Classifications

• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
  • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
  • G10L21/0208—Noise filtering
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L19/00—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis
  • G10L19/02—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders
  • G10L19/0204—Speech or audio signals analysis-synthesis techniques for redundancy reduction, e.g. in vocoders; Coding or decoding of speech or audio signals, using source filter models or psychoacoustic analysis using spectral analysis, e.g. transform vocoders or subband vocoders using subband decomposition
• • G—PHYSICS
  • G10—MUSICAL INSTRUMENTS; ACOUSTICS
  • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
  • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
  • G10L21/003—Changing voice quality, e.g. pitch or formants
  • G10L21/007—Changing voice quality, e.g. pitch or formants characterised by the process used
  • G10L21/013—Adapting to target pitch
  • G10L2021/0135—Voice conversion or morphing

Definitions

• the present invention relates to a differentiated digital processing of voice and music, noise filtering, the creation of special effects as well as a device for implementing said method.
• Its purpose is more particularly to transform the voice in a realistic or original way and, more generally, to process in real time the voice, the music, the ambient noise and to record the results obtained on a support. computer science.
• the voice signal is made up of a mixture of very complex transient signals (noises) and quasi-periodic signal parts (harmonic sounds).
• the noises can be small explosions: P, B, T, D, K, GU; soft diffuse noises: F, V, J, Z or intense CH, S; as for harmonic sounds, their spectrum varies with the type of vowel and with the speaker.
• the intensity ratios between noises and vowels vary depending on whether it is a conversational voice, a conference-type spoken voice, a loud shouted voice or a sung voice. The loud voice and the sung voice favor the vowel sounds to the detriment of the noises.
• the voice signal simultaneously transmits two types of messages: a semantic message conveyed by speech, a verbal expression of thought, and an aesthetic message perceptible through the aesthetic qualities of the voice (timbre, intonation, flow, etc.).
• the semantic content of speech is practically independent of the qualities of the voice; it is conveyed by temporal acoustic forms; a whispered voice consists only of flow noises; an “intimate” or proximity voice is made up of a mixture of harmonic sounds in the low frequencies and flow noises in the high range; the voice of a speaker or singer has a rich and intense harmonic vocal spectrum.
• the musical range and the spectral content are not directly linked; some instruments have the maximum energy included in the range; others have a well-circumscribed maximum energy zone, located at the acute limit of the range and beyond; others still have very spread energy maxima which go far beyond the acute limit of the range.
• analog processing of these complex signals for example their amplification, induces, in an inescapable manner, an increasing degradation as and when said processing and this in an irreversible manner.
• the originality of digital technologies is to introduce a dete ⁇ i-iinism (that is to say a priori knowledge) as important as possible at the level of the signals processed so as to carry out particular processing which will reside in the form of calculations.
• this signal will be processed without undergoing degradation such as noise of background, distortion and bandwidth limitation; moreover, it can be processed to create special effects such as voice transformation, suppression of ambient noise, modification of the breath of the voice, differentiation of the voice and the music.
• bitrate reduction methods are mainly used for digital storage (with the aim of reducing the bit volume) and for transmission (with the aim of reducing the necessary bitrate). These methods include processing prior to storage or transmission (coding) and processing upon return (decoding).
• bitrate reduction methods those using the perceptual methods with loss of information are the most used and in particular the MPEG Audio method.
• This process is based on the mask effect of human hearing, i.e. the disappearance of weak sounds in the presence of loud sounds, equivalent to a displacement of the hearing threshold caused by the loudest sound and depending on the frequency and level difference between the two sounds.
• the number of bits per sample is defined according to the mask effect since the weak sounds and the quantization noise are inaudible.
• the audio spectrum is divided into a certain number of sub-bands, thus making it possible to specify the level of mask in each of the sub-bands and to perform a binary allocation for each of them. between them.
• the MPEG audio process thus consists in:
• This technique consists in transmitting a variable bit rate according to the instantaneous composition of the sound.
• this process is rather suitable for processing music and not for the voice signal; it does not detect the presence of voice or music, separate the vocal or musical signal and noise, modify the voice in real time to synthesize a different but realistic voice, synthesize breath (noise) to create special effects, encode a voice signal with one voice, reduce ambient noise.
• the invention therefore more particularly aims to eliminate these drawbacks.
• this process for transforming voice, music and ambient noise essentially involves:
• FIG. 1 is a simplified flow diagram of the method according to the invention.
• FIG. 2 is a flow diagram of the analysis step
• Figure 3 is a flow diagram of the synthesis step
• Figure 4 is a flow diagram of the coding step; and Figure 5 is a block diagram of a device according to the invention.
• the differentiated digital voice and music processing method according to the invention comprises the following steps:
• the analysis of the voice signal and the coding of the parameters constitute the two functionalities of the analyzer (block A); similarly, decoding parameters, special effects and synthesis constitute the functions of the synthesizer (block C).
• the differentiated digital processing method for voice and music essentially comprises four processing configurations: • the first configuration (path I) comprising the analysis, followed by the coding of the parameters, followed by the saving and reading the parameters, followed by the decoding of the parameters, followed by the special effects, followed by the synthesis,
