WO2007107659A2 - Quantification vectorielle contrainte - Google Patents
Quantification vectorielle contrainte Download PDFInfo
- Publication number
- WO2007107659A2 WO2007107659A2 PCT/FR2007/050908 FR2007050908W WO2007107659A2 WO 2007107659 A2 WO2007107659 A2 WO 2007107659A2 FR 2007050908 W FR2007050908 W FR 2007050908W WO 2007107659 A2 WO2007107659 A2 WO 2007107659A2
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- vectors
- dictionary
- signal
- vector
- code vectors
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Classifications
-
- H—ELECTRICITY
- H03—ELECTRONIC CIRCUITRY
- H03M—CODING; DECODING; CODE CONVERSION IN GENERAL
- H03M7/00—Conversion of a code where information is represented by a given sequence or number of digits to a code where the same, similar or subset of information is represented by a different sequence or number of digits
- H03M7/30—Compression; Expansion; Suppression of unnecessary data, e.g. redundancy reduction
- H03M7/3082—Vector coding
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04N—PICTORIAL COMMUNICATION, e.g. TELEVISION
- H04N19/00—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals
- H04N19/90—Methods or arrangements for coding, decoding, compressing or decompressing digital video signals using coding techniques not provided for in groups H04N19/10-H04N19/85, e.g. fractals
- H04N19/94—Vector quantisation
Definitions
- the present invention relates to quantization dictionaries for coding and decoding signals such as audio, video signals, and more generally multimedia signals, for storage or transmission.
- the present invention proposes a solution to the problem of unstructured statistical vector quantization integrating a priori knowledge of a distortion threshold.
- QV vector quantization
- This finite set is called the reproduction alphabet or dictionary or directory, and its elements are code vectors, code words, exit points, or representatives.
- vector quantization a block of n samples of a signal is treated as a vector of dimension n.
- the vector is encoded by choosing, in a finite dictionary, the code vector that "most" resembles it. For example, an exhaustive search is made among all elements of the dictionary to select the dictionary element that minimizes a measure of distance between that element and the input vector.
- vector quantization dictionaries are designed from the analysis of samples of the signal to be coded forming a training sequence, using statistical methods such as generalized Lloyd-Max algorithms (GLA: Generalized Lloyd-Max Algorithm) or LBG (for Linde, Buzo and Gray).
- GLA Generalized Lloyd-Max Algorithm
- LBG for Linde, Buzo and Gray.
- each iteration includes a step of classifying the training vectors to build quantization regions of the training sequence according to the rule of closest neighbor, and a step of improving the dictionary by replacing the old code vectors by the centroids of the regions, according to the so-called centroid rule.
- dictionaries, or statistical vector quantizers thus obtained have no particular structure, which makes their exploration expensive in calculations and greedy in memory.
- the use, particularly in real time, of such dictionaries obtained by conventional statistical vector quantization is often limited to small coding and / or low bit rates of the order of 4 to 10 bits per vector.
- Algebraic vector quantization uses strongly structured dictionaries, derived from regular networks of points or error correcting codes.
- a simple example of a regular network is the cubic network, formed for example of a set of points with integer coordinates. Thanks to the algebraic properties of their dictionaries, the algebraic vector quantizers are simple to implement and do not have to be stored in memory.
- One of the main parameters of a regular network is its minimum quadratic distance between two points on which the code vectors are placed.
- the dictionary is generally chosen as the intersection of the network with a regular polytope, such as a hypercube or hypersphere, or with its surface. The points outside this intersection are then outliers. For a given bit rate, that is to say a fixed number of points of the intersection, limiting the number of outliers leads to increasing the minimum distance between two code vectors.
- algebraic vector quantizers are less complex to implement and require less memory, they are only optimal for encoding signals with uniform distribution. It should also be noted that the operations of indexing the code vectors, or numbering, and the reverse decoding operations require more calculations than in the case of statistical vector quantizers, for which these operations are performed by simple reads of tables.
- Another form of vector quantization combines algebraic quantization and statistical quantization.
- This quantization comprises two quantifiers.
