Patent Grant
7457742
This application is the U.S. national phase of the PCT/FR2003/003870 filed Dec. 22, 2003, which claims the benefit of French Application No. 03 00164 filed Jan. 8, 2003, the entire content of which is incorporated herein by reference.
The invention relates to devices for coding and decoding audio signals, intended in particular to sit within applications of transmission or storage of digitized and compressed audio signals (speech and/or sounds).
More particularly, this invention pertains to audio coding systems having the capacity to provide varied bit rates, also referred to as multirate coding systems. Such systems are distinguished from fixed rate coders by their capacity to modify the bit rate of the coding, possibly during processing, this being especially suited to transmission over heterogeneous access networks: be they networks of IP type mixing fixed and mobile access, high bit rates (ADLS), low bit rates (RTC, GPRS modems) or involving terminals with variable capacities (mobiles, PCs, etc.).
Essentially, two categories of multirate coders are distinguished: that of “switchable” multirate coders and that of “hierarchical” coders.
“Switchable” multirate coders rely on a coding architecture belonging to a technological family (temporal coding or frequency coding, for example: CELP, sinusoidal, or by transform), in which an indication of bit rate is simultaneously supplied to the coder and to the decoder. The coder uses this information to select the parts of the algorithm and the tables relevant to the bit rate chosen. The decoder operates in a symmetric manner. Numerous switchable multirate coding structures have been proposed for audio coding. Such is the case for example with mobile coders standardized by the 3GPP organization (“3rd Generation Partnership Project”), NB-AMR (“Narrow Band Adaptive Multirate”, Technical Specification 3GPP TS 26.090, version 5.0.0, June 2002) in the telephone band, or WB-AMR (“Wide Band Adaptive Multirate”, Technical Specification 3GPP TS 26.190, version 5.1.0, December 2001) in wideband. These coders operate over fairly wide bit rate ranges (4.75 to 12.2 kbit/s for NB-AMR, and 6.60 to 23.85 kbit/s for WB-AMR), with a fairly sizeable granularity (8 bit rates for NB-AMR and 9 for WB-AMR) . However, the price to be paid for this flexibility is a rather considerable complexity of structure: to be able to host all these bit rates, these coders must support numerous different options, varied quantization tables etc. The performance curve increases progressively with bit rate, but the progress is not linear and certain bit rates are in essence better optimized than others.
In so-called “hierarchical” coding systems, also referred to as “scalable”, the binary data arising from the coding operation are distributed into successive layers. A base layer, also called the “kernel”, is formed of the binary elements that are absolutely necessary for the decoding of the binary train, and determine a minimum quality of decoding.
The subsequent layers make it possible to progressively improve the quality of the signal arising from the decoding operation, each new layer bringing new information which, utilized by the decoder, supplies a signal of increasing quality at output.
One of the particular features of hierarchical coding is the possibility offered of intervening at any level whatsoever of the transmission or storage chain so as to delete a part of the binary train without having to supply any particular indication to the coder or to the decoder. The decoder uses the binary information that it receives and produces a signal of corresponding quality.
The field of hierarchical coding structures has given rise likewise to much work. Certain hierarchical coding structures operate on the basis of one type of coder alone, designed to deliver hierarchized coded information. When the additional layers improve the quality of the output signal without modifying the bandwidth, one speaks rather of “embedded coders” (see for example R. D. Lacovo et al., “Embedded CELP Coding for Variable Bit-Rate Between 6.4 and 9.6 kbit/s, Proc. ICASSP 1991, pp. 681-686). Coders of this type do not however allow large gaps between the lowest and the highest bit rate proposed.
The hierarchy is often used to progressively increase the bandwidth of the signal: the kernel supplies a baseband signal, for example telephonic (300-3400 Hz), and the subsequent layers allow the coding of additional frequency bands (for example, wide band up to 7 kHz, HiFi band up to 20 kHz or intermediate, etc.). The subband coders or coders using a time/frequency transformation such as described in the documents “Subband/transform coding using filter banks designs based on time domain aliasing cancellation” by J. P. Princen et al. (Proc. IEEE ICASSP-87, pp. 2161-2164) and “High Quality Audio Transform Coding at 64 kbit/s”, by Y. Mahieux et al. (IEEE Trans. Commun., Vol. 42, No. 11, November 1994, pp. 3010-3019), lend themselves particularly to such operations.
Moreover, a different coding technique is frequently used for the kernel and for the module or modules coding the additional layers, one then speaks of various coding stages, each stage consisting of a subcoder. The subcoder of the stage of a given level will be able either to code parts of the signal that are not coded by the previous stages, or to code the coding residual of the previous stage, the residual is obtained by subtracting the decoded signal from the original signal.
The advantage of such structures it that they make it possible to go down to relatively low bit rates with sufficient quality, while producing good quality at high bit rate. Specifically, the techniques used for low bit rates are not generally effective at high bit rates and vice versa.
Such structures making it possible to use two different technologies (for example CELP and time/frequency transform, etc.) are especially effective for sweeping large bit rate ranges.
However, the hierarchical coding structures proposed in the prior art define precisely the bit rate allocated to each of the intermediate layers. Each layer corresponds to the encoding of certain parameters, and the granularity of the hierarchical binary train depends on the bit rate allocated to these parameters (typically a layer can contain of the order of a few tens of bits per frame, a signal frame consisting of a certain number of samples of the signal over a given duration, the example described later considering a frame of 960 samples corresponding to 60 ms of signal).
Moreover, when the bandwidth of the decoded signals can vary according to the level of the layers of binary elements, the modification of the line bit rate may produce artifacts that impede listening.
The present invention has the aim in particular of proposing a multirate coding solution which alleviates the drawbacks cited in the case of the use of existing hierarchical and switchable codings.
The invention thus proposes a method of coding a digital audio signal frame as a binary output sequence, in which a maximum number Nmax of coding bits is defined for a set of parameters that can be calculated according to the signal frame, which set is composed of a first and of a second subset. The proposed method comprises the following steps:
The allocation and/or the order of ranking of the Nmax−N0 coding bits are determined as a function of the coded parameters of the first subset. The coding method furthermore comprises the following steps in response to the indication of a number N of bits of the binary output sequence that are available for the coding of said set of parameters, with N0<N≦Nmax:
The method according to the invention makes it possible to define a multirate coding, which will operate at least in a range corresponding for each frame to a number of bits ranging from N0 to Nmax.
It may thus be considered that the notion of pre-established bit rates which is related to the existing hierarchical and switchable codings is replaced by a notion of “cursor”, making it possible to freely vary the bit rate between a minimum value (that may possibly correspond to a number of bits N less than N0) and a maximum value (corresponding to Nmax). These extreme values are potentially far apart. The method offers good performance in terms of effectiveness of coding regardless of the bit rate chosen.
Advantageously, the number N of bits of the binary output sequence is strictly less than Nmax. What is noteworthy about the coder is then that the allocation of the bits that is employed makes no reference to the actual output bit rate of the coder, but to another number Nmax agreed with the decoder.
It is however possible to fix Nmax=N as a function of the instantaneous bit rate available on a transmission channel. The output sequence of a switchable multirate coder such as this may be processed by a decoder which does not receive the entire sequence, so long as it is capable of retrieving the structure of the coding bits of the second subset by virtue of the knowledge of Nmax.
Another case where it is possible to have N=Nmax is that of the storage of audio data at the maximum coding rate. When reading N′ bits of this content stored at lower bit rate, the decoder would be capable of retrieving the structure of the coding bits of the second subset as long as N′≧N0.
The order of ranking of the coding bits allocated to the parameters of the second subset may be a preestablished order.
In a preferred embodiment, the order of ranking of the coding bits allocated to the parameters of the second subset is variable. It may in particular be an order of decreasing importance determined as a function of at least the coded parameters of the first subset. Thus the decoder which receives a binary sequence of N′ bits for the frame, with N0≦N′≦N≦Nmax, will be able to deduce this order from the N0 bits received for the coding of the first subset.
The allocation of the Nmax−N0 bits to the coding of the parameters of the second subset may be carried out in a fixed manner (in this case, the order of ranking of these bits will be dependent at least on the coded parameters of the first subset).
In a preferred embodiment, the allocation of the Nmax−N0 bits to the coding of the parameters of the second subset is a function of the coded parameters of the first subset.
Advantageously, this order of ranking of the coding bits allocated to the parameters of the second subset is determined with the aid of at least one psychoacoustic criterion as a function of the coded parameters of the first subset.
The parameters of the second subset pertain to spectral bands of the signal. In this case, the method advantageously comprises a step of estimating a spectral envelope of the coded signal on the basis of the coded parameters of the first subset, and a step of calculating a curve of frequency masking by applying an auditory perception model to the estimated spectral envelope, and the psychoacoustic criterion makes reference to the level of the estimated spectral envelope with respect to the masking curve in each spectral band.
In a mode of implementation, the coding bits are ordered in the output sequence in such a way that the N0 coding bits of the first subset precede the N−N0 coding bits of the selected parameters of the second subset and that the respective coding bits of the selected parameters of the second subset appear therein in the order determined for said coding bits. This makes it possible, in the case where the binary sequence is truncated, to receive the most important part.
The number N may vary from one frame to another, in particular as a function for example of the available capacity of the transmission resource.
The multirate audio coding according to the present invention may be used according to a very flexible hierarchical or switchable mode, since any number of bits to be transmitted chosen freely between N0 and Nmax may be selected at any moment, that is to say frame by frame.
The coding of the parameters of the first subset may be at variable bit rate, thereby varying the number N0 from one frame to another. This allows best adjustment of the distribution of the bits as a function of the frames to be coded.
In a mode of implementation, the first subset comprises parameters calculated by a coder kernel. Advantageously, the coder kernel has a lower frequency band of operation than the bandwidth of the signal to be coded, and the first subset furthermore comprises energy levels of the audio signal that are associated with frequency bands higher than the operating band of the coder kernel. This type of structure is that of a hierarchical coder with two levels, which delivers for example via the coder kernel a coded signal of a quality deemed to be sufficient and which, as a function of the bit rate available, supplements the coding performed by the coder kernel with additional information arising from the method of coding according to the invention.
Preferably, the coding bits of the first subset are then ordered in the output sequence in such a way that the coding bits of the parameters calculated by the coder kernel are immediately followed by the coding bits of the energy levels associated with the higher frequency bands. This ensures one and the same bandwidth for the successively coded frames as long as the decoder receives enough bits to be in possession of information of the coder kernel and coded energy levels associated with the higher frequency bands.
In a mode of implementation, a signal of difference between the signal to be coded and a synthesis signal derived from the coded parameters produced by the coder kernel is estimated, and the first subset furthermore comprises energy levels of the difference signal that are associated with frequency bands included in the operating band of the coder kernel.
A second aspect of the invention pertains to a method of decoding a binary input sequence so as to synthesize a digital audio signal corresponding to the decoding of a frame coded according to the method of coding of the invention. According to this method, a maximum number Nmax of coding bits is defined for a set of parameters for describing a signal frame, which set is composed of a first and a second subset. The input sequence comprises, for a signal frame, a number N′ of coding bits for the set of parameters, with N′≦Nmax. The decoding method according to the invention comprises the following steps:
The allocation and/or the order of ranking of the Nmax−N0 coding bits are determined as a function of the recovered parameters of the first subset. The decoding method furthermore comprises the following steps:
This method of decoding is advantageously associated with procedures for regenerating the parameters which are missing on account of the truncation of the sequence of Nmax bits that is produced, virtually or otherwise, by the coder.
A third aspect of the invention pertains to an audio coder, comprising means of digital signal processing that are devised to implement a method of coding according to the invention.
Another aspect of the invention pertains to an audio decoder, comprising means of digital signal processing that are devised to implement a method of decoding according to the invention.
The coder represented in
The second coding stage 2, making it possible to increase the bandwidth towards the wide band (50-7000 Hz), operates on the coding residual E of the first stage, supplied by a subtractor 3 in the diagram of
The audio signal to be coded S has for example a bandwidth of 7 kHz, while being sampled at 16 kHz. A frame consists for example of 960 samples, i.e. 60 ms of signal or two elementary frames of the coder kernel G.723.1. Since the latter operates on signals sampled at 8 kHz, the signal S is subsampled in a factor 2 at the input of the coder kernel 1. Likewise, the synthetic signal S′ is oversampled at 16 kHz at the output of the coder kernel 1.
The bit rate of the first stage 1 is 6.4 kbit/s (2×N1=2×192=384 bits per frame). If the coder has a maximum bit rate of 32 kbit/s (Nmax=1920 bits per frame), the maximum bit rate of the second stage is 25.6 kbit/s (1920−384=1536 bits per frame). The second stage 2 operates for example on elementary frames, or subframes, of 20 ms (320 samples at 16 kHz).
The second stage 2 comprises a time/frequency transformation module 5, for example of MDCT (“Modified Discrete Cosine Transform”) type to which the residual E obtained by the subtractor 3 is addressed. In practice, the manner of operation of the modules 3 and 5 represented in
The resulting spectrum is distributed into several bands of different widths by a module 6. By way of example, the bandwidth of the G.723.1 codec may be subdivided into 21 bands while the higher frequencies are distributed into 11 additional bands. In these 11 additional bands, the residual E is identical to the input signal S.
A module 7 performs the coding of the spectral envelope of the residual E. It begins by calculating the energy of the MDCT coefficients of each band of the difference spectrum. These energies are hereinbelow referred to as “scale factors”. The 32 scale factors constitute the spectral envelope of the difference signal. The module 7 then proceeds to their quantization in two parts. The first part corresponds to the telephone band (first 21 bands, from 0 to 3450 Hz), the second to the high bands (last 11 bands, from 3450 to 7225 Hz) . In each part, the first scale factor is quantized on an absolute basis, and the subsequent ones on a differential basis, by using a conventional Huffman coding with variable bit rate. These 32 scale factors are quantized on a variable number N2(i) of bits P2 for each subframe of rank i (i=1, 2, 3).
The quantized scale factors are denoted FQ in
A module 8 normalizes the MDCT coefficients distributed into bands by the module 6, by dividing them by the quantized scale factors FQ respectively determined for these bands. The spectra thus normalized are supplied to the quantization module 9 which uses a vector quantization scheme of known type. The quantization bits arising from the module 9 are denoted P3 in
An output multiplexer 10 gathers together the bits P1, P2 and P3 arising from the modules 1, 7 and 9 to form the binary output sequence Φ of the coder.
In accordance with the invention, the total number of bits N of the output sequence representing a current frame is not necessarily equal to Nmax. It may be less than the latter. However, the allocation of the quantization bits to the bands is performed on the basis of the number Nmax.
In the diagram of
The manner of operation of the latter module 11 is as follows. It firstly determines an approximate value of the original spectral envelope of the signal S on the basis of that of the difference signal, such as quantized by the module 7, and of that which it determines with the same resolution for the synthetic signal S′ resulting from the coder kernel. These last two envelopes are also determinable by a decoder which is provided only with the parameters of the aforesaid first subset. Thus the estimated spectral envelope of the signal S will also be available to the decoder. Thereafter, the module 11 calculates a spectral masking curve by applying, in a manner known per se, a model of band by band auditory perception to the original estimated spectral envelope. This curve 11 gives a masking level for each band considered.
The module 12 carries out a dynamic allocation of the Nmax−N0 remaining bits of the sequence Φ among the 3×32 bands of the three MDCT transformations of the difference signal. In the implementation of the invention set forth here, as a function of a criterion of psychoacoustic perceptual importance making reference to the level of the spectral envelope estimated with respect to the masking curve in each band, a bit rate proportional to this level is allocated to each band. Other ranking criteria would be useable.
Subsequent to this allocation of bits, the module 9 knows how many bits are to be considered for the quantization of each band in each subframe.
Nevertheless, if N<Nmax, these allocated bits will not necessarily all be used. An ordering of the bits representing the bands is performed by a module 13 as a function of a criterion of perceptual importance. The module 13 ranks the 3×32 bands in an order of decreasing importance which may be the decreasing order of the signal-to-mask ratios (ratio between the estimated spectral envelope and the masking curve in each band). This order is used for the construction of the binary sequence Φ in accordance with the invention.
As a function of the desired number N of bits in the sequence Φ for the coding of the current frame, the bands which are to be quantized by the module 9 are determined by selecting the bands ranked first by the module 13 and by keeping for each band selected a number of bits such as is determined by the module 12.
Then the MDCT coefficients of each band selected are quantized by the module 9, for example with the aid of a vector quantizer, in accordance with the allocated number of bits, so as to produce a total number of bits equal to N−N0.
The output multiplexer 10 builds the binary sequence Φ consisting of the first N bits of the following ordered sequence represented in
By placing first (a and b) the G.723.1 parameters and the scale factors of the high bands it is possible to retain the same bandwidth for the signal restorable by the decoder regardless of the actual bit rate beyond a minimum value corresponding to the reception of these groups a and b. This minimum value, sufficient for the Huffman coding of the 3×11=33 scale factors of the high bands in addition to the G.723.1 coding, is for example 8 kbit/s.
The method of coding hereinabove allows a decoding of the frame if the decoder receives N′ bits with N0≦N′≦N. This number N′ will generally be variable from one frame to another.
A decoder according to the invention, corresponding to this example, is illustrated by
A module 23 calculating the masking curve, identical to the module 11 of the coder of
According to the information supplied by the modules 24 and 25, the module 26 extracts the bits P3 of the input sequence Φ′ and synthesizes the normalized MDCT coefficients relating to the bands represented in the sequence Φ′. If appropriate (N′<Nmax), the standardized MDCT coefficients relating to the missing bands may furthermore be synthesized by interpolation or extrapolation as described hereinbelow (module 27). These missing bands may have been eliminated by the coder on account of a truncation to N<Nmax, or they may have been eliminated in the course of transmission (N′<N).
The standardized MDCT coefficients, synthesized by the module 26 and/or the module 27, are multiplied by their respective quantized scale factors (multiplier 28) before being presented to the module 29 which performs the frequency/time transformation which is the inverse of the MDCT transformation operated by the module 5 of the coder. The temporal correction signal which results therefrom is added to the synthetic signal S′ delivered by the decoder kernel 21 (adder 30) to produce the output audio signal Ŝ of the decoder.
It should be noted that the decoder will be able to synthesize a signal Ŝ even in cases where it does not receive the first N0 bits of the sequence.
It is sufficient for it to receive the 2×N1 bits corresponding to the part a of the listing hereinabove, the decoding then being in a “degraded” mode. Only this degraded mode does not use the MDCT synthesis to obtain the decoded signal. To ensure the switching with no break between this mode and the other modes, the decoder performs three MDCT analyses followed by three MDCT syntheses, allowing the updating of the memories of the MDCT transformation. The output signal contains a signal of telephone band quality. If the first 2×N1 bits are not even received, the decoder considers the corresponding frame as having been erased and can use a known algorithm for conceiving erased frames.
If the decoder receives the 2×N1 bits corresponding to part a plus bits of part b (high bands of the three spectral envelopes), it can begin to synthesize a wide band signal. It can in particular proceed as follows.
In the case where the decoder also receives part at least of the low spectral envelope of the difference signal (part c), it may or may not take this information into account to refine the spectral envelope in step 3.
If the decoder 10 receives enough bits P3 to decode at least the MDCT coefficients of the most important band, ranked first in the part d of the sequence, then the module 26 recovers certain of the normalized MDCT coefficients according to the allocation and ordering that are indicated by the modules 24 and 25. These MDCT coefficients therefore need not be interpolated as in step 5 hereinabove. For the other bands, the process of steps 1 to 6 is applicable by the module 27 in the same manner as previously, the knowledge of the MDCT coefficients received for certain bands allowing more reliable interpolation in step 5.
The bands not received may vary from one MDCT subframe to the next. The “known neighborhood” of a missing band may correspond to the same band in another subframe where it is not missing, and/or to one or more bands closest in the frequency domain in the course of the same subframe. It is also possible to regenerate an MDCT spectrum missing from a band for a subframe by calculating a weighted sum of contributions evaluated on the basis of several bands/subframes of the “known neighborhood”.
Insofar as the actual bit rate of N′ bits per frame places the last bit of a given frame arbitrarily, the last coded parameter transmitted may, according to case, be transmitted completely or partially. Two cases may then arise:
| Number | Date | Country | Kind |
|---|---|---|---|
| 03 00164 | Jan 2003 | FR | national |
| Filing Document | Filing Date | Country | Kind | 371c Date |
|---|---|---|---|---|
| PCT/FR03/03870 | 12/22/2003 | WO | 00 | 7/1/2005 |
| Publishing Document | Publishing Date | Country | Kind |
|---|---|---|---|
| WO2004/070706 | 8/19/2004 | WO | A |
| Number | Name | Date | Kind |
|---|---|---|---|
| 4949383 | Koh et al. | Aug 1990 | A |
| 6016111 | Park et al. | Jan 2000 | A |
| 6370507 | Grill et al. | Apr 2002 | B1 |
| 20040010407 | Kovesi et al. | Jan 2004 | A1 |
| 20050010395 | Chiu et al. | Jan 2005 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20060036435 A1 | Feb 2006 | US |