• the analysis phase of the audio signal comprises the following steps: shaping of the input signal (block 1), calculation of the time envelope (block 2 ), time interpolation detection (block 3), detection of the audible signal (block 4), calculation of the time interpolation (block 5), calculation of the signal dynamics (block 6), detection of an inaudible frame after a frame higher energy
• TRF (block 11), preprocessing of the time signal (block 12), calculation of the TRF on processed signal (block 13), calculation of the signal to noise ratio (block 14), test of the doppler variation of the pitch (block 15 ) calculation of the TRF on unprocessed signal (block 16), calculation of the signal to noise ratio (block 17), comparison of the signal to noise ratios with and without preprocessing
• TRF fast Fourier transform
• the analysis of the voice signal is carried out essentially in four stages • calculation of the signal envelope (block 2), • calculation of the “pitch” and its variation (block 12), • application to the time signal of the inverse variation of the pitch (block 12),
• TRF fast Fourrier transform
• thresholds make it possible to detect respectively the presence of inaudible signal, the presence of inaudible frame, the presence of a pulse, the presence of disturbing sector signal (50 Hz or 60 Hz ).
• a fifth threshold makes it possible to carry out the rapid Fourrier transform (TRF) on the unprocessed signal as a function of the characteristics of the "pitch" and of its variation.
• a sixth threshold makes it possible to restore the result of the fast Fourrier transform (TRF) with preprocessing as a function of the signal to noise ratio.
• Two frames are used in the audio signal analysis method, a so-called “current” frame, of fixed periodicity, containing a certain number of samples corresponding to the voice signal, and a so-called “analysis” frame, the number of which of samples is equivalent to that of the current frame or the double, and being able to be shifted, according to the temporal interpolation, compared to the aforesaid current frame.
• the shaping of the input signal (block 1) consists in performing high pass filtering in order to improve the future coding of the frequency amplitudes by increasing their dynamics; said high pass filtering increases the dynamic frequency amplitude by avoiding that a low audible frequency occupies all the dynamics and makes disappear frequencies of low amplitude but nevertheless audible.
• the filtered signal is then directed to block 2 for the determination of the time envelope.
• the time difference to be applied to the analysis frame is then calculated by looking on the one hand for the maximum of the envelope in said frame and then on the other hand for two indices corresponding to the values of the envelope which are lower by a certain percentage than the maximum value.
• the temporal interpolation detection (block 3) makes it possible to correct the two offset indices of the analysis frame found in the previous calculation, and this taking into account the past.
• a first threshold (block 4) detects or not the presence of an audible signal by measuring the maximum value of the envelope; in the affirmative, the analysis of the frame is completed; otherwise, continuous processing.
• a calculation is then made (block 5) of the parameters associated with the time offset of the analysis frame by determining the interpolation parameter of the modules which is equal to the ratio of the maximum envelope in the current frame to that of the offset frame .
• the signal dynamics are then calculated (block 6) for its normalization in order to reduce the calculation noise; the signal normalization gain is calculated from the highest sample in absolute value in the analysis frame.
• a second threshold (block 7) detects or not the presence of an inaudible frame by mask effect caused by the previous frames; if so, the analysis is complete; otherwise, processing continues.
• a third threshold (block 8) then detects or not the presence of a pulse; if so, specific processing is carried out (blocks 9, 10); otherwise, the signal parameter calculations (block 11) used for the preprocessing of the time signal (block 12) will be performed.
• the repetition of the pulse (block 9) is carried out by creating an artificial pitch, equal to the duration of the pulse, so as to avoid masking of the useful frequencies during the fast transform of Fourrier (TRF).
• the fast Fourier transform (TRF) (block 10) is then carried out on the repeated pulse while retaining only the absolute value of the complex number and not the phase; the calculation of the frequencies and of the frequency data modules (block 20) is then carried out.
• the signal parameters (block 11) are calculated, which parameters relate to: - the calculation of the pitch and its variation,
• the calculation of the "pitch” is carried out beforehand by a differentiation of the signal from the analysis frame, followed by a low-pass filtering of the components of high rank, then by a cubic rise of the result of said filtering; the value of the pitch is determined by calculating the minimum distance between a portion of a high energy signal with the continuation of the subsequent signal, since the said minimum distance is the sum of the absolute value of the differences between the samples the template and the samples to be correlated; then, the main part of a "pitch” centered around one and a half times the value of the "pitch” is sought at the start of the analysis frame in order to calculate the distance of this portion of "pitch” over the entirety of the analysis frame; thus, the minimum distances define the positions of the "pitches", the "pitch” being the average of the "pitches” detected; then the variation of the "pitch” is calculated using a straight line which mutates the mean square error of the successions of the "pitchs” detected; the “pitch” estimated at the start and end
• the variation in pitch found and validated previously, will be subtracted from the time signal in block 12 of time preprocessing, using only the first order of said variation.
• Subtracting the variation in “pitch” consists in sampling the oversampled analysis frame with a sampling step varying with the inverse value of said variation in “pitch”.
• the oversampling, in a ratio two, of the analysis frame is carried out by multiplying the result of the fast Fourrier transform (TFR) of the analysis frame by the factor exp (-j * 2 * PI * k / (2 * L_frame), so as to add a delay of half a sample to the time signal used for the calculation of the fast Fourrier transform; the fast inverse Fourier transform is then carried out in order to obtain the time signal shifted by half a sample.
