EP1665234B1 - Information flow transmission method whereby said flow is inserted into a speech data flow, and parametric codec used to implement same - Google Patents

Abstract

For the transmission of a secondary information flow between a transmitter and a receiver, the secondary information flow is inserted at a parametric vocoder of the transmitter which generates a main information flow. The main information flow is a speech data flow encoding a speech signal and is transmitted from the transmitter to the receiver. Bits from the secondary information flow are inserted into only some of the frames of the main information flow, these frames being selected by a frame mask which is known to the transmitter and the receiver, and/or into a determined frame of the main information flow, by imposing a constraint on only some of the bits of the frame, these bits being selected by a bit mask known to the emitter and the receiver.

Description

The present invention relates generally to the field of speech coding, and in particular to a method of inserting an information flow within a speech data stream, the information flow. inserted may be a stream of lower speed speech data or a transparent data stream.

The invention finds applications, in particular, in public or professional mobile radiocommunication systems (PMR systems, "Professional Mobile Radiocommunication").

A speech signal is an acoustic signal emitted by a human voice apparatus.

A codec is a hardware and / or software unit for coding and decoding a digital stream. Its coding function makes it possible to transcode a digital stream of quantized samples in the time domain of a source signal (for example a speech signal) into a compressed digital stream. Its decoding function makes it possible to perform a pseudo-inverse operation in order to restore attributes representative of the source signal, for example perceptible attributes in a receiver such as the human ear.

A speech data stream is a data stream generated by a speech codec, from the encoding of a speech signal. A transparent data stream is a binary digital sequence whose content type is unspecified, whether it is a computer data stream or a speech data stream. The data are said to be transparent in that, from an external point of view, all the bits have equal importance vis-à-vis, for example of the correction of the transmission errors so that an error-correcting coding must therefore be uniform on all the bits. Conversely, if the stream is a speech data stream, some bits are more important to protect than others.

A speech codec, also called vocoder (in English "Vocoder""SpeechCodec" or "Voice Codec") is a specialized codec which is adapted to the encoding of a quantized speech signal and to the decoding of a stream of frames. lyrics. In particular, it presents for its coding function a sensitivity which depends on the characteristics of the speech of the speaker and a low bit rate associated with a more limited frequency band than the general audio frequency band (20 Hz-20 kHz).

There are several families of speech coding techniques, including coding techniques of the speech signal waveform (eg ITU-T G.711 MIC A / mu law coding), coding techniques for speech coding. source model (the best known being the CELP coding, Code-Excited Linear Prediction), perceptual encodings, and hybrid techniques based on the combination of techniques belonging to at least two of the above families.

The invention is directed to the application of "source model" coding techniques. These techniques are also called parametric coding techniques, because they are based on the representation of excitation parameters of the speech source and / or parameters describing the spectral envelope of the signal emitted by the speaker (for example according to a model of linear prediction coding exploiting the correlation between the consecutive values of the parameters associated with a synthesis filter, or according to a cepstral model) and / or source-dependent acoustic parameters, for example the amplitude and the perceived fundamental central frequency ( "Pitch" in English), the period ("Pitch period" in English) and the amplitude of the energy peaks of the first harmonics of a pitch frequency at different intervals, its degree of voicing ("voicing rate" in English). ), its melody and its sequences.

A vocoder is called a vocoder that implements digital speech coding using a parametric model of the speech source. In practice, such a vocoder associates several parameters to each frame of the speech stream. First linear prediction spectral parameters also called, for example, LP coefficients (Linear Prediction) or Linear Prediction Coding (LPC) coefficients, which define the vocoder linear prediction filter ( short-term filter). Secondly, adaptive excitation parameters associated with one or more adaptive excitation vectors, also known as LTP (Long Term Predictor) parameters or adaptive prediction coefficients, which define a long term in the form of a first excitation vector and a gain associated with applying the input of the synthesis filter. And, thirdly, fixed excitation parameters associated with (or more) fixed excitation vector (s), also called algebraic parameters or stochastic parameters which define a second excitation vector and an associated gain to be applied at the input of the synthesis filter.

Of the document EP-A-1,020,848 there is known a method for transmitting auxiliary information in a main information stream corresponding to a speech signal, said auxiliary information being inserted in the vocoder CELP which encodes the speech signal, replacing the index of the adaptive excitation vector and / or index of the fixed excitation vector. More specifically, the auxiliary information bits are inserted in the vocoder of the transmitter in place of the bits normally encoding the corresponding index, and the value of the gain is set to zero in order to inform the vocoder of the receiver.

According to a drawback, the insertion of an auxiliary information stream into the stream is not discrete, in that it is sufficient to note the null value of the gain to know that the bits normally allocated to the coding of the associated index actually contain the auxiliary information. This is considered a disadvantage for the implementation of the method in a system in which the confidentiality of the transmissions is important.

The document US 2001/038643 discloses a method of inserting a secondary information stream into a main information stream, wherein sub-bands of an audio signal corresponding to the main stream that can contain data of the secondary stream are determined. This selection of subbands is performed according to the characteristics of the audio signal in question, such as the signal-to-noise ratio in the subbands under consideration. Then, for a selected subband, the number of coding bits available for inserting data of the secondary stream is determined. Again, this determination makes use of audio signal characteristics such as the difference between the scale factor and the noise floor level in the subband. In the end, values associated with subbands are masked with data of the secondary information stream to be transmitted.

The main object of the invention is to allow the discrete insertion of a secondary stream into a main stream corresponding to a speech stream. Other objects of the invention are aimed at maximizing the rate of the insertable secondary stream, while at the same time preserving the performance of the coding of the main stream with respect to attributes of the source (ie by preserving the perceived quality of the source stream). hearing during speech flow synthesis). Another object of the invention is also to simultaneously preserve the performance of the coding of the secondary flow with respect to attributes of the source of the secondary flow, especially when it is also a speech flow.

Some or all of these objects are achieved according to a first aspect of the invention by a method of transmitting a secondary information stream between an emitter and a receiver according to claim 1.

