audio, i.e. one whose coefficients have been produced by an algorithm comprising a regression or iteration operation aimed at minimizing an error. The coefficients thus produced are the operating coefficients used to define an operational F1 filter. The invention therefore makes it possible to generate a “learned model” in order to subsequently use a characteristic filter F, within the equipment to correct the audio flow distorted by the presence of the mouthpiece in the mouth. The inverse problem can be written: SÀt) = W“p) . S,(t)- Linear model Which can be written in terms of dynamical systems theory as the following equation according to a linear MOD model, where x(t) is an internal variable of the system: <1)=Gx(1) +F-S1r) S1(1)=H- dr) According to one embodiment, in order to solve this system, a first hypothesis is to consider a sufficiently smooth signal S.(t) in order to consider a derivable function having a set of derivatives constituting different orders of the function S; (D) = y(t) where y(t) is the output of the MOD model, or MOD, Thus, the set of derivatives is written [y], "n" being the order of the derivative of the sampled signal and being greater than or equal to 1. The desired output can be modeled by a vector z defined by the system of differential equations that we are trying to estimate: z=[y®, y®, y9), .. y(#P]T and y«(t) = Theta(t) This is the set of first, second, third, and so on derivatives up to the nth derivative. The number of derivatives considered defines the dimension of the system. Considering an order of the retained derivatives equal to 2 to simplify the calculations, it is possible to define a simple MOD model of the system which is efficient. The following is described for an order n in order to illustrate a more general mode of implementation. We can define the following equivalent system: 2= MA -1+ Phi(r) v=Cz In which MA is defined by the following matrix: 0 0 0 0 0 0 0 MA= ... 0 CO 0 80 0 0 0 0 0 And “C” is a vector defined by the following vector: C=[1,0,0...,0] And is a vector defined by the following vector: Phi(t) = [0, …, Theta(t)]T Joh ; À = $, We note paqetrhe = x” pod sign(x} where alpha is strictly greater than 0 and x is a real. The homogeneous finite-time differentiator is defined by the following system for i = 2, …ra di ; al 8-8. [8 ->] At At 18 =Â s 15.-v tk [8 r] To To an L=-k4 [E->] Where * is the estimate of z The coefficients k,, …, k, form a Hurwitz polynomial. It is possible in the resolution of this system to choose appropriately to ensure the homogeneity of the system the coefficients (a,, …a,) in the space of strictly positive real numbers of dimension n. According to an example, for a given A in the interval [1 — 1 / (n-1), 1], the sequences (r 1» +++» Fa) and (a;, …A,) can be chosen as follows: ri=1l -{i-1)-A,with1 <i<n a;= -lA ,withl <i<n In order to ensure finite-time convergence, according to one example, the method can be modeled by taking into account the conditions necessary for homogeneity theory. Under this assumption, it is possible to write that $é-1(t) = 2,(t), j = 1, …, n after a finite time of transients. The dynamics of differentiation errors takes the form of the following system: 4 dd ë,=ek; [e]”, i=2 .…., R-1 iA(ét) é=eu-k [e] pt ê,= rhëoft)- In pe 4 eh We know, having chosen the coefficients which can define a Hurwitz polynomial that there exists A in the interval [1-1 / (n-1), 1] sufficiently close to 1 such that equation (1) is globally stable in finite time. One of the main advantages of stable homogeneous systems in finite time is their rapid convergence rate and their robustness with respect to different perturbations. Due to the term Theta(t), it is impossible to obtain the convergence of the error to Zero without having additional knowledge about this signal Theta(t). This problem can be overcome by assuming that y{t) is locally polynomial and that over a small time interval, Theta(t) = 0. In this case, it is then possible to recover the derivatives. According to one embodiment, the method comprises a number of iterations aimed at repeating the sequence of calculating the coefficients until a calculated error is less than a given threshold. In order to quantify the performance of the results obtained with the final algorithm, an error is defined based on the norm of the difference between the reconstructed and original sounds. According to one embodiment, at each iteration a first condition Ç, is verified. According to one example the first condition C, corresponds to the verification that the error is less than a predefined threshold. According to one example, a second condition C» is verified. The second condition Cz is the verification that a predefined number of iterations has been carried out. Advantageously, the method includes a step of generating the filter coefficients as soon as one of the two conditions C, or C, is verified. The succession of iteration steps corresponds to the learning step of the algorithm which results in producing filter coefficients making it possible to reconstruct the signal Sp(t) from an input signal S,(t) according to the processing method PROC, of the invention. Nonlinear model According to an alternative