• TFR fast Fourrier transform
• a frame of double length is thus produced by alternately using a sample of the original frame with a sample of the frame offset by half a sample.
• the said "pitch” seems identical over the entire analysis window, which will give a result of the fast Fourrier transform (TRF) without spreading frequencies; the fast Fourier transform (TRF) can then be carried out in block 13 in order to know the frequency domain of the analysis frame; the method used makes it possible to quickly calculate the module of the complex number at the expense of the signal phase.
• TRF fast Fourrier transform
• the signal-to-noise ratio is calculated on the absolute value of the result of the fast Fourier transform (TRF); the above ratio is in fact the ratio of the difference of the signal energy and the noise to the sum of the signal energy and the noise; the numerator of the above ratio corresponds to the logarithm of the difference between two energy peaks, signal and noise respectively, the energy peak being that which is either greater than the four adjacent samples corresponding to the harmonic signal, or less than the four adjacent samples corresponding to noise; the denominator is the sum of the logarithms of all the signal and noise peaks; moreover, the calculation of the signal to noise ratio is done by sub-band, the highest sub-bands, in terms of level, are averaged and give the desired ratio.
• TRF fast Fourier transform
• the calculation of the signal-to-noise ratio defined as the signal-to-noise-to-signal-plus-noise-ratio, carried out in block 14, makes it possible to determine whether the signal analyzed is a voiced signal or music, in the case of a high ratio, or noise, in case of a low ratio.
• the calculation of the signal to noise ratio is then carried out in block 17, so as to transmit to block 20 the results of the fast Fourrier transform (TRF) without preprocessing, in the case of a variation of the zero pitch, or, in the opposite case to restore the results of the fast Fourier transform (TRF) with pretreatment (block 19).
• TRF fast Fourrier transform
• the calculation of the frequencies and of the frequency data modules of the fast Fourrier transform (TRF) is carried out in block 20.
• the fast Fourier transform (TRF), previously cited with reference to blocks 10, 13, 16, is performed, for example, on 256 samples in the case of an offset frame or a pulse, or on the double samples in the case of a centered frame without pulse.
• a weighting of the samples located at the ends of the samples known as HAMMING, is carried out in the case of the fast Fourrier transform (TRF) on n samples; on 2n samples, we use the HAMMING weighting window multiplied by the square root of the HAMMING window.
• the ratio between two adjacent maximum values is calculated, each representing the product of the amplitude of the frequency component by a cardinal sine; by successive approximations, this ratio between the maximum values is compared to values contained in tables containing this same ratio, for N frequencies (for example 32 or 64) distributed uniformly over a half sample of the fast Fourier transform (TRF) .
• N frequencies for example 32 or 64
• the index of said table which defines the ratio closest to that to be compared gives on the one hand the modulus and on the other hand the frequency for each maximum of the absolute value of the fast Fourier transform (TRF).
• the calculation of the frequencies and of the frequency data modules of the fast Fourier transform (TRF), carried out in block 20, also makes it possible to detect a DTMF (multifrequency dual tone) signal in telephony.
• TRF fast Fourier transform
• the signal to noise ratio is the essential criterion which defines the type of signal.
• the signal extracted from block 20 is categorized into four types in block 21, namely:
• the pitch and its variation can be non-zero; the noise applied to the synthesis will be of low energy; the coding of the parameters will be carried out with maximum precision.
• the pitch and its variation are zero; the noise applied to the synthesis will be of high energy; the coding of the parameters will be carried out with the minimum precision.
• - type 2 voiced signal or music.
• the pitch and its variation are zero; the noise applied to the synthesis will be of medium energy; the parameters will be coded with intermediate precision.
• this type of signal is decided at the end of the analysis when the signal to be synthesized is zero.
• a detection of the presence or non-presence of a disturbing signal at 50 Hz (60 Hz) is carried out in block 22; the level of the detection threshold is a function of the level of the signal sought so as to avoid confusing the electromagnetic disturbance (50, 60 Hz) and the fundamental of a musical instrument.
• the frequency plane is subdivided into several parts, each of which has several ranges of amplitude differentiated as a function of the type of signal detected at block 21.
• the temporal interpolation and the frequency interpolation are suppressed at the level of block 24; these were carried out to optimize the quality of the signal.
• Frequency interpolation depends on the variation of the pitch; this will be deleted depending on the offset of a certain number of samples and the direction of the variation of the pitch.
• the suppression of the inaudible signal is then carried out in block 25. Indeed, certain frequencies are inaudible because masked by other signals of higher amplitude.
• the amplitudes situated below the lower limit of the amplitude range are eliminated, then the frequencies whose interval is less than a frequency unit are moved away, defined as the sampling frequency by sample unit. Then, the inaudible components are eliminated using a test between the amplitude of the frequency component to be tested and the amplitude of the others. adjacent components multiplied by an attenuating term depending on the difference between their frequency.
• the number of frequency components is limited to a value beyond which the difference in the result obtained is not perceptible.