The transmitter and the receiver, as well as the transmission, must be interpreted in their broadest sense. In an exemplary application to a radio system, the transmitter and the receiver are terminal equipment of the system, and the transmission is a radio transmission.

The insertion is performed at a parametric vocoder transmitter that produces said main information stream, without changing the bit rate of the latter compared to what it would be without insertion. In other words, the secondary information flow is interpreted as a sequence of constraints on the sequence of values of certain parameters of the parametric coding model of the main information flow. Compared to the insertion method known in the prior art, the method according to the invention has the advantage that nothing in the main information stream that is transmitted betrays the presence of the secondary information flow inserted. Moreover, by limiting the insertion to certain frames and / or bits in a frame only, the intelligibility of the speech signal encoded in the main information stream is preserved, which is not the case with the aforementioned known insertion method.

In order to reinforce the discretion of the insertion, and therefore the robustness vis-à-vis attempts to hack the transmission, the frame mask can be variable. It is then generated according to a common parallel algorithm in the transmitter and in the receiver, in order to synchronize the coding and the decoding of the main information stream, respectively in the transmitter and in the receiver.

The frame mask may advantageously define a subset of groups of consecutive frames in each of which bits of the secondary information stream are inserted, in order to take advantage of the sliding effect of the coding which results from the storage of the frames in the frame. parametric vocoder. This helps to preserve the fidelity of the main information flow to the speech signal.

Preferably, the number of frames of a group of consecutive frames is then substantially equal to the depth of storage of the frames in the parametric vocoder.

When the source model of the parametric vocoder provides, for at least some of the frames of the main information stream, different bit classes according to their sensitivity to the quality of speech coding, the mask of bits may be such that bits of the secondary information stream are inserted into these frames by imposing a priority constraint on the bits belonging to the least sensitive bit class. This also helps to preserve the fidelity of the main information flow to the speech signal.

The secondary information stream may be a speech data stream having a lower rate than the main information rate. This is the case when the secondary information stream leaves another vocoder having a lower rate than the parametric vocoder bit rate.

Of course, the secondary information flow can also be a transparent data stream.

When the rate of the secondary information flow to be inserted is too high compared to the rate of the parametric vocoder, it may be necessary to remove bits from the secondary information stream, if this is compatible with the application. Conversely, if the secondary information flow rate is too low, it is possible to repeat certain bits or to introduce stuffing bits.

The secondary information flow is subjected to error correction coding before insertion into the main information flow. This makes it possible to overcome the fact that, in the context of parametric vocoders, certain bits of the frames of the main information stream are weakly or even not subject to an error correcting coding (forming channel coding) prior to transmission.

In one possible embodiment, bits of the secondary information stream are inserted by imposing values on bits that belong to excitation parameters of a filter of the source model of the parametric vocoder, for example parameters of adaptive excitation and / or fixed excitation parameters of the linear prediction filter of a CELP vocoder. The fact of not imposing a constraint on the bits of the linear prediction parameters preserves the intelligibility of the main information flow. For this purpose also, it is preferred to impose constraints on the bits forming the adaptive excitation parameters rather than on those forming the fixed excitation parameters.

In one embodiment, bits of the secondary information stream may also be inserted into silence frames of the main information stream, instead of or in addition to insertion into speech frames.

In another embodiment, bits of the secondary information stream may be inserted by imposing constraints on unencrypted bits as end-to-end encryption of the main information stream. This allows a receiving equipment can, after extraction, decode the secondary information stream although not having the ability to decrypt as such. Of course, the bits concerned may nevertheless undergo one or more encryption / decryption operations in another respect, for example link ciphers or radio interface ciphers.

For example, the insertion constraint may be a constraint of equality of the bits of the frame of the main information stream with the bits of the inserted secondary information stream.

A second aspect of the invention relates to a parametric vocoder according to claim 13, adapted for carrying out the method according to the first aspect. With respect to its coding function, such a parametric vocoder includes insertion means for inserting a secondary information stream into a main information stream that is generated by the vocoder parametric from a speech signal.

For its decoding function, the vocoder includes means for extracting the secondary information stream from the main information stream.

A third aspect of the invention further relates to a terminal equipment of a radio communication system comprising a parametric vocoder according to the second aspect.

• la figure 1 est un diagramme illustrant un exemple de flux de données de paroles codées (flux de parole) organisé en trames et sous-trames ;
• la figure 2 est un schéma synoptique partiel d'un exemple d'équipement émetteur selon l'invention ;
• la figure 3 est un schéma synoptique partiel d'un exemple d'un vocodeur selon l'invention ; et
• la figure 4 est un schéma synoptique partiel d'un exemple de vocodeur utilisé dans équipement récepteur selon l'invention.

Other features and advantages of the invention will become apparent on reading the description which follows. This is purely illustrative and should be read in conjunction with the attached drawings in which:

• the figure 1 is a diagram illustrating an example of coded speech data flow (speech flow) organized into frames and subframes;
• the figure 2 is a partial block diagram of an exemplary transmitter equipment according to the invention;
• the figure 3 is a partial block diagram of an example of a vocoder according to the invention; and
• the figure 4 is a partial block diagram of an example of vocoder used in receiver equipment according to the invention.

The figure 1 is a diagram illustrating the general principle of inserting a secondary data stream DS2 into a main data stream DS1 encoding a speech signal VS1. This insertion is carried out at a transmitter which, after multiplexing and channel coding, transmits the DS1 stream, and therefore the DS2 stream that it contains, to a remote receiver. Such a transmitter and such a receiver are for example mobile terminals of a public radio system such as GSM or UMTS, or a professional radio system such as TETRA or TETRAPOL.

The stream DS1 is generated by a vocoder 10 from the speech signal VS1, which is produced by a speech source 1 such as the voice apparatus of an individual. For this purpose, the speech signal VS1 is digitized according to a linear encoding (pulse modulation coding) PCM, and segmented into frames called speech frames. In addition, each frame is generally segmented at vocoder 10 into a fixed number M of segments called subframes in the time domain (CELP model) or in the frequency domain (MBE model). Excitation"). Typically M is between 2 and 6, depending on the vocoders). Each frame comprises a determined number N of bits.

