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
  • 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
  • G10L19/04—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 predictive techniques
  • G10L19/08—Determination or coding of the excitation function; Determination or coding of the long-term prediction parameters
• • 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
  • G10L2019/0001—Codebooks
  • G10L2019/0011—Long term prediction filters, i.e. pitch estimation

Definitions

• the present invention relates to a method for predicting, in a so-called CELP speech coder, the residual vector signal, or residual vector, of the short-term analysis, said signal containing the periodicity information present in an initial speech signal to be coded. broken down into successive subframes, said prediction being carried out on the basis of optimal excitations predicted for the preceding subframe.
• CELP speech coder comprising on the one hand a short-term analysis filter, which receives an initial voice signal to be coded broken down into successive sub-frames and delivers a residual vector signal defining the information of periodicity present in the initial voice signal, and on the other hand a device for predicting this residual signal and a circuit for estimating a prediction error by difference between this residual vector signal and the predicted vector signal, more particularly said communication device prediction.
• the words emitted by a phonation organ constitute a vocal signal which presents two types of properties: on the one hand those related to the mechanism of the perception of this signal by the human auditory system (finite bandwidth, finite resolution in frequency, sensitivity at resonance frequencies, insensitivity to the phase of the frequency components of the signal, etc.), and on the other hand those linked to the operating mechanism of the phonation organ (pseudo-periodicity of the sounds, resonant structure of the signal, ).
• the voice message itself can be considered as the combination of content information and additional so-called expression information, which reflects individual variations in the acoustic presentation of the message. It is obvious that an effective transmission of such a message would undoubtedly imply defining a criterion of loyalty. It is however more realistic, in general, to be content with defining a perceptual criterion, which makes it possible to recognize the absence of discernible differences between a message sent and the corresponding received signal.
• the voice signal is, in fact, constituted by variations in air pressure, generated by the vocal tract under the action of the respiratory system which supplies the energy necessary for the production of speech.
• the air flow out of the lungs is modulated at a so-called fundamental frequency F with which the production of vowels is associated.
• This frequency which varies from 70 to 150 hertz approximately for men and from 150 to * ⁇ 00 hertz approximately for women, characterizes so-called voiced sounds (an example of representation of the amplitude A of a voiced sound as a function of time t is given in Figure 1).
• the air flow then excites in forced oscillations of the cavities of the vocal tract, to the shape of which correspond natural frequencies F. ,, F 2 , F- ,, etc ... called formants.
• the voice signal also includes signals which do not have the coherence of voiced sounds, but which are similar to noise, emitted by a source without natural frequency, and which do not excite the natural frequencies of the vocal tract. (these sounds are linked to the production of most consonants).
• the characteristics of the voice signal can be highlighted by a spectral analysis, which shows that the emitted spectrum comprises on the one hand a spectrum of lines (periodic excitations) for the production of the voiced sounds and on the other hand a continuous spectrum (excitations inconsistent) for producing unvoiced sounds.
• a further analysis of a voice signal ultimately shows that its processing with a view to a faithful transmission of the bandwidth which it represents leads to acoustically manipulating a considerable data stream.
• This coder is based on a principle called analysis by synthesis (in English: linear prediction analysis-by-synthesis coding), which comprises on the one hand an analysis step, to determine the coefficients of the synthesis filter, and on the other hand, a step of analysis by synthesis, which consists in finding or calculating a sequence of excitations minimizing a determined error criterion (we often use the criterion of least squares).
• analysis by synthesis in English: linear prediction analysis-by-synthesis coding
• CELP coding a term which will be adopted in the following description
• CELP coding is based on a simplified model of the speech production mode, according to which, as a first approximation, the voice signal can be modeled by a short-term (voice path) and long-term (voice source) correlation filter having a signal as input excitation.
• CELP CELP
• the draft recommendation G.723 cited above also uses so-called harmonic filtering.
• Short-term analysis is based on a predictive method in which the basic idea implemented is, knowing an input voice signal or observed signal s (n)
• the index n denotes the rank of the sample
• the speech coder described in the document G.723 cited above receives a signal consisting of blocks, or frames, comprising 240 samples, with a sampling frequency of 8 kHz, and each frame is supposed to be divided into four sub-frames. of 60 samples each.
• a filter 1 / A (z) called a synthesis filter, which, applied to an excitation signal x (n), makes it possible to obtain a signal p (n) as close as possible to the sampling signal s (n), it is equivalent to searching for a filter A (z), said to be of analysis, whose coefficients are such that the output signal of the filter is as bleached as possible when this filter is attacked by the signal to be transmitted (in an ideal coder, the output signal would be real white noise).