• the calculation of the pitch and the validation of the pitch are performed at block 26; in fact the “pitch” calculated in block 11 on the time signal was determined in the time domain in the presence of noise; calculating the pitch in the frequency domain will improve the precision of the pitch and detect a pitch that the calculation on the time signal, performed in block 11, would not have determined because of the noise ambient. Furthermore, the calculation of the “pitch” on the frequency signal must make it possible to decide whether it should be used for coding, knowing that the use of the “pitch” for coding makes it possible to greatly reduce the coding and to make the voice more natural to synthesis; it is also used by the noise filter.
• the principle of calculating the "pitch” consists in synthesizing the signal by a sum of cosines having originally zero phases; thus the shape of the original signal will be reconstituted without the disturbances of the envelope, the phases and the variation of the "pitch".
• the value of the frequency pitch is defined by the value of the time pitch which is equivalent to the first synthesis value having a maximum greater than the product of a coefficient by the sum of the modules used for local synthesis (sum of the cosines said modules); this coefficient is equal to the ratio of the signal energy, considered to be harmonic, to the sum of the noise energy and the signal energy; the aforesaid coefficient is all the more low as the “pitch” to be detected is drowned in the noise; for example, a signal to noise ratio of 0 decibels corresponds to a coefficient of 0.5.
• the validation information for the frequency pitch is obtained using the ratio of the synthesis sample, at the location of the pitch, to the sum of the modules used for the local synthesis; this ratio, synonymous with the energy of the harmonic signal over the total energy of the signal, is corrected as a function of the approximate signal to noise ratio calculated in block 14; the pitch validation information depends on exceeding the threshold of this report.
• the local synthesis is calculated twice; a first time by using only the frequencies whose modulus is high, in order to get rid of the noise for the calculation of the "pitch”; a second time with all the modules limited in maximum value, in order to calculate the signal to noise ratio which will validate the "pitch”; indeed the limitation of the modules gives more weight to the non harmonic frequencies with weak module, in order to decrease the probability of validation of a "pitch" on music.
• the values of said modules are not limited for the second local synthesis, only the number of frequencies is limited by taking into account only those which have a significant module in order to limit the noise.
• a second method of calculating the “pitch” consists in selecting the “pitch” which gives the maximum energy for a sampling step of the synthesis equal to the “pitch” sought; this process is used for music or a sound environment with several voices.
• a decision will be made by the user if he wishes to perform noise filtering or generate special effects (block 27), from the analysis, without going through the synthesis . Otherwise, the analysis will end with the following processing consisting in attenuating the noise, in block 28, by decreasing the frequency components which are not a multiple of the "pitch”; after attenuation of said frequency components, the inaudible signal will be removed again, as described above, at block 25.
• the attenuation of said frequency components depends on the type of signal as defined previously by block 21.
• phase of synthesis of the audio signal (block C3), represented according to FIG. 3, comprises the following steps:
• the synthesis consists in calculating the samples of the audio signal from the parameters calculated by the analysis; the phases and the noise will be calculated artificially according to the context.
• the formatting of the modules (block 31) consists in eliminating the attenuation of the input filter of the samples of the analysis (block 1 of block Al) and taking into account the direction of the variation of the "pitch" because the synthesis is performed temporally by a phase increment of a sine.
• the pitch validation information is deleted if the music synthesis option is validated; this option improves the phase calculation of the frequencies by avoiding to synchronize the phases of the harmonics between them according to the "pitch".
• the noise reduction (block 32) is carried out if it has not been previously carried out during the analysis (block 28 of block A1).
• the signal upgrade removes the normalization of the modules received from the analysis; this upgrade consists in multiplying the modules by the inverse of the normalization gain defined in the calculation of the signal dynamics (block 6 of block Al) and in multiplying said modules by 4 in order to eliminate the effect of the HAMMING window, and that only half of the frequency plane is used.
• the modules are saturated (block 34) if the sum of the modules is greater than the dynamic range of the signal of the output samples; it consists in multiplying the modules by the ratio of the maximum value of the sum of the modules to the sum of the modules, in the case where said ratio is less than 1.
• the pulse is re-generated by realizing the sum of sines in the pulse duration; the pulse parameters are modified (block 35) according to the variable synthesis speed.
• the frequency phase calculation is then performed (block 36); its purpose is to give phase continuity between the frequencies of the frames or to re-synchronize the phases between them; it also makes the voice more natural.
• Phase synchronization is performed each time a new signal in the current frame appears to be separated in the time domain or in the frequency domain of the previous frame; this separation corresponds to: • the transition from a noisy signal to a non-noisy signal,
• Phase continuity consists in finding the frequencies of the current frame at the start of the frame which are closest to the frequencies at the end of the frame of the previous frame; then the phase of each frequency becomes equal to that of the nearest previous frequency, knowing that the frequencies at the start of the current frame are calculated from the central value of the frequency modified by the variation of the "pitch".
• the phases of the harmonics will be synchronized with that of the pitch by multiplying the phase of the "pitch” by the index of the harmonic of the "pitch”; as for phase continuity, the phase of the pitch at the end of the frame is calculated as a function of its variation and of the phase at the origin of the frame; this phase will be used for the start of the next frame.