- the first high-resolution, or fine-resolution, quantizer is generally well structured and uncomplicated to implement, while the second, coarse-resolution quantizer is an unstructured statistical vector quantizer.
- the first floor serves as a pre-calculation to speed up search in the second floor.
- the quantizer "Fine-coarse” suffers from the same disadvantages as those indicated above for the statistical vector quantization. Indeed, the link between the fine-resolution quantizer and the quantizer Coarse resolution is a simple look-up table and therefore, even if the high-resolution vector quantizer is a finite subset of a regular network with a minimal distance between code vectors, this constraint is not imposed on the vector quantizer with coarse resolution that is constructed by conventional techniques of statistical vector quantization.
- the invention has the particular advantage of defining a method for obtaining a dictionary by statistical vector quantization adapted to the signal to be encoded and taking into account the perceptual thresholds.
- the subject of the invention is a method for generating a vector quantization dictionary of a signal, of the type comprising a step of statistical analysis of training vectors representative of the signal determining a finite set of vectors of a signal. code representing said drive vectors, characterized in that the method further comprises a step of modifying said finite set of code vectors to impose a minimum distance between the modified code vectors two by two, these modified vectors forming said dictionary.
- the invention results in a dictionary adapted to the signal because of a step of statistical analysis of the drive vectors but with a minimum distance between the code vectors, which makes it possible to take into account the fact that vectors too close code are indistinguishable at the level of perception.
- the statistical analysis step comprises a sub-step of creating an initial dictionary, a substep of classification of the training vectors from the initial dictionary for forming quantization regions and a sub-step for determining a centroid for each region, said centroids being said code vectors.
- said substep of creating an initial dictionary is made from said training vectors and comprises replacing each training vector with a rounded vector chosen on a regular network of points and the selective elimination of vectors. rounded according to their frequency of appearance.
- the statistical analysis step further comprises a sub-step of calculating the distortion of the initial dictionary and a substep of comparing this distortion with a tolerance threshold, said statistical analysis step being repeated if said comparison is negative.
- the modifying step includes replacing each of said code vectors with a neighbor code vector selected on a regular network of points.
- the method further comprises a step of eliminating duplicates among said modified code vectors, which has the effect of reducing the size of the dictionary.
- said duplicate elimination step includes a sub-step of searching for duplicates by comparing the modified code vectors with each other and a sub-step of deleting duplicates.
- the method comprises a step of merging the modified code vectors with the drive vectors having at least said minimum distance between them and with said modified code vectors in order to improve the distribution of the code vectors of the dictionary.
- this melting step includes a substep of replacing each training vector with a rounded vector chosen over a regular network of points and a substep of combining modified code vectors and rounded training vectors.
- the rounded training vectors that are combined with the modified code vectors also have a minimum distance of two to two.
- said merging step further comprises scheduling the rounded training vectors according to their occurrence frequency and selecting a finite number of these vectors to combine them with the modified code vectors.
- said combining sub-step includes adding to said modified code vectors a finite number of rounded training vectors separate from said modified code vectors.
- the regular network of points has a step corresponding to a sensitivity level of a recipient receiver of said signal, which makes it possible to impose a minimum distance between two code vectors corresponding to a perceptual reality.
- the method further comprises an optimization phase of said dictionary.
- the optimization phase comprises a step of calculating the distortion of the dictionary and a step of comparing this distortion with a tolerance threshold, said steps of statistical analysis and modification being repeated if said comparison is negative.
- the optimization phase comprises at least one iteration of the method implemented from the dictionary formed of said modified code vectors.
- the subject of the invention is also a computer program comprising code instructions for implementing the method as mentioned above, as well as a device comprising means such as, for example, a working memory and a device. processor, to execute this computer program and thus generate a vector quantization dictionary within the meaning of the invention.
- the present invention also relates to encoding and decoding methods, and a decoder using a dictionary obtained according to the invention.
- FIG. 1 is a flowchart of the method of the invention
- FIGS. 2 and 3 are representations of a dictionary at different stages of the method of the invention.