• the first performs a decorrelation of adjacent samples: its purpose is to define the coefficients of the input filter most suitable for obtaining after filtering the signal known, of a residual signal as close as possible to white noise.
• t (.) and p (.) denote respectively, the sample concerned in the signal to be modeled and the predicted sample
• ⁇ is a gain value
• OLP in English, Open Loop Pitch
• the determination of the period OLP and that of the gain ⁇ suffices to implement the prediction represented by the expression (3) -
• this determination has the form of a direct dependence: to within a gain factor, the sample to be predicted would be equal to one of the samples already occurring. In fact, the principle adopted is even more general.
• each of these groups of five vectors is called "V-vector".
• the first component of the first vector of each V-vector is obtained by shifting from
• the operation corresponding to expression (4) is therefore a prediction with linear combination of samples, during which the search for the solution vector is refined by varying the gain ⁇ , for example, in the case of the G project. 723 cited, by giving it five distinct values ( ⁇ is the gain vector formed by these five values), and also by adjusting the OLP quantity by a small value.
• the selection, during this research, of the best possible vector solution is made by including in the course of the determination process a minimization step, in the least squares sense, the difference between the vector t (n) of the analysis filter output (whose coefficients will be transmitted) and the solution vectors resulting from the implementation of expression (4).
• the analysis filter which, having received the voice signal to be transmitted, only delivers a residual signal constituting the periodicity information of this initial voice signal, it is this signal depleted t (n) for which the prediction explained below will be implemented with reference to FIG. 2.
• the adaptive dictionary therefore contains the excitation vectors candidates for the construction of an approximation of this vector t (n) .
• FIG. 2 shows, in the case of the G.723 project. an example of a prediction device making it possible to implement the principles of determination which have just been described.
• This device firstly comprises a circuit 20 for storing excitation vectors (this is the adaptive dictionary mentioned above), constructed from the optimal excitation of the preceding sub-frame, that is to say -to say excitations selected during a previous implementation of the same prediction method for previous samples.
• excitation vectors this is the adaptive dictionary mentioned above
• OLP + ⁇ CLP (in English, Closed Loop Pitch)
• circuit 20 is followed, in the case where one is in an even sub-frame and where ⁇ takes for example the three values - 1, 0 and + 1 (this is the case shown in FIG.
• channels 30 here three identical channels 30a, 30b, 30c (circuit 20 is followed by four identical channels, referenced in a similar way, in the case where one is in an odd sub-frame, ⁇ then taking the four values - 1, 0, + 1, + 2).
• Each of these channels processes the V-vector of the adaptive dictionary which corresponds to the step ⁇ of the channel considered, and includes in this firstly a series effect filter 31. having as response the impulse response of the synthesis filter (defined above).
• a circuit 32 also receiving the target vector to be modeled t (n) is then provided for the calculation of a vector with twenty components V ( ⁇ ) consisting of five correlation terms between the filtered vectors and the residual vector (in English: cross-products), given by the scalar products of the five filtered vectors of the adaptive dictionary by the target vector t (n), of five energy terms , given by the scalar products of the five filtered vectors of the adaptive dictionary by themselves, and ten two-by-two correlation terms between the filtered vectors. From these correlations, it is possible to determine to what extent the residual vector or target vector t (n) can be modeled from the V-vectors of the adaptive dictionary.
• the gains which are quantized, are provided by a memory * t0, or quantization table, which contains the possible values for the different gains (170 to 5.3 kbits / s., And 85 or 170 to 6.3 kbits / s., the 170 vectors of the table used in one or the other case then being the same).
• the information relating to the gains is given, in this quantification table, in the form of vectors each having twenty terms (as previously) defined as follows: five gain values, five values equal to the square of these gains, ten values corresponding to the ten two by two of these five gain values.
• the optimal gain vector ⁇ in the sense of expression (5) is the one which cancels the derivative of ⁇ with respect to each of the components of the vector, and which, by there, maximizes the scalar product of V ( ⁇ ) by a vector from the quantification table.
• a circuit 50 then allows the selection of the maximum scalar product among the three or four scalar products available at the output of these three (or four) circuits, maximum scalar product to which an optimal value of the step correspondent (stored in a memory 110) and an optimal value of the gain vector j3.
• the optimal value of ⁇ is of course one of the three (or four) values used in the three (or four) channels, and the value thus selected makes it possible to control a switch 60, comprising as many inputs as there are of tracks (three, or four). This switch, provided at the output of the filters 31. makes it possible to select the filtered V-vector constituting the best representative for the desired solution vector.
• This selected filtered vector is then presented at the input of an amplifier 70 whose gain vector j3, delivered by the selection circuit 50, had been stored in a memory 80 present at the output of this circuit 50.
• a first aim of the invention is therefore to propose a simpler prediction method, with practically equal quality, than that previously described.