• a second solution consists in no longer applying the variation of the "pitch” to the "pitch” in order to know the new phase; it is enough to resume the phase of the end of the previous frame of the "pitch”; moreover, during the synthesis, the variation of the "pitch” is applied to the interpolation of the synthesis carried out without variation of the "pitch".
• the generation of the breath is then carried out (block 37).
• any sound signal in the interval of a frame is the sum of sines of fixed amplitude and the frequency of which is modulated linearly as a function of time, this sum being temporally modulated by the envelope of the signal, the noise being added to this signal before said sum.
• the voice is metallic because the elimination of weak modules, carried out in block 25 of block A3, essentially concerns breath. Furthermore, the estimation of the signal to noise ratio carried out in block 14 of block A3 is not used; noise is calculated as a function of the type of signal, the modules and the frequencies.
• the principle of the noise calculation is based on a filtering of white noise by a transversal filter whose coefficients are calculated by the sum of the sines of the signal frequencies whose amplitudes are attenuated as a function of the values of their frequency and their amplitude.
• a HAMMING window is then applied to the coefficients to reduce the secondary lobes.
• the filtered noise is then saved in two separate parts. A first part will make the link between two successive frames; the connection between two frames is made by overlapping these two frames, each of which is weighted linearly and in the opposite direction; said overlap is effected when the signal is sinusoidal; it does not apply when it is uncorrelated noise; thus the saved part of the filtered noise is added without weighting on the overlapping area.
• the second part is for the main body of the frame.
• the link between two frames must on the one hand allow a smooth passage between two noise filters of two successive frames, and on the other hand to prolong the noise of the following frame beyond the overlapping part of the frames if a start word (or sound) is detected.
• the smooth passage between two frames is achieved by the sum of the white noise filtered by the filter of the previous frame weighted by a linear downward slope, and the same white noise filtered by the noise filter of the current frame weighted by the slope reverse amount of that of the filter of the previous frame.
• the energy of the noise will be added to the energy of the sum of the sines, according to the proposed method.
• the generation of a pulse differs from a signal without a pulse; indeed, in the case of the generation of an impulse, the sum of the sines is not realized than on a part of the current frame to which is added the sum of the sines of the previous frame.
• the synthesis with the new frequency data (block 39) consists in performing the sum of the sines of the frequency components of the current frame; varying the length of the frame makes it possible to perform synthesis at variable speed; nevertheless the values of the frequencies at the beginning and at the end of the frame must be identical, whatever the length of the frame, for a given speed of synthesis.
• the phase associated with the sine, a function of the frequency will be calculated by iteration; indeed for each iteration, the sine multiplied by the module is calculated; the result is then summed for each sample according to all the frequencies of the signal.
• Another synthesis method consists in carrying out the inverse of the analysis by recreating the frequency domain from the cardinal sine produced with the module, the frequency and the phase, and then by performing a fast inverse Fourier transform (TFR), followed by the product of the inverse of the HAMMING window to obtain the time domain of the signal.
• TFR fast inverse Fourier transform
• the reverse of the analysis is carried out again by adding the variation of the "pitch" to the oversampled time frame.
• the phases at the origin of the frequency data are maintained at the value 0.
• the calculation of the sum of the sines is also carried out on a portion preceding the frame and on the same portion following the frame; the parts at the two ends of the frame will then be summed with those of the adjacent frames by linear weighting.
• the sum of the sines is performed in the time interval of generation of the pulse; in order to avoid the creation of spurious pulses following discontinuities in the calculation of the sum of the sines, a certain number of samples located at the beginning and at the end of the sequence are weighted respectively by an upward slope and a downward slope.
• the pitch pitch harmonic frequencies the phases were previously calculated to be synchronized, they will be generated from the index of the corresponding harmonic.
• the synthesis by the sum of the sines with the data of the previous frame (block 41) is carried out when the current frame contains a pulse to be generated; indeed, in the case of music or noise, if the synthesis is not carried out on the previous frame, serving as background signal, the pulse will be generated on a silence, which is detrimental to a good quality of the obtained result ; moreover, the continuity of the previous frame is inaudible, even in the presence of a signal progression.
• the application of the envelope to the synthesis signal (block 42) is carried out from the sampled values of the envelope previously determined.
• the length of the frame varies in steps in order to be homogeneous with the sampling of the envelope.
• the frame edge is saved (block 47) so that said frame edge can be added at the start of the next frame.
• the parameter coding phase (block A2), represented according to FIG. 4, comprises the following steps:
• the coding of the parameters (block A2) calculated in the analysis (block Al) in the method according to the invention consists in limiting the quantity of useful information in order to reproduce in the synthesis (block C3) after decoding (block Cl) a hearing equivalent to the original audio signal.
• each coded frame has its own number of information bits; the audio signal being variable, more or less information will be coded.
• the coding of the parameters can be either linear, the number of bits being a function of the number of values, or of HUFFMAN type, the number of bits being a statistical function of the value to be coded (the more frequent the data, the less it uses bits and vice versa).
• the type of signal as defined during the analysis (block 21 of block A1), provides the noise generation information and the quality of the coding to be used; the signal type is coded first (block 51).