The figure 2 illustrates a digitized speech signal segmented into successive F [i] frames, for i ranging from 0 to infinity. In addition, at least for certain parameters, each frame F [i] can be segmented into M subframes SF [m], for m between 1 and M. In the figure, D is the duration of a frame .

Back to the figure 1 , the vocoder 10 may be an EFR vocoder (of the English "Enhanced Full Rate") of the GSM system (see specification ETSI EN 300 726 GSM 06.60), a vocoder AMR (of the English "Adaptive Multi-Rate"") of the UMTS system (see ETSI specification 3GPP TS26.101) for which D = 20 ms and M = 4, a vocoder of a TETRA radiocommunication terminal in accordance with EN 300 395-2 of the ETSI, or a 6 kbit / s TETRAPOL vocoder (referenced in ITU-R report M.2014) for which the number D = 20 ms, M = 3 and N = 120.

The secondary data stream DS2 is for example generated by a codec 20, which receives a data stream to be coded from a source 2. In an exemplary application of the invention, the source 2 also transmits a speech signal, the codec 20 then being a bitrate vocoder less than that of the cocodener 10. In this case, the stream DS2 is also a stream of speech frames. In this application, the invention allows the discrete insertion of a secondary communication in a main communication. The codec 20, more specifically the vocoder 20, may be a MF / MELP type vocoder (of the "Multi-Frame - Mixed Excitation Linear Prediction") at 1200/2400 bits / s described in NATO STANAG 4591.

Optionally, the stream DS2 may be subjected to an error-correcting coding, for example a CRC (of the English "Cyclic Redundancy Code") or a convolutional coding, which forms a channel coding for transmission through the transmission channel. In fact, it is known that certain bits of the frames of the DS1 speech stream are little or not protected by channel coding, so that specific protection of the bits of the DS2 information stream may be required, depending on the applications.

The vocoder 10 comprises an encoder 100 which implements a source model coding algorithm (or parametric model), for example of the CELP or MELP type. In such a case, the parameters corresponding to the coding of a transmitter-side speech frame include, among others, excitation vectors which are subjected, on the receiver side, to a filter whose response models the speech.

Parametric coding algorithms use parameters calculated either directly as a function of the stream of incoming speech frames and an internal state of the vocoder, or calculated by iterations (on successive frames and / or subframes) by optimizing a criterion given. Typically, the first parameters include the linear prediction (LP) parameters defining a short-term filter, and the second parameters include the adaptive excitation (LTP) parameters defining a long-term filter and the fixed excitation parameters. Each iteration corresponds to the coding of a sub-frame in a frame of the input stream.

Thus, for example, the adaptive excitation parameters and the fixed excitation parameters are selected by successive iterations to minimize the quadratic error between the synthesized speech signal and the original VS1 speech signal. In the Anglo-Saxon literature, this iterative selection is sometimes called "Codebook search" or "Analysis by Synthesis Search", or "Error Minimization Loop" or "Closed Loop Pitch Analysis".

In general, the adaptive excitation parameters and / or the fixed excitation parameters may each comprise, on the one hand, an index corresponding to a value of a vector in the adaptive dictionary (depending on the subframe) or in a fixed dictionary, respectively, and on the other hand a gain value associated with said vector. Nevertheless, in certain vocoders such as the TETRAPOL vocoder, the parameters of at least one of the adaptive and fixed excitations directly define the excitation vector to be applied, that is to say without addressing a dictionary by an index. In what follows, no distinction is made between the mode of definition of the excitation vectors. The constraints imposed by the bits of the stream DS2 applying either to the index relating to the value of the excitation vector in the dictionary, or to the value of the excitation itself.

In addition to the main data stream (speech frame stream) VS1 and the secondary data stream DS2, the vocoder 10 receives, according to the invention, a stream TS of frame masks, and / or a stream BS of bit masks .

The stream FS is generated by a frame mask generator 3, from a bit stream received from a pseudo-random generator 5, which operates from a secret key Kf known to the transmitter and the receiver. A frame mask has the function of selecting, from among a certain number of frames of the stream of DS1 speech frames, those in which only the bits of the secondary data stream DS2 are inserted.

For this purpose, the generator 3 executes the following process. Let the sequence of the frames F [i] of the main stream DS1, let h be a numerical function with integer values, and let k be a determined integer, which is preferably substantially equal to the depth of storage of successive frames in the vocoder 10 ( see further, number P, with reference to the diagram of the figure 3 ), then the frames F [h (i)], F [h (i) +1], ..., F [h (i) + k] define what is called here a subset of groups of frames of the sequence of frames F [i].

According to a preferred embodiment of the invention, the frames undergoing the insertion constraint are frames belonging to a subset of groups of consecutive frames of the main stream DS1. This makes it possible to take advantage of the sliding effect of the speech coding resulting from the storage of frames provided in the vocoder 10, in order to preserve the quality of the coding of the speech signal VS1 in the main stream DS1. For this reason, the number k, which corresponds to the frame length of a group of frames, is preferably equal to or at least close to the storage depth R of the vocoder 10, as it has been said above.

For example, by choosing h (i) = 10 xi and k = 5, then the frames F [0] to F [5] undergo the insertion constraint, the frames F [6] to F [9] do not undergo the insertion constraint, the frames F [10] to F [15] undergo the insertion constraint, the frames F [16] to F [19] do not undergo the insertion constraint, etc. In other words, in this example, 6 out of 10 consecutive frames undergo the insertion constraint.

The stream BS is generated by a bit mask generator 4, from a bit stream received from a pseudo-random generator 6, which operates from a secret key Kb, also known from the transmitter and receiver. A bit mask has the function of selecting, from among the N bits of a frame of the speech frame stream DS1 selected by virtue of the frame mask associated with the current F [i] frame, those which are only constrained by bits of the secondary data stream DS2.