- FIG. 4 is a flowchart of an optimization method
- FIGS. 5 and 6 are block diagrams of an encoder and a decoder implementing the dictionary obtained by the method of the invention.
- this method is implemented on training vectors representative of a digitized audiophonic signal such as a speech signal.
- This method firstly comprises a step 10 of statistical analysis of the training vectors.
- step 10 corresponds to the application of the generalized Lloyd-Max algorithm.
- Step 10 begins with a sub-step 12 of creating an initial dictionary marked CO.
- This dictionary is created from the analysis of the training vectors or by other means known per se such as a random selection of a finite number of the training vectors, or so-called “splitting" algorithms, "LBG algorithm”, or other.
- This initial dictionary CO is of a determined size.
- the method then comprises a substep 14 for initializing common variables and in particular for initializing an ITER variable corresponding to an iteration number and a DITER variable corresponding to a value of distortion.
- Step 10 then comprises a substep 16 of classification of the training vectors with respect to the initial dictionary CO to form classification regions and a sub-step 18 of determining a centroid for each region. More precisely, during the sub-step 16 of classification, for each training vector, the whole of the current dictionary, that is to say of the CO dictionary modified by successive iterations, is scanned, and one chooses the dictionary code vector minimizing quadratic error with drive vector. Sub-step 18 is performed as follows: for each class or region Vi, the new centroids are calculated.
- Q () is the quantization function
- yi is a code vector of the dictionary.
- a distortion value for the current dictionary is calculated.
- the drive vectors are quantized with the current dictionary provided by the preceding steps.
- the quadratic error between the drive vectors and their quantized version is the distortion measure.
- the substeps 16, 18 and 20 are then repeated iteratively until a maximum number of ITERMAX iterations is reached or the distortion is less than a tolerance threshold, the tests being carried out at a later time. sub-step 22.
- this step 10 in particular with the repetition of substeps 16, 18, 20 and 22, corresponds to the application of the generalized Lloyd-Max algorithm.
- the method thus delivers a CITER dictionary formed by an iterative statistical analysis of the training vectors.
- a two-dimensional representation of the CITER dictionary is given in Figure 2 in which the code vectors are represented by a scatter plot.
- the gray dots correspond to the training vectors or vectors of the signal and the black dots to the vectors of the dictionary. It can be seen that the distribution of the CITER dictionary vectors follows the density of the drive vectors and that many vectors are close to each other.
- the method then comprises a step of modifying the CITER dictionary to impose a minimum distance between the two-to-two code vectors.
- This modification step comprises the replacement of each of the code vectors by a neighboring vector chosen on a regular network of points such as a Cartesian grid. These modified vectors are also called rounded vectors.
- the neighboring vector chosen is the vector placed on the regular network closest to the code vector. Depending on the embodiments and types of vectors, criteria other than the distance between the vectors may be retained to determine the neighboring vector of a code vector.
- step 30 delivers a new dictionary C obtained by rounding the CITER dictionary obtained in step 10, on the regular network.
- the method then comprises a step 32 of eliminating duplicates among the modified code vectors forming the dictionary C.
- Step 32 comprises a substep of searching for duplicates performed on the set of dictionary C by comparing each of the vectors code to the other code vectors of the dictionary and then a sub-step of deleting duplicates.
- the dictionary C may comprise fewer code vectors than the CITER dictionary.
- the method then comprises a step 34 of melting the modified dictionary with the rounded training vectors in order to complete the dictionary if its size has been reduced.
- This step 34 begins with a substep 36 of replacing each training vector with a neighboring vector, called a rounded training vector, chosen on a regular network of points.
- a rounded training vector chosen on a regular network of points.
- the rounded training vectors are then ordered according to their frequency of occurrence in descending order, advantageously after eliminating the duplicates.
- the method then comprises a sub-step 38 of combining the dictionary C of modified code vectors with the rounded training vectors.
- the dictionary C is completed by the addition of the rounded training vectors not present in the dictionary, advantageously following the order of decreasing occurrence frequencies.