• a test is then performed (block 52) allowing, in the case of type 3 of the signal, as defined in block 21 of the analysis (block A1), not to carry out coding of the parameters; the synthesis will include null samples.
• the compression type coding (block 53) is used in the case where the user wishes to act on the bit rate of the coding data, to the detriment of the quality; this option can be advantageous in telecommunication mode associated with a high compression rate.
• the coding of the normalization value (block 54) of the signal of the analysis frame is of the HUFFMAN type.
• a test on the presence of pulse (block 55) is then carried out, allowing in the event of synthesis of a pulse, to code the parameters of said pulse.
• the coding, according to a linear law, of the parameters of said pulse (block 56) will be carried out at the start and the end of said pulse in the current frame.
• the coding of the Doppler variation of the "pitch" (block 57), it will be carried out according to a logarithmic law, taking into account the sign of said variation; this coding will not be carried out in the presence of a pulse or if the type of signal is unvoiced.
• a limitation of the number of frequencies to be coded (block 58) is then carried out in order to avoid that a frequency of high value does not exceed the dynamic bounded by the sampling frequency, since the Doppler variation of the "pitch" varies the frequencies during the synthesis.
• the coding of the envelope sampling values depends on the variation of the signal, the type of compression, the type of signal, the normalization value and the possible presence of a pulse; said coding consists in coding the variations and the minimum value of said sampling values.
• the validation of the pitch is then coded (block 60), followed by a validation test (block 61) requiring, if so, to code the harmonic frequencies (block 62) according to their index relative to the frequency of the "Pitch". As for non-harmonic frequencies, they will be coded (block 63) according to their whole part.
• the coding of harmonic frequencies (block 62) consists in performing a logarithmic coding of the pitch, in order to obtain the same relative precision for each harmonic frequency; the coding of said harmonic indices is carried out as a function of their presence or of their absence in a packet of three indices according to the coding of HUFFMAN.
• a non-harmonic frequency changes position relative to a harmonic frequency
• the non-harmonic frequency which is too close to the harmonic frequency is removed, knowing that it has less weight in the sense audible; thus the suppression takes place if the non-harmonic frequency is higher than the harmonic frequency and the fraction of the non-harmonic frequency due to the coding of the whole part, makes said non-harmonic frequency lower than the near harmonic frequency.
• the coding of non-harmonic frequencies (block 63) consists in coding the number of non-harmonic frequencies, then the whole part of the frequencies, then the fractional parts when the modules are coded; concerning the coding of the whole part of the frequencies, only the differences between said whole parts are coded; moreover, the lower the modulus and the lower the precision on the fractional part; this is to decrease the bit rate.
• a maximum number of deviations between two frequencies is defined.
• the coding of the module dynamics uses a HUFFMAN law as a function of the number of ranges defining said dynamic and of the signal type.
• the energy of the signal is located in the low frequencies; for the other types of signal, the energy is distributed uniformly in the frequency plane, with a decrease towards the high frequencies.
• the coding of the highest module (block 65) consists in coding, according to a law of HUFFMAN, the entire part of said highest module taking into account the statistics of said highest module.
• the coding of the modules is only carried out if the number of modules to be coded is greater than 1, since otherwise it is alone by being the highest module.
• the suppression of the inaudible signal eliminates the modules lower than the product of the module by the corresponding attenuation; thus a module is necessarily located in a zone of the module / frequency plane depending on the distance which separates it from its two adjacent modules as a function of the frequency difference of said modules adjacent.
• the value of the module is approximated relative to the previous module as a function of • the frequency deviation and the corresponding attenuation which depends on the type of signal, the normalization value and the type of compression; said approximation of the value of the module is made with reference to a scale whose pitch varies according to a logarithmic law.
• the attenuation (block 67) provided by the sample input filter is coded, then is followed by the removal of the normalization (block 68) which makes it possible to recalculate the highest module as well as the corresponding frequency.
• the coding of the frequency fractions of the non-harmonic frequencies completes the coding of the whole parts of the said frequencies.
• the accuracy of the coding will depend:
• the number of coding bytes (block 70) is coded at the end of the coding of the various aforementioned parameters, stored in a dedicated coding memory.
• FIG. 1 representing a simplified flowchart of the method according to the invention
• the phase of noise filtering and generation of special effects, from the analysis, without going through the synthesis is indicated by block D.
• the filtering of the brait is carried out from the voice parameters calculated in the analysis (block Al of block A), taking the path IV indicated on said simplified flowchart of the method according to the invention.
• the objective of noise filtering is therefore to reduce all kinds of brait such as: ambient noise from car, engine, crowd, music, other voices if these are lower than those to be kept, as well than the calculation braits of any vocoder (for example: ADPCM, GSM, G723).
• the filtering of the brait (block D) for a voiced signal consists in carrying out the sum for each sample, of the original signal, of the original signal shifted by a "pitch" in positive value and of the original signal shifted by a "pitch” in negative value . This requires knowing for each sample the value of the pitch and its variation.
• the two offset signals are multiplied by the same coefficient, and the original signal not offset by a second coefficient; the sum of said first coefficient added to itself and of said second coefficient is equal to 1, denominated so as to retain a