For this purpose, the generator 4 executes the following process. It produces a flow of a fixed number Smax bits, where Smax denotes the maximum number of bits of a current frame Fi of the main stream DS1 that can be constrained by bits of the secondary stream DS2. A determined number S of among these Smax bits, where S is less than or equal to Smax (S≤Smax), have the logic value 1, the others having the logic value 0. These Smax bits are inserted in an N-bit string at predefined and fixed positions which are provided in the vocoder software 10, so as to form a bit mask on the frame. This mask, called a bit mask, thus comprises S bits equal to 1. In one example, when a bit of the bit mask is equal to 1, it indicates a position for inserting a bit of the secondary stream DS2 into the current frame Fi of the main stream DS1.

The number Smax is fixed by making a compromise between the maximum number of bits of the secondary stream DS2 that can be inserted into a frame of the main stream DS1, on the one hand, and the concern to preserve the quality of the coding of the speech signal. VS1 in the main stream DS1, on the other hand. Since the number Smax is fixed, the number S depends on the flow rate of the secondary stream DS2. The S / N ratio defines what can be called the insertion rate of the secondary stream DS2 in the main stream DS1 for the current frame F [i], the ratio Smax / N defining the maximum insertion rate.

In an example where a TETRAPOL vocoder (for which N = 120) is used with h (i) = 10 xi, k = 5 and S = 50, a mean flow channel of 1215 bit is obtained for insertion of the secondary flow. / s. Such a bit rate allows the insertion of a secondary data stream generated by a MF-MELP type codec at 1200 bit / s (requiring 81 bits in 67.5 ms) described in NATO STANAG 4591. In other words, the rate of The insertion obtained is sufficient to discretely transmit a secondary stream which is also a speech stream generated by a secondary vocoder 20 of a rate lower than that of the main vocoder 10.

An example of an insertion constraint consists in replacing (ie, overwriting) the bits of the main stream DS1 normally generated according to the standard coding algorithm implemented by the vocoder 10 from the speech signal VS1, by bits of the stream secondary DS2. In other words, the constraints applied to the speech coding parameters of the main stream are equality constraints with the bits of the second stream, combined with logical AND operation selection constraints applying a bit mask on the bits forming the main stream. .

This example is the simplest, but it is not the only one. Indeed, algorithms on the main stream and the secondary stream using any contextual grammar or linear or non-linear algebra, including Boolean algebra and Allen's time algebra (see the article " Maintaining Knowledge about Temporal Intervals ", Communications of the ACM, 26/11/1983, pp. 832-84 ), any auxiliary memories and dependent on the value of third parameters, allow the skilled person to define complex constraints, which respect for example statistical properties imposed by the speech model of the main stream.

Note in particular that the set of indices of excitations in a dictionary generally has a distribution of bits at 0 and at 1 completely neutral with respect to a statistical analysis of occurrences. It is usually possible to encrypt the DS2 secondary stream in a pseudo-random form before insertion, without changing the statistical distribution of the 0s and 1s in the modified bits of the main stream. Assuming a speech coding model leading to a coded stream of which some subframes would have a correlation to 0 or 1, the aforementioned pseudo-random generator or a secondary stream encryption algorithm will also have this bias.

As will be understood, the number of bits constrained during coding varies from one frame to another according to a known evolution law of the transmitter and the receiver, which are supposed to be synchronized.

The synchronization of the transmitter and the receiver with regard to the application of the frame masks and / or bit masks results from the general synchronization between these two devices. Typically, this synchronization is ensured by the labeling of the frames using values generated by a frame counter. In a known manner, the general synchronization between the transmitter and the receiver may also originate, wholly or in addition, synchronization elements (particular bit patterns) inserted into the main stream DS1.

The encoder 100 of the transmitter and the decoder of the receiver share the same initial information making it possible to determine the subsequence of the groups frames and subframes where the insertion of the secondary flow takes place. This information may comprise an initialization vector of the pseudo-random generators 5 and 6. It may be fixed. It may also depend, for example, on the average rate imposed by the secondary stream, or may depend on unconstrained parameters of the main codec 10 calculated during the coding of the main stream.

As he is represented at figure 3 , the encoder 100 comprises a module 11 which is a hardware and / or software module for synthesis of the linear prediction parameters, receiving as input the speech signal VS1 and outputting an LP information corresponding to the linear prediction parameters (filter coefficients short-term linear prediction). The LP information is passed to the input of a logical unit 12, for example a multiplexer, which is controlled by the frame mask stream FS and the bit mask stream BS. The unit 12 outputs LP information LP 'corresponding to the LP information, some bits of which at least for at least some frames, have been altered by applying the constraints resulting from the secondary stream DS2 via the frame mask and the bit mask associated with the current frame. A memorization of the information LP ', with a depth of memorization corresponding to a predetermined number P of successive frames, may be provided for the module 11.

The encoder 100 also comprises a module 21 which is a hardware and / or software module for synthesizing the adaptive excitation parameters, receiving as input the information LP 'and outputting an LTP information corresponding to the adaptive excitation parameters (defining a first quantization vector and an associated gain for the short-term synthesis filter). The LTP information is passed to the input of a logic unit 22, for example a multiplexer, which is controlled by the frame mask stream FS and the bit mask stream BS. The unit 22 outputs LTP 'information corresponding to the LTP information of which at least some bits for certain frames and / or for at least some subframes have been altered by applying the constraints resulting from the secondary stream DS2 via the frame mask and the bit mask associated with the current frame. A storage of the information LTP ', with a storage depth corresponding to a given number Q of successive subframes of the current frame (Q≤M-1), may be provided for the module 21.

The encoder 100 finally comprises a module 31 which is a hardware and / or software module for the synthesis of the fixed excitation parameters, receiving as input the information LTP 'and outputting a FIX information corresponding to the fixed excitation parameters (defining a second quantization vector and an associated gain for the short-term synthesis filter). The FIX information is passed to the input of a logical unit 32, for example a multiplexer, which is controlled by the frame mask stream FS and the bit mask stream BS. The unit 32 outputs FIX information corresponding to the FIX information of which at least some bits for certain frames and / or for at least some subframes have been altered by applying the constraints resulting from the secondary stream DS2 via the frame mask and the bit mask associated with the current frame. A storage of the information FIX ', with a storage depth corresponding to a determined number R of successive subframes of the current frame (R≤M-1), is provided for the module 21. In addition, a storage of the information FIX ', with a storage depth corresponding for example to a determined number W of successive subframes of the current frame (W≤M-1), may be provided for the module 21.