- the dictionary C is constructed in such a way that a minimum distance is guaranteed between the centroids forming the code vectors of the dictionary so that it is called constraint
- the parameters to be quantified are taken from a logarithmic scale, a constant step in this domain is equivalent to a logarithmic step in the field of energies, which corresponds to the sensitivity of the human ear.
- FIG. 3 shows the dictionary C "in the form of a two-dimensional point cloud. It will be noted in FIG. 3 that the dictionary C" obtained according to the method described above covers a large area of the drive vectors. and has a minimum distance between the two-to-two code vectors.
- the method further comprises a phase 40 for optimizing the dictionary C "whose flowchart is shown with reference to FIG. 4.
- This optimization phase consists, in the embodiment described, in taking over the same processing steps by applying them to the code vectors of the dictionary C.sub.H However, in the optimization phase described, at each iteration, the classification , the determination of the centroids, the modification of the code vectors and the elimination of the duplicates are carried out.
- the phase 40 begins with a classification step 42 applied to the code vectors of the dictionary C "followed by a step 44 of determining the centroids. These steps 42 and 44 are similar to the steps 16 and 18 described above.
- the optimization phase comprises a step 50 of measuring the distortion and a test 52 to determine if the distortion is less than a tolerance threshold or if a maximum number of iterations has been reached. Such an optimization phase makes it possible to further improve the quality of the dictionary.
- the initial dictionary is created by pruning rounded training vectors on the regular network. That is to say, the creation of this initial dictionary comprises the replacement of the drive vectors by neighboring vectors selected on the regular network and then the selective elimination, according to their frequency of appearance, of the vectors of rounded training.
- the optimization phase consists in repeating the method of FIG. 1 identically, using as initial dictionary the dictionary obtained previously.
- the melting step is implemented between the modified code vectors and the unrounded training vectors.
- the drive vectors used are chosen to ensure the maintenance of the minimum distance between the code vectors of the dictionary.
- the method of generating a dictionary as described above can be implemented through a program for any type of computer and computer.
- encoders and decoders are part of a global hierarchical subband audio coding and decoding system that operates at three possible bit rates: 8, 12 or 13.65 Kbit / s.
- the encoder always operates at the maximum rate of 13.65 kbit / s, while the decoder can receive the heart at 8 kbit / s, and one or two enhancement layers, respectively 12 or 13.65 kbit / s.
- the encoder 60 shown in FIG. 5 comprises two channels receiving the input signal.
- the first channel comprises a filter unit 62 able to extract a low band BF from 0 to 4000 Hz of the signal.
- the second channel comprises a similar unit 64 capable of extracting a high band HF spreading from 4000 to 8000Hz.
- These units 62 and 64 comprise, for example, filters and decimation modules made in a conventional manner.
- the low band signal is then coded according to a conventional CELP coding such as a CELP encode of 8 to 12 Kbps in a coding unit 66.
- a conventional CELP coding such as a CELP encode of 8 to 12 Kbps in a coding unit 66.
- the high band signal is coded in particular by using a vector quantization in a unit 70.
- the unit 70 is a parametric coding unit in which the coded parameters are time and frequency envelopes of the corresponding signal respectively. to the root mean square (rms) value per subframe or subband of the high band signal. These envelopes are passed in a logarithmic domain as follows: where ⁇ is the value of the envelope.
- parametric analysis modules 72 and 74 For each frame of the input signal of the time envelope and frequency envelope parameters are extracted in parametric analysis modules 72 and 74. These parameters in the logarithmic domain are then jointly quantized in a vector quantization module 76. using dictionaries generated according to the invention.
- the coding of the parameters is carried out by a vector quantization called Cartesian product with suppressed average.
- the mean ⁇ of the time envelope vector is calculated and then this average is quantized.
- the quantized mean is noted ⁇ q.
- the temporal envelope and frequency envelope vectors are then centered on ⁇ q before being separately quantized and multiplexed.
- the two channels HF and BF are then multiplexed in a multiplexer 80 to form the output signal of the encoder 60.
- the corresponding decoder 100 is described with reference to FIG.