• the number of samples spaced from a time pitch is not limited to three samples; the more samples used for the brait filter, the more the filter decreases the brait.
• the number of three samples is adapted to the highest temporal pitch encountered in the voice and to the filter delay. In order to keep a fixed filter delay, the lower the temporal "pitch", the more samples offset by a "pitch" can be used to perform the filtering; which amounts to keeping the bandwidth around a harmonic, more or less constant; the higher the fundamental, the higher the attenuated bandwidth.
• noise filtering does not concern signals in the form of a pulse; it is therefore necessary to detect the presence of possible pulses in the signal.
• Brait filtering (block D) for an unvoiced signal consists in attenuating said signal by a coefficient less than 1.
• the filtering of brait makes it possible to generate special effects; said generation of special effects makes it possible to obtain:
• the noise filtering and special effects generation phase from the analysis, without going through the synthesis, may not include the calculation of the variation of the "pitch"; this makes it possible to obtain a hearing quality close to that previously obtained by the aforementioned method; in this operating mode, the functions defined by blocks 11, 12, 15, 16, 17, 18, 19, 25 and 28 are deleted.
• phase of generation of special effects, associated with the synthesis (block C3) is indicated by block C2 of block C.
• the said phase of generation of special effects, associated with the synthesis makes it possible to transform the voice or the music: • either by modifying according to certain laws, the decoded parameters coming from block Cl (path II),
• the modified parameters are: the "pitch”, the variation of the "pitch”, the validation of the "pitch”, the number of frequency components, "the frequencies, the modules, • the clues.
• the "Transform" function consists in multiplying all the frequencies of the frequency components by a coefficient.
• the changes in the voice depend on the value of this coefficient, namely:
• this artificial rendering of the voice is due to the fact that the modules of the frequency components are unchanged and that the spectral envelope is deformed.
• the "Transvoice” function consists in recreating the harmonic modules from the spectral envelope, the original harmonics are abandoned knowing that the non harmonic frequencies are not modified; as such, said "Transvoice” function calls on the "Formant” function which determines the form.
• the transformation of the voice is carried out realistically because the form is preserved; a multiplication coefficient of harmonic frequencies greater than 1 rejuvenates the voice, even feminizes it; conversely, a multiplication coefficient of harmonic frequencies less than 1 makes the voice lower.
• the new amplitudes will be multiplied by the ratio of the sum of the modules at the input of said "Transvoice" function to the sum of the modules at the output.
• the "Formant" function consists in determining the spectral envelope of the frequency signal; it is used to keep the modules of the frequency components constant when the frequencies are modified.
• the determination of the envelope is carried out in two stages, namely:
• Said "Formant” function can be applied when coding the modules, frequencies, amplitude ranges and frequency fractions, by performing said coding only on the essential parameters of the formant, the "pitch" being validated.
• the frequencies and the modules are recalculated from the "pitch” and the spectral envelope respectively.
• the bit rate is reduced; however, this approach is only applicable by voice.
• the so-called "Transform” and “Transvoice” functions, described above, involve a constant frequency multiplication coefficient. This transformation can be non-linear and can make the voice artificial.
• this multiplication coefficient is a function of the ratio between the new "pitch” and the actual “pitch"
• the voice will be characterized by a fixed “pitch” and a variable forming; it will thus be transformed into a robot voice associated with a spatial effect.
• this multiplication coefficient varies periodically or randomly, at low frequency, the voice is aged associated with a very low frequency.
• a final solution is to perform fixed rate coding.
• the type of signal is reduced to a voiced signal (type 0 and 2 with the validation of the "pitch” to 1), or to brait (type 1 and 2 with the validation of the "pitch” to 0). Since type 2 is for music, it is eliminated in this case, since this coding can only code for voice.
• Fixed rate coding consists of:
• the "pitch" provides all the harmonics of the voice; their amplitudes are those of the trainer.
• frequencies of the unvoiced signal frequencies spaced apart are calculated by an average value to which is added a random deviation; the amplitudes are those of the trainer.
• the device according to the invention essentially comprises: “a computer 71, of DSP type, making it possible to carry out digital processing of the signals, A keyboard 72 for selecting the voice processing menus,
• a read-only memory 73 of EEPROM type, containing the voice processing software
• a random access memory 74 of the flash or “memory stick” type, containing the recordings of the voice processed
• An encoder / decoder 76 of coded type, ensuring the input / output links of the audio devices,
• the device may include:
• a telephone connector allowing the device according to the invention to replace a telephone handset
• the device may include:
• analysis means making it possible to determine parameters representative of said sound signal, the aforesaid analysis means comprising: • means for calculating the envelope of the signal, • means of calculating the pitch and its variation,
• TRF rapid Fourier transform
• the aforesaid synthesis means comprising: • means for summing the sinuses whose amplitude of the frequency components varies as a function of the envelope of the signal,
• means for generating special effects associated with the synthesis comprising:
• the device may include all the elements mentioned above, in professional or semi-professional version; some elements, such as the display, can be simplified in the basic version.
• the device according to the invention described above, will be able to exploit the method of differentiated digital processing of voice and music, of noise filtering and the creation of special effects.