For each current frame, the information LP '(F [i]) corresponding to the parameters of linear prediction of the frame, the information LTP' (SF [1]), ..., LTP '(SF [M] corresponding to adaptive excitation parameters respectively for each of the subframes SF [1] to SF [M] of the frame, and the information FIX '(SF [1]), ..., FIX' (SF [M] corresponding to fixed excitation parameters respectively for each of the subframes SF [1] to SF [M] of the frame, are transmitted at the input of a multiplexer 41 which concatenates them to form a frame of the main stream DS1.

The memorizations mentioned above make it possible to attenuate the effect of the constraints applied to the bits of the linear prediction parameters, the adaptive excitation parameters and / or the fixed excitation parameters, with respect to the fidelity of the main stream DS1 to the source speech signal VS1. Indeed, these memorizations allow a sliding effect in the calculation of the parameters, so that, for a given frame, the stresses applied to the first parameters are at least partially compensated, from the perceptual point of view, by the calculation of calculated parameters. then from a speech synthesis based on said first parameters.

• 1°) LP'(F[i])=f(LP'(F[i-1]), LP'(F[i-2]),..., LP'(F[i-P]) ;
• 2°) LTP'(SF[i])=f(LTP'(SF[i-1]),..., LTP'(SF[i-R]), FIX'(SF[i-1]),..., FIX'(SF[i-W]) ;
• 3°) FIX'(SF[i])=f(FIX'(SF[i-1]),..., FIX'(SF[i-W]).

More specifically, we can write the following relations, where f denotes a function translating the analysis by synthesis:

• 1 °) LP '(F [i]) = f (LP' (F [i-1]), LP '(F [i-2]), ..., LP' (F [iP]);
• 2 °) LTP '(SF [i]) = f (LTP' (SF [i-1]), ..., LTP '(SF [iR]), FIX' (SF [i-1]) ,. .., FIX '(SF [iW]);
• 3 °) FIX '(SF [i]) = f (FIX' (SF [i-1]), ..., FIX '(SF [iW]).

These compensations, and also the fact that the insertion of the bits of the secondary flow is not random, make it possible to achieve in practice, for some vocoders, insertion rates of the order of 10% without generating degradation ( from the perceptual point of view) of the speech signal VS1 greater than what generates a residual bit error rate (after channel coding) of the order of a few%.

The implications of the receiver-side method will now be described.

Note first of all that, for a receiver equipment not processing the secondary stream DS2, the decoding of the frames of the stream DS1 received, is only performed according to the standard synthesis algorithm of the vocoder 10 of the transmitting equipment.

For a receiving equipment processing the secondary stream DS2, the recovery of the information coded by the bits of this secondary stream requires a synchronization of the equipment with the sending equipment, means for extracting the secondary stream DS2 from the stream main DS1. identical to the codec 20 of the sending equipment.

We refer to the diagram of the figure 4 , which schematically shows the means of a receiver device vocoder 10a for the processing of the secondary stream transmitted by the method according to the invention.

The vocoder 10a, if necessary after demultiplexing and decoding the channel, receives the main stream DS1 at the input, and delivers a speech signal VS1 'at the output.

The signal VS1 'is less faithful to the source speech signal VS1 ( figure 3 ) that it would be in the absence of implementation of the insertion method according to the invention. This reflects the loss of quality of the coding performed on the transmitter side, because of the external constraints applied to the vocoder 1 of the transmitting equipment.

The receiving equipment may also include means for reproducing the speech signal VS1 ', for example a loudspeaker or the like.

As already mentioned above, the known transmission protocols provide for a general synchronization of the receiving equipment with the transmitting equipment. The implementation of the invention does not require any particular means in this respect.

For the extraction of the secondary stream, the vocoder 10a comprises a frame mask generator 3a and a bit mask generator 4a, respectively associated with a pseudo-random generator 5a and a pseudo-random generator 6a, which are identical and arranged in the same way as the means 3, 4, 5 and 6 respectively of the vocoder 10 of the sending equipment ( figure 3 ). Note that the generators 5a and 6a of the receiver receive the same secret key, respectively Kf and Kb, as the generators 5 and 6 of the vocoder 10 of the transmitter equipment. These keys are stored in an ad hoc memory of equipment. The generators 3a and 4a respectively generate a frame mask stream FSa and a stream of bit masks BSa, which are supplied at the input of a decoder 100a of the vocoder 10a.

The extraction of the bits of the secondary stream DS2 is done by synchronous application (for example via logical AND operations) of the frame masks and input bit masks of the decoder 100a (for example via logical AND operations), without that affects the decoding of the main stream DS1 by the latter. For this purpose, the stream DS1 is provided at the input of the decoder 100a via a logic unit 7a, which extracts the secondary information stream DS2 from the main information stream DS1 under the control of the frame mask stream FSa and the stream of information. BSa bit masks.

The receiving equipment may also include a secondary codec identical to the codec 20 of the transmitting equipment for decoding the secondary stream DS2. When this stream is a speech stream, the secondary codec generates a speech signal that can be output via a speaker or the like.

It will be noted that the fluctuation of the bit transmission rate of the secondary stream DS2 does not pose any particular problem on the receiver side, since the secondary stream DS2 is supplied as input to a variable rate secondary codec, as is the case with all the vocoders of the market. Indeed, such a codec comprises an input buffer ("Input Buffer" in English) in which DS2 stream data are stored for decoding. Just make sure that the input buffer is never empty. For this purpose, the appropriate insertion rate is determined, taking into account, in particular, the bit rate of the encoder 100 and the secondary vocoder 20 and the objectives of preserving the fidelity of the main stream VS1 to the speech signal VS1. Given the high insertion rates obtained in practice (of the order of 10%), this question of the secondary vocoder supply of the receiving equipment should not be a problem, with a main vocoder 10 AMR type in its 12.2 kbit / s coding mode and a secondary rate vocoder about ten times less.