- This decoder firstly comprises a demultiplexer 102 separating the corresponding channels at the low band and high band portions of the signal.
- the channel corresponding to the low-band signal is introduced into a CELP decoder 104.
- This CELP decoder also provides excitation parameters calculated in a conventional manner per se.
- the channel corresponding to the high band signal is introduced into a decoding unit 1 10 using in particular a reverse vector quantization 1 12.
- This module 1 12 is supplied to a temporal shaping module 114.
- This module 1 14 also receives a synthetic excitation signal generated by a module 116 from the CELP excitation parameters supplied by the decoder CELP 104
- the unit 1 10 finally comprises a frequency shaping module
- the decoding is carried out by denormalization of the temporal and frequency envelopes dequantized by the dequantized average.
- the two paths are processed in processing modules 120 and 122 before being recombined with each other in a mixer 124.
Landscapes
- Engineering & Computer Science (AREA)
- Multimedia (AREA)
- Signal Processing (AREA)
- Theoretical Computer Science (AREA)
- Compression, Expansion, Code Conversion, And Decoders (AREA)
Abstract
Description
Claims
Priority Applications (5)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| JP2009500891A JP4981122B2 (ja) | 2006-03-21 | 2007-03-09 | 抑制されたベクトル量子化 |
| KR1020087025757A KR101370018B1 (ko) | 2006-03-21 | 2007-03-09 | 제한된 벡터 양자화 |
| CN2007800099157A CN101467459B (zh) | 2006-03-21 | 2007-03-09 | 信号的矢量量化字典生成方法、编解码器及编解码方法 |
| EP07731724A EP2005756A2 (fr) | 2006-03-21 | 2007-03-09 | Quantification vectorielle contrainte |
| US12/225,312 US8285544B2 (en) | 2006-03-21 | 2007-03-09 | Restrained vector quantisation |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| FR0602452 | 2006-03-21 | ||
| FR0602452 | 2006-03-21 |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| WO2007107659A2 true WO2007107659A2 (fr) | 2007-09-27 |
| WO2007107659A3 WO2007107659A3 (fr) | 2008-12-18 |
Family
ID=36763212
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/FR2007/050908 Ceased WO2007107659A2 (fr) | 2006-03-21 | 2007-03-09 | Quantification vectorielle contrainte |
Country Status (6)
| Country | Link |
|---|---|
| US (1) | US8285544B2 (fr) |
| EP (1) | EP2005756A2 (fr) |
| JP (1) | JP4981122B2 (fr) |
| KR (1) | KR101370018B1 (fr) |
| CN (1) | CN101467459B (fr) |
| WO (1) | WO2007107659A2 (fr) |
Cited By (2)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2009111674A (ja) * | 2007-10-30 | 2009-05-21 | Nippon Telegr & Teleph Corp <Ntt> | ベクトル量子化方法,装置およびそれらのプログラムとそれを記録したコンピュータ読み取り可能な記録媒体 |
| CN113779103A (zh) * | 2021-03-02 | 2021-12-10 | 北京沃东天骏信息技术有限公司 | 用于检测异常数据的方法和装置 |
Families Citing this family (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| TWI376960B (en) * | 2009-07-31 | 2012-11-11 | Univ Nat Pingtung Sci & Tech | Codebook generating method for image compression |
| CN102307372B (zh) * | 2011-08-26 | 2014-12-17 | 电信科学技术研究院 | 一种基于Lloyd-Max量化器的数据压缩方法和设备 |
| ES2960582T3 (es) * | 2012-03-29 | 2024-03-05 | Ericsson Telefon Ab L M | Cuantificador vectorial |