embodiment, a nonlinear model MOD: can be used to implement the invention. xXt)=Gx{t) +F-IF(r) S1(1)=Hx(t) + NN(x(t), S1(#)) Where NN denotes a non-linear filter. In this implementation the filter F, = NN corresponds to an additional correction which allows to correct the modeling error of the linear model. In this case, the error is minimized by taking into consideration the nonlinear part NN. In the case of an implementation of a nonlinear model MOD, the filter F, is a nonlinear filter. According to an example, a method for calculating the coefficients of the filter nonlinear F, can be implemented from a neural network. For this purpose, a forward propagation neural network is produced with the result obtained with the linear filter. According to one example, the invention comprises the implementation of a CNN-type network, defining a convolutional neural network. In this exemplary implementation, a conventional feedforward network with a hidden layer that uses sigmoidal activation functions achieves good learning performance. In the case of a non-linear filter, learning also includes the phase of calculating the network coefficients by a regression carried out at each iteration. According to other examples, other neural networks can be used. The method of the invention is based on the implementation of an algorithm allowing the calculation of coefficients integrating iterative phases including the calculation of an error. This algorithm is called a "learning algorithm" or "learning function" insofar as it takes into account intermediate steps allowing convergence towards a result leading to the construction of the characteristic filter F. Once the filter F is determined, the set of parameters of the MOD model, for the linear model and the MOD model; for the nonlinear model are determined. The models can then be applied to predict the output Sp(t) in [Fig.2]. Figures 3 and 4 represent the step of calculating the error er(t) at each iteration considering S-(t) acquired directly and the output y(t) estimated. According to one embodiment, an initial filter Fi is pre-trained and recorded in a memory of a device. Such an initial filter Fi can be trained with different voice profiles such as female voice profiles, male voice profiles, or child voice profiles. Certain vocal ranges or timbres can be used to pre-train an initial filter Fi within the device. One advantage is to define initial coefficients making it possible to improve the convergence of the error below a given threshold during training. In order to define an initial filter Fi, the CARAC characterization method of the invention can be used with a first configuration. The first configuration can benefit from additional computing resources, so the order of the system can be increased, as well as the samples or even a greater number of training sentences can be available. In this configuration, a nonlinear model MOD; can be used to define the initial filter Fi and one or the linear model can be used to train the model with the real voice of the diver. According to another embodiment, the PROCI processing method of the invention can be implemented to correct an error also during the diving phase, for example with supervised or unsupervised training. For example, a repeated sentence or a signal indicating misunderstanding can be used to label model outputs to improve the construction of an F2 filter during the PROC processing process. According to an exemplary embodiment, the method of the invention may comprise three successive trainings of the model or models which may be carried out at times spaced apart from each other. A first training of the model makes it possible to define an initial filter Fi from a first configuration of the CARAC characterization method, for example with a pre-selected voice corresponding to a first signal S2(t). A second training of the model allows to define a characteristic filter F, from a second configuration of the CARAC characterization process, with the diver's voice corresponding to a second signal S2(t). This second training can be carried out with initial model coefficients which were calculated from the first training. Finally, a third training can be carried out during communications using the PROCI processing method and automatic labeling of outputs which can be deduced from the analysis of the audio stream of one or both divers, such as for example: Repetitions of phrases spoken at the same speeds or at different speeds different; Identification of predefined sentences such as "not understood", "I don't “don’t understand”, “can you repeat”, etc.; A change in breathing or detection of a bubble rate; Etc. Conversion of the filtered and transmitted signal for bone conduction According to one embodiment, the output filtered by the filter F, characterized by the CARAC characterization method, of the filter is transmitted via a wireless communication interface to another communication device of another swimmer. According to one example, the filtered audio output is modulated on a carrier frequency. The invention is compatible with any form of wireless transmission by electromagnetic means such as a radio link or by emission of sound or ultrasonic waves. Different modulations can be used, whether frequency modulation, amplitude modulation. According to one embodiment, the filtered audio signal is transmitted by being modulated by a frequency evasion technique. The mouthpiece 1 may comprise signal transmission means connected to the microphone 5 and / or to the vibrating element 3. The transmission means may comprise electronic cables and / or a conduction sheet. These transmission means may be at least partially integrated into an elastic covering. When the signal transmitted from one device to another is demodulated upon reception, the audio signal can be converted so as to vibrate a bone conduction plate. [Fig. 5] shows an example of an acoustic conduction plate 20 arranged so as to be bitten by the user wearing the mouthpiece 1. According to this example, the mouthpiece 1 further comprises a transducer-type vibrating element. The vibrating element is connected to the conduction plate so as to transmit the vibrations of the vibrating element to the acoustic conduction plate. The vibrating element 3 is configured to convert an audio input into an acoustic output and acoustically couple to the diver's upper and / or lower teeth to conduct the acoustic output from the diver's upper teeth through the skull to propagate the acoustic signal to the diver's inner ear via the skull and jaw bones when the diver wears the mouthpiece in the mouth. The acoustic conduction plate 20 is designed and arranged to acoustically couple to the user's upper teeth to conduct the acoustic output of the diver's upper teeth through the skull to the cochlea to generate audible sound in at least one of the diver's inner ears when the diver wears the mouthpiece 1. The acoustic conduction plate 20 is configured to engage and acoustically couple to the surface of the diver's teeth and is configured to conduct vibrations of the vibrating element 3 in response to an electrical signal. The vibration of the vibrating element 3 produces an acoustic output signal that is acoustically conducted to the diver's teeth, via the acoustic conduction plate, and then through the bones of the diver's jaw and skull to the inner ear, including the cochlea, where it is perceived as sound.
Claims
Claims
1. Method for characterizing (CARAC) a filter (F,) for processing (PROC) a voice of a given subject, said voice being acquired by a microphone arranged in an underwater breathing aid device intended to be worn near the mouth of a subject, said method comprising: Acquisition (ACQI) of at least one first signal (S,) cor- corresponding to the pronunciation of a first set of phonemes by said given subject; said first signal (S;) cor- responding to a subject's voice acquired by the microphone of the underwater breathing aid device when the latter is carried by said subject; Acquisition (ACQ2) of at least one second signal (S;) cor- corresponding to the pronunciation of the first set of phonemes (PHR,) by said subject; said second signal (S,) cor- responding to a subject's voice acquired by a microphone without that the subject is wearing a breathing aid device under navy; For the first signal (S,), perform the following steps: Sampling (ECH,) of the first signal (S,) for generate a plurality of samples from the first signal (S,); Estimation (EST) of a first set (ENS) of average values of the amplitude of the first signal (S 1) on a plurality of first predefined windows, each first window comprising a plurality of samples of the first signal (S1); Estimation (EST-) of a second set (ENS,) of signal speed values (S,) from a finite-time homogeneous differentiator method on the plurality of first predefined windows; Calculation (EST;) of a first set of coefficients of a system of differential equations representing a first model (MOD, MOD) whose solutions are the first and second set of values {ENS,, ENS»); Calculation of an error (Er) between at least one set of samples of the second signal (S;) and the samples obtained at the output of the first model (MOD,,MOD:;) executed with values of the first signal (S,) and the estimated coefficients (G, F, H), Iteration(s) of the calculation of the coefficients of the first model (MOD,,MOD:) by minimizing the said error at each iteration (Er) calculated, the number of iterations being configured to minimize the error below a predefined threshold in a manner to define operating coefficients, said coefficients operating system defining the filter coefficients (F,).
2. Characterization method according to claim |, characterized in that that it comprises a step of filtering the first signal (S,) carried out prior to sampling.
3. Characterization method according to any one of claims 1 to 2, characterized in that the sampling of the first signal (S,) includes the following steps: Analysis of the spectral density of the signal (S,) in order to de- terminate characteristic thresholds of amplitude or power ; Selection of a set of characteristic frequencies for which the said determined thresholds are exceeded; Sampling segments of the first signal (S;) around each characteristic frequency selected; Generation of a sample set (ECH) for the processing of the first signal (S,) from each segment sampled.
4. Characterization method according to any one of claims 1 to 3, characterized in that the sampling (ECH,) produces between 400 and 96000 samples for 1 second.
5. Characterization method according to any one of claims 1 to 4 characterized in that the method is repeated for a plurality of acquisitions of first signals (S,), each first signal (S,) acquired corresponding to the pronunciation of a new set of phonemes by the same subject, the calculation of the error (Er) and its minimization being carried out for each new first signal (S,) acquired.
6. Characterization method according to any one of claims 1 to 5 characterized in that the system of differential equations is of the first order.
7. Characterization method according to any one of claims 1 to 6 characterized in that it comprises: Estimation (EST) of a third set (ENS-) of values acceleration of the signal (S,) from the method of the differ- homogeneous finite-time reciprocator on the plurality of first predefined windows; Calculation (EST+) of a first set of coefficients of a system of first-order differential equations re- presenting a first model (MOD,,MOD);) whose solutions are the first and second set of values (ENS,, ENS, ENS).
8. A characterization method according to any one of claims 1 to 7 characterized in that it comprises: Estimation (ESTy) of an Nth set (ENSy) of values of an Nth order derivative of the signal (S;) from the method of the homogeneous differentiator in finite time on the plurality of first predefined windows; Calculation (EST) of a first set of coefficients of a system of differential equations of the Nwe order representing a first model (MOD,,MOD;) whose solutions are the first and second set of values (ENS,, ENS, ENSy )
9. A characterization method according to any one of claims 1 to 8 characterized in that the first model is a linear model {MODI1) and is written: dx / dt{t) = Gx(t) + FS(t), vb = Hx(t) where G, F and H are matrices whose coefficients are to be determined, the dimension of the matrix being determined by the order of the linear system of differential equations, v(t) is the signal that we seek to identify approaching the second signal (S-), and x(t) is an internal state variable to the first model (MOD)).
10. A characterization method according to any one of claims 1 to 8 characterized in that the first model is a non-linear model {MOD;) and is written: dx / dt{t) = Gx{t) + FS(t), y{t) = Hx(t)} + NN(x({t), S;(t)) where G, F and H are matrices whose coefficients are to be determined, the dimension of the matrix being determined by the order of the system of differential equations, NN denotes a non-linear filter, y(t) is the signal that we seek to identify approaching the second signal (S,), and x(t) is an internal state variable of the first model (MOD:,), the method further comprising a calculation (ESTy») of a set of coefficients of a neural network from of a regression configured to minimize the error (Er).
11. A characterization method according to any one of claims 1 to 10 characterized in that the calculation of the minimization step (EST,, ESTyx) of the error (Er) is realized from the implementation of a neural network whose coefficients are calculated by regression to minimize the error (Er).
12. A characterization method according to any one of claims 1 to 11 characterized in that at each iteration a first condition is verified, said first condition being the verification that the error is lower than a predefined threshold and / or that a second condition is verified, said second condition being the verification that a number of predefined iterations was carried out, the method comprising a step of generation of the filter coefficients as soon as one of the two conditions is met verified.
13. Processing method (PROC,) of a voice of a given subject including: Acquisition of a signal (Sa(t)) corresponding to a voice acquired by a microphone arranged in a res-aid device underwater breathing designed to be carried in the mouth of a subject, the acquired voice being defined by an acquired audio signal in real time (S,) corresponding to a subject's voice deformed by the presence of the mouthpiece in the mouth; Application (APP) of a first trained model (MOD) according to the method of characterization of the filter (F,) of one any of claims 1 to 12, said application (APP, ) aiming to obtain a second signal (Sg) of transformed voice at from the first signal (S,) of the voice distorted by a underwater breathing aid device (EB).
14. Communication device comprising a mouthpiece (1) intended to be carried in the mouth of a subject including: a microphone to acquire an acoustic signal coming from the voice of the subject; an electronic card containing an audio filter, a converter analog-digital, a calculator to sample the signal digitized, a memory for recording the coefficients of a filter (F,) to apply to the sampled signal (Sa(t)) and a calculator allowing generate an output signal (SB(t)), said recorded coefficients being calculated from a method according to any of the res- dication | at 12.
15. Computer program product comprising instructions which, when the program is executed by the communication device, lead the said communication device to implement the steps of the method of any one of claims 1 to 12.