Moreover, in the case where the secondary stream is a speech stream and in order to provide the second decoder a regular stream of frames, it is possible optionally to memorize the sequences and not to start the decoding immediately.

In the case where the secondary stream is a transparent data stream, it is proposed to concatenate and treat them as if they had been transmitted by means of a short messenger of maximum length (SMS service in GSM, for example), and to add a convolutive error correction code. Alternatively, the transparent data stream may be sent to an encryption module or to a "Text-to-Speech" type of transcoding and synthesis module.

Returning now to the general description of the methods of implementation of the transmission method according to the invention.

The choice of the bits of a given frame of the main stream that undergo the application of the constraint of the secondary stream is determined according to the particularities of each application. The following are several possible embodiments in this respect, as well as other features and advantages of the invention.

In one possible embodiment, constraints are imposed during the coding on the value of zero, several or all the bits of the frame which are associated with an excitation vector of determined, adaptive or fixed type before performing the iterations making it possible to calculate the parameters which depend on said excitation vector by virtue of the memorizations made in the vocoder. These constrained value bits are then the information of the secondary stream transported by the frame and constitute the channel of the secondary information stream DS2. In other words, the secondary stream is inserted by imposing values on bits forming the parameters of the adaptive or fixed excitation vectors. This may possibly be extended by applying constraints simultaneously to the excitation vectors of the other type, respectively fixed or adaptive.

When the transmission between the transmitter and the receiver provides partial encipherment of the frames of the main stream (i.e., encryption of only some bits in each frame), the bit mask may advantageously coincide with a set of unencrypted bits of a frame. This allows the receiving equipment acting as gateway to extract the secondary stream inserted in the main stream without having the means to decrypt the main stream.

This is particularly useful while preserving the confidentiality of the main stream, under the approximate linearity of the vocoder speech model, ie considering that the residual or excitation parameters of the vocal cords are uncorrelated. the coefficients describing the spectral envelope of response of the vocal tract.

In other words, this embodiment of the method is characterized in that the secondary information stream is inserted by imposing constraints on unencrypted bits of parameters of the speech pattern of the main stream.

This mode of implementation is illustrated by an example concerning an EFR vocoder (see above) used as the main codec. One chooses to use bits among the unprotected bits of each frame as a channel for the secondary stream, by overwriting their value calculated by the source stream encoding algorithm of the main stream by applying a bit mask on the unprotected 78 bits. of each frame. These unprotected 78 bits are identified in Table 6 (entitled "Ordering of Enhanced Full Speech Parameters for the Channel Encoder" in ETSI EN 300 909 V8.5.1 GSM 05.03 "Channel Coding") and relate to a subset of bits describing the fixed excitation vectors. With these 78 class 2 bits per 20 ms frame, a secondary channel with a nominal bit rate of 3900 bit / s is obtained. The least sensitive bits of the 12.2 kbit / s encoding mode of the AMR codec (see above) preferably identified in order of sensitivity can be used in Table B.8 (entitled "Ordering of the Speech Encoder Bits from the 12.2 kbit / s Mode "in the 3GPP specification TS26.101" Adaptive Multi-Rate (AMR) Speech Codec Frame Structure ").

It is therefore also possible to introduce, in the 12.2 kbit / s coding mode of the AMR codec, the stream of a secondary codec, for example the 1200/2400 bit / s MELP coder described in NATO STANAG 4591, requiring 81 bits per 67.5 ms at 1200 bits / s (respectively 54 bits by 22.5 ms 2400 bits / s), embedded in its own error correcting coding (2/3 FEC rate), for example, which protects 100% of the bits at 1200 bit / s (respectively 50% of the bits at 2400 bit / s), and or embedded in NATO-defined Future Narrow Band Digital Terminal (FNBDT) security interoperability frames, or a lighter type of security protocol.

In another embodiment, applicable to vocoders using an algorithm based on the selection of quantized excitations in a dictionary, the constraint consists in imposing a determined excitation value, taken from the dictionary. In a variant, the dictionary is partitioned into several sub-dictionaries, and the constraint consists in imposing one of the sub-dictionaries. Another variant comprises the combination of the two types of constraint above. During the decoding of the main stream on the receiver side, the knowledge of the received excitation makes it possible to identify the sub-dictionary and / or the excitation concerned, and to deduce the constraint which determines the bits of the secondary stream. Note that at a permutation close to the excitations, the imposition constraint of the sub-dictionary can be equivalent to the application of the constraints on the low-order bits of the excitation indices in the dictionary.

In another embodiment, the secondary stream defines a differential coding of the excitation vector indices, for example fixed excitation vectors, in the subsequence of successive frames of the main stream.

In another embodiment, the constrained bits may be the least significant bits of the fixed excitations (i.e. nonadaptive excitations) for each speech frame and possibly for each subframe defined in FIG. speech frame within the meaning of the vocoder coding algorithm 10.

In another embodiment, the number and position of the constrained bits are identified for each successive frame according to an algorithm for calculating a mask and a known secret element of the transmitter and the receiver, to increase the chances of non-detection of the existence of the secondary stream by a third party.

Another embodiment, applicable to an encoding algorithm requiring several fixed excitation vectors per frame or subframe, such as the CELP codec for the speech of an MPEG-4 stream (defined in the ISO / IEC specification). IEC 14496-3 Subpart 3) for which some fixed excitations of a frame are chosen from previous calculations and where other fixed excitations of the same frame are calculated by synthesis analysis on a dictionary (see the ISO specification / IEC 14496-3 §7.9.3.4 "Multi-Pulse Excitation for the bandwidth extension tool"), consists in imposing the constraint on the dictionary choice of the first fixed excitation and then using the iterations of analysis by synthesis on the second fixed excitation to catch the error imposed by the stress on the first fixed excitation.

In another embodiment, the subset of the frames of the main stream that are concerned by the insertion of the secondary stream includes only the frames that have sufficient energy and speech in the vocoder sense. In an alternative applicable for example to MELP vocoders (which define several levels of voicing) or vocoder HVXC (Harmony Vector excitation Codec), which are parametric vocoders of an MPEG-4 speech stream defined in the ISO / IEC 14496-3 Subpart specification 2) the subsequence only applies to the unvoiced or totally unvoiced segments of the frames.