• the invention relates to a prediction method comprising for each of said subframes the following steps: (1) for different values of a step ⁇ said of determining said periodicity information, and with a view to selecting, from said previously predicted optimal excitations, an optimal gain vector p opt and the corresponding value of not optimal, a step of carrying out, in series and for each value of ⁇ , the following substeps:
• (c) a preselection sub-step, for determining an initial gain vector ⁇ ( ⁇ ) init , squares of the components of said vector, and of the products of these same components taken two by two, delivering a second vector P ⁇ ⁇ i n i t •
• Another object of the invention is to propose a speech coder similar to that which has just been described, but with reduced complexity and while retaining practically equivalent quality.
• the invention relates, in an encoder as defined in the preamble to the description, a prediction device comprising:
• (A) a vector storage circuit, called an adaptive dictionary, containing the optimal excitations predicted for the previous sub-frame;
• (D) a memory, called a quantization table, which contains the components of the candidate gain vectors, as well as their squares and their products in pairs;
• (E) a selector switch, in connection with said optimal value of the step, of the corresponding optimal vector V ( ⁇ ) composed of correlation terms, given by the scalar products of the vectors filtered by the residual vector of energy terms, given by the dot products of the vectors filtered by themselves, and of correlation terms, given by the dot products two by two of the filtered vectors; (F) a circuit for selecting, in said quantization table, the candidate gain vectors;
• (I) a subtractor for estimating said prediction error by difference between said residual signal t (n) and the predicted signal p (n) delivered by said amplifier.
• the solution according to the invention makes it possible to make a much lower number of courses (four in the case encoder according to this document), with practically insensible quality degradation.
• the basic idea of the structure thus proposed is indeed the following: by making the simplifying assumption that the predictors are decorrelated, or, which is equivalent, that the correlations between filtered vectors taken two by two, previously defined, are zero, we can, for each sub-frame and for each channel corresponding to a step ⁇ , define an initial gain vector P ( ( 5) init , without costly matrix inversion calculation. The components of this vector are then, for the channel considered, the successive ratios of the correlation terms between the filtered vectors and the target vector and the energy terms of the vector V ( ⁇ ) defined above.
• an initial vector gain ⁇ ( ⁇ J-init llows then determine an initial ⁇ value (and suboptimal) route requiring only the quantization table per subframe
• the optimal initial step determined for each subframe by the calculation subset provided at the output of the channels, is then that which maximizes one of the three, or four depending on whether the subframe is even or odd , scalar products of ⁇ ⁇ ) ⁇ n it ⁇ ar V ( ⁇ ), determined by the calculation circuit provided in each channel after the calculation of V ( ⁇ ) (we recall here that V ( ⁇ ) was defined previously, while that ⁇ (° " ) i n it is a vector which includes the same number of components as V ( ⁇ ), namely, in this case: the five
• the vector of the optimal gain for each sub-frame is then obtained using the circuit provided at the output of the quantization table and which performs the search for the maximum value of the scalar product of the vector V ( ⁇ ) corresponding to the optimal value the step ⁇ which has just been calculated by each of the vectors (here 170 or 85 depending on the flow) of said table.
• This determination of the optimal gain vector requires only one scan of the quantization table per subframe, ie four per frame (instead of three or four per subframe depending on whether the subframe is even or odd. , or fourteen per frame), which results in a significant reduction in complexity.
• FIG. 1 shows an example of representation of the amplitude of a voiced sound as a function of time
• FIG. 2 and 3 show the structure of a closed-loop voicing period prediction device, respectively in the case of the cited G723 document and in the case of the present invention.
• the prediction device presents, with that of FIG. 2, common elements, namely the circuit 20 for storing the candidate excitation vectors (or adaptive dictionary), the filters 31. the circuits 32 for calculating the correlation and energy terms, the memory 40 (or quantification table), the switch 60, the amplifier 70, the memory 80, the subtractor 90, and the memory 110.
• the search for the optimal gain vector is then carried out by the selection circuit 150, and the vector thus selected is kept in memory 80.
• the switch 60 provided at the output of the filters 31 and whose position is controlled by the value of ⁇ issue from memory 110, sends the selected filtered V-vector to amplifier 70.
• the optimal filtered V-vector thus amplified is the prediction vector p (n), sent, as in the case of FIG. 2, to the subtractor 90.

Landscapes

• Engineering & Computer Science (AREA)
• Computational Linguistics (AREA)
• Signal Processing (AREA)
• Health & Medical Sciences (AREA)
• Audiology, Speech & Language Pathology (AREA)
• Human Computer Interaction (AREA)
• Physics & Mathematics (AREA)
• Acoustics & Sound (AREA)
• Multimedia (AREA)
• Compression, Expansion, Code Conversion, And Decoders (AREA)

Priority Applications (3)

Applications Claiming Priority (4)

Publications (1)

Family

ID=26144072

Family Applications (1)

Country Status (4)

Cited By (1)

Citations (2)

• 1996
  • 1996-10-11 EP EP96934902A patent/EP0796490B1/fr not_active Expired - Lifetime
  • 1996-10-11 WO PCT/FR1996/001596 patent/WO1997014139A1/fr active IP Right Grant
  • 1996-10-11 DE DE69609592T patent/DE69609592T2/de not_active Expired - Fee Related
  • 1996-10-11 JP JP9514782A patent/JPH11500837A/ja active Pending

Patent Citations (2)

Non-Patent Citations (4)

Also Published As

Similar Documents

Legal Events