Landscapes

• Engineering & Computer Science (AREA)
• Physics & Mathematics (AREA)
• Acoustics & Sound (AREA)
• Multimedia (AREA)
• Health & Medical Sciences (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Human Computer Interaction (AREA)
• Signal Processing (AREA)
• Computational Linguistics (AREA)
• Quality & Reliability (AREA)
• Spectroscopy & Molecular Physics (AREA)
• Compression, Expansion, Code Conversion, And Decoders (AREA)
• Electrophonic Musical Instruments (AREA)
• Soundproofing, Sound Blocking, And Sound Damping (AREA)
• Noise Elimination (AREA)
• Signal Processing Not Specific To The Method Of Recording And Reproducing (AREA)

Applications Claiming Priority (3)

Publications (2)

Family

ID=32696232

Family Applications (1)

Country Status (7)

Families Citing this family (28)

Family Cites Families (11)

• 2003
  • 2003-01-30 FR FR0301081A patent/FR2850781B1/fr not_active Expired - Fee Related
• 2004
  • 2004-01-27 WO PCT/FR2004/000184 patent/WO2004070705A1/fr active Application Filing
  • 2004-01-27 DE DE602004025903T patent/DE602004025903D1/de not_active Expired - Lifetime
  • 2004-01-27 AT AT04705433T patent/ATE460726T1/de not_active IP Right Cessation
  • 2004-01-27 ES ES04705433T patent/ES2342601T3/es not_active Expired - Lifetime
  • 2004-01-27 EP EP04705433A patent/EP1593116B1/de not_active Expired - Lifetime
  • 2004-01-27 US US10/544,189 patent/US8229738B2/en not_active Expired - Fee Related

Non-Patent Citations (1)

Also Published As

Similar Documents

Legal Events