When the constraint is applied to the excitation parameters, for example to the fixed excitation indices, the parameters of a subframe of the main stream DS1 remain entirely in accordance with the speech coding model of vocoder 10. Nevertheless, the sequence of modified fixed excitations may be statistically atypical for human speech or possibly atypical for the speaker recognition method, depending on the constraints applied and the desired fidelity objective. In order to prevent the presence of the secondary flux in these excitations from being detected in a receiver equipment, a parameter processing comprising a smoothing of the gains of the fixed excitations associated with a treatment of the pulses isolated from the excitation vectors followed by a post-excitation filtering after speech synthesis, can be applied to decoding. These treatments make it possible to exclude acoustic sequences appearing after transmission in a noisy channel, which would be impossible to pronounce by a human voice device in the ambience of a microphone. It is for example certain sequences of rattling, hissing, screeching, hissing or other, in the background noise that the standard vocoder would not have sufficiently filtered during speech synthesis because of the constraints imposed. Thus undesired unvoiced sounds can be made imperceptible, which would be correlated with the fixed excitation sequences constrained according to the method of the invention.

Nevertheless, when the application of constraints risks leading to the perception of unwanted voiceless sounds correlated with an atypical fixed excitation sequence of human speech and unfiltered by the filtering of the standard decoder of the vocoder, the sub-sequel frames on which the constraints are applied can be defined according to preliminary statistical analyzes on the values of the consecutive parameters of the vocoder speech model, for example by taking advantage of the texture of the speech parameters, defined by an inertia, entropy or energy derived from the probability of parameter value sequences, for example in eight consecutive frames representative of the duration of a phoneme.

For each mode of implementation, the performance of the synthesis of the main stream DS1, that is to say the fidelity to the signal VS1, is inversely proportional to the relative flow rate of the secondary stream DS2. The subjective fidelity performance at source 1 of speech signal VS1 can however be reached when the proposed method keeps invariant some subjective attributes (for example certain psychoacoustic criteria) of source 1. It can be measured by statistical measurements ( "Mean Opinion Score", or MOS) according to a standardized scale (see ITU-T Recommendation P.862 "Perceptual evaluation of speech quality -PESQ").

In some embodiments, the degradation of the subjective quality of the DS1 speech stream from the vocoder 10, which is due to the insertion of the secondary stream DS2, is assumed to be acceptable to justify the application of the proposed method. This is particularly the case when the secondary stream is also a speech stream and the auditory content of the main stream is much less important than the content of the secondary stream for the legitimate listener. Indeed, the psycho-acoustic perception of the possible presence of the flow secondary when listening to the main stream decoded and restored does not help to locate the secondary stream in the main stream and therefore to provide a formal proof of its existence. This is particularly the case for a low rate vocoder 10 used in a noisy environment, since the decoding and the rendering of the main stream DS1 provide speech sequences conforming to the vocoder model 10. This is also the case in some psychoacoustic limits, when the minimum flow rate of the secondary flow must be ensured to the detriment of the quality of return of the main flow.

In order to best preserve the intelligibility of the synthesis of the main stream DS1, it is preferred not to apply constraints on the linear prediction spectral parameters (LP) defining the short-term filter, and not to disturb the long-term parameters too much ( LTP) adapted to each subframe, in order to maintain subjective characteristics deemed essential in the speech signal VS1. In particular, an implementation mode consists in applying the constraints on subframes different from the sub-frames on which the long-term analysis windows of the frame are concentrated, namely, for example, the second one. and the fourth subframe for the 12.2 kbit / s coding mode of the AMR vocoder discussed above (see 3GPP specification TS 26.090 V5.0.0, §5.2.1 "Windowing and auto-correlation computation"). In particular, it will avoid disturbing many voiced segments, generally carrying the majority of speaker identification characteristics.

As a developed example, in the 12.2 kbit / s coding mode of the AMR vocoder, it is possible to impose a constraint on the choice of the adaptive excitation by imposing initial values on the samples u (n) n = 0, ..., 39, in the recursive equation (38) for calculating the adaptive vector described in paragraph 5.6.1 (entitled "Adaptive Codebook Search") of the specification 3GPP TS 26.090 mentioned above, substituting for the values of the LP residue, calculated in equation (36), 40 values extracted from the secondary stream. The error between the signal of the main stream and the signal synthesized by the short-term filter with the contribution of the constrained adaptive vector is compensated by the choice of the fixed excitation vector which tries to catch the residual error (for example the error residual quadratic) of the long-term prediction on the same subframe, as well as the excitation vectors of the successive subframes. Thus the constrained excitation vectors encode the secondary flow as an adaptive residue above the response of the short-term synthesis filter of the main flux corrected by the fixed residue.

In another example, for a speech model of the parametric vocoder type STC (English "Sinusoidal Transform Coding") or type MBE ("Multi Band Excitation") for example according to the standard specifications ANSI / TIA / EIA 102 .BABA ("APCO Project 25 Vocoder Description"), an implementation mode leads to interest in the least significant bits of the harmonic amplitude parameters of the frame segments or the sample amplitude parameters of the sample. spectral envelope. In an MBE codec, the excitation parameters are the fundamental frequency as well as the voiced / unvoiced decision for each frequency band.

In the foregoing, embodiments have been described providing for the insertion of the bits of the secondary stream into speech frames of the main stream. Nevertheless, it is known that the main stream DS1 also contains frames of silence, which are frames coded by the vocoder 10 with a lower bit rate and transmitted with a periodicity less than the speech frames, to synthesize the periods of silence contained in the speech signal VS1. These frames of silence synthesize what is called a noise of comfort.

However, one embodiment of the method may provide, alternatively or in addition, the insertion of the secondary flow via numerical constraints on the values of the descriptors parameters of the comfort noise to be generated under the main flow.

This mode of implementation is illustrated by an example concerning an EFR or AMR codec (see above) used as the main codec. In the GSM and UMTS systems, the frames carrying comfort noise (silence frames) are called SID frames (see, for example, the specification 3GPP TS 26.092 "Mandatory Speech Codec Speech Processing Functions"; AMR Speech Codec; Comfort Noise Aspects ";'AND IF). More specifically, Frames considered here are SID-UPDATE frames that contain 35-bit comfort noise parameters and 7-bit error correction code.

In a GSM or UMTS system, it is the source that controls the transmission of the silence frames, that is to say the codec of the transmitter (subject to interactions with the voice activity detection process and discontinuous transmission, particularly on the downlink from the relay to the mobile terminal). It is therefore possible to proceed by inserting the second stream according to a method similar to that applicable to a frame containing sufficient speech energy (speech frame).

Alternatively, it is possible to control the transmission of a particular silence frame from the digitized analog input of the codec by generating the comfort analog noise representative of the 35 bits of the secondary stream. In GSM and UMTS systems, the frequency of the silence frames is controlled by the source or the relay and corresponds either to a silence frame every 20 ms or to a silence frame every 160 ms, or to a frame of silence every 480 ms for the codec EFR of the GSM system. This determines the maximum flow rate for the secondary flow in this variant of the process.

In a particular mode, it is possible to use the duplex transmission channel to send frames of silence when the speaker is a second participant in the communication or in silences in a first conversation, that is to say between groups of phonemes emitted according to the main flow.

It should be noted that the specification 3GPP TS 26.090 specifies that the size of the coding field of the comfort noise of the EFR codec, namely 35 bits per silence frame, is identical to the size of the fixed excitation parameter for this same codec. This means that one can apply the same constraints and obtain a permanent minimum insertion rate using all the frames regardless of the nature, speech or silence, of the main stream.

Claims (18)

1. Method of transmitting a secondary information stream (DS2) between a sender and a receiver, the method including inserting said secondary information stream in a parametric vocoder (100) of the sender generating a main information stream (DS1) that is a voice data stream coding a voice signal and is transmitted from the sender to the receiver, in which method bits of the secondary information stream are inserted:

   • into only some of the frames (F[i]) of the main information stream selected by a frame mask known to the sender and to the receiver, and

   • into a selected frame of the main information stream, at predefined positions by imposing a constraint on only some of the bits of the frame selected by a bit mask known to the sender and to the receiver;
   wherein the frame mask defines a subseries (SF[m]) of groups of consecutive frames in each of which bits of the secondary information stream are inserted; and,
   the length in frames of a group (M) of consecutive frames is substantially equal to the depth of storage of the frames in the parametric vocoder.
2. Method according to claim 1, wherein the frame mask is variable and is generated in parallel in the sender and in the receiver using a common algorithm.
3. Method according to any preceding claim, wherein, the source model of the parametric vocoder providing, for at least some of the frames of the main information stream, different classes of bits as a function of their sensitivity to the quality of voice signal coding, the bit mask is such that bits of the secondary information stream are inserted into these frames, by imposing a constraint as a matter of priority on the bits belonging to the least sensitive bit class.
4. Method according to any one of claims 1 to 3, wherein the secondary information stream is a voice data stream from another vocoder (20) having a lower bit rate than the parametric vocoder.
5. Method according to any one of claims 1 to 3, wherein the secondary information stream is a transparent data stream.
6. Method according to anyone of the preceding claims, wherein the secondary information stream is subjected to error corrector coding before inserting it into the main information stream.
7. Method according to anyone of the preceding claims, wherein bits of the secondary information stream are inserted by imposing values on bits that belong to excitation parameters of a filter of the source model of the parametric vocoder.
8. Method according to anyone of the preceding claims, wherein bits of the secondary information stream are inserted into silence frames of the main information stream.
9. Method according to anyone of the preceding claims, wherein bits of the secondary information stream are inserted by imposing constraints on unencrypted bits in relation to end-to-end encryption of the main information stream.
10. Method according to any preceding claim, wherein the constraint is a constraint of equality of the bits of the frame of the main information stream and the inserted bits of the secondary information stream.
11. Parametric vocoder (100) including, for inserting a secondary information stream (DS2) into a main information stream (DS1) that is generated by the parametric vocoder from a voice signal, insertion means adapted to insert bits of the secondary information stream:

    • into only some of the frames (F[i]) of the main information stream selected by a particular frame mask known to the sender and to the receiver, and/or

    • into a selected frame of the main information stream, at predefined positions by imposing a constraint on only some of the bits of the frame selected by a particular bit mask known to the sender and to the receiver ;
    wherein the frame mask defines a subseries of consecutive (SF[m]) frames into each of which bits of the secondary information stream are inserted; and,
    wherein the length in frames of the subseries of consecutive frames is substantially equal to the depth of storage of the frames in the parametric voice codec.
12. Parametric vocoder according to claim 13, wherein the frame mask is variable and is generated by an algorithm based on a secret key.
13. Parametric vocoder according to any one of claims 11 and 12, wherein, the source model of the parametric vocoder providing, in at least some of the frames of the main information stream, different classes of bits as a function of their sensitivity to the quality of voice signal coding, the bit mask is such that bits of the secondary information stream are inserted into these frames, by imposing a constraint as a matter of priority on the bits belonging to the least sensitive bit class.
14. Parametric vocoder according to any one of claims 11 to 13, further including means for subjecting the secondary information stream to error corrector coding before inserting it into the main information stream.
15. Parametric vocoder according to any one of claims 11 to 14, wherein the insertion means are adapted to insert bits of the secondary information stream by imposing values on bits that belong to excitation parameters of a filter of the source model of the parametric vocoder.
16. Parametric vocoder according to any one of claims 11 to 15, wherein the insertion means are adapted to insert bits of the secondary information stream into silence frames of the main information stream.
17. Parametric vocoder according to any one of claims 11 to 16, wherein the insertion means are adapted to insert bits of the secondary information stream by imposing constraints on unencrypted bits in relation to end-to-end encryption of the main information stream.
18. Terminal equipment of a radio system including a parametric vocoder according to any one of claims 11 to 17.

Images