| US11347941B2 (en) * | 2019-04-30 | 2022-05-31 | Marvell Asia Pte, Ltd. | Methods and apparatus for compressing data streams |
Family Cites Families (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US4817157A (en) * | 1988-01-07 | 1989-03-28 | Motorola, Inc. | Digital speech coder having improved vector excitation source |
| JP3264242B2 (ja) | 1997-02-28 | 2002-03-11 | 日本電気株式会社 | 認識辞書学習方法及びその装置並びにプログラムを記録した機械読み取り可能な記録媒体 |
| US6393394B1 (en) * | 1999-07-19 | 2002-05-21 | Qualcomm Incorporated | Method and apparatus for interleaving line spectral information quantization methods in a speech coder |
| JP3483513B2 (ja) * | 2000-03-02 | 2004-01-06 | 沖電気工業株式会社 | 音声録音再生装置 |
| JP3983532B2 (ja) * | 2001-12-05 | 2007-09-26 | 日本放送協会 | 場面抽出装置 |
| CN1906855B (zh) * | 2004-01-30 | 2014-04-02 | 法国电信 | 空间矢量和可变分辨率量化 |
-
2007
- 2007-03-09 CN CN2007800099157A patent/CN101467459B/zh not_active Expired - Fee Related
- 2007-03-09 WO PCT/FR2007/050908 patent/WO2007107659A2/fr not_active Ceased
- 2007-03-09 JP JP2009500891A patent/JP4981122B2/ja not_active Expired - Fee Related
- 2007-03-09 EP EP07731724A patent/EP2005756A2/fr not_active Withdrawn
- 2007-03-09 US US12/225,312 patent/US8285544B2/en not_active Expired - Fee Related
- 2007-03-09 KR KR1020087025757A patent/KR101370018B1/ko not_active Expired - Fee Related
Non-Patent Citations (3)
| Title |
|---|
| ANONYMOUS: "Just noticeable difference" WIKIPEDIA ARTICLE, [Online] 18 mars 2006 (2006-03-18), XP002394603 Extrait de l'Internet: URL:http://en.wikipedia.org/w/index.php?title=Just_noticeable_difference&oldid=44358293> [extrait le 2006-08-11] * |
| SCHEUNDERS P: "A genetic Lloyd-Max image quantization algorithm" PATTERN RECOGNITION LETTERS, NORTH-HOLLAND PUBL. AMSTERDAM, NL, vol. 17, no. 5, 1 mai 1996 (1996-05-01), pages 547-556, XP004017944 ISSN: 0167-8655 * |
| TED PAINTER, ANDREAS SPANIAS: "Perceptual Coding of Digital Audio"[Online] avril 2000 (2000-04), XP002394604 Extrait de l'Internet: URL:http://www.eas.asu.edu/~spanias/papers/paper-audio-tedspanias-00.pdf> [extrait le 2006-08-11] * |
Cited By (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2009111674A (ja) * | 2007-10-30 | 2009-05-21 | Nippon Telegr & Teleph Corp <Ntt> | ベクトル量子化方法,装置およびそれらのプログラムとそれを記録したコンピュータ読み取り可能な記録媒体 |
| CN113779103A (zh) * | 2021-03-02 | 2021-12-10 | 北京沃东天骏信息技术有限公司 | 用于检测异常数据的方法和装置 |
| CN113779103B (zh) * | 2021-03-02 | 2024-04-09 | 北京沃东天骏信息技术有限公司 | 用于检测异常数据的方法和装置 |
Also Published As
| Publication number | Publication date |
|---|---|
| KR101370018B1 (ko) | 2014-03-06 |
| US8285544B2 (en) | 2012-10-09 |
| KR20090005027A (ko) | 2009-01-12 |
| WO2007107659A3 (fr) | 2008-12-18 |
| JP4981122B2 (ja) | 2012-07-18 |
| US20100228808A1 (en) | 2010-09-09 |
| EP2005756A2 (fr) | 2008-12-24 |
| JP2009530940A (ja) | 2009-08-27 |
| CN101467459A (zh) | 2009-06-24 |
| CN101467459B (zh) | 2011-08-31 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| EP2374123B1 (fr) | Codage perfectionne de signaux audionumeriques multicanaux | |
| EP1709743A1 (fr) | Quantification vectorielle en dimension et resolution variables | |
| RU2326450C2 (ru) | Способ и устройство для векторного квантования с надежным предсказанием параметров линейного предсказания в кодировании речи с переменной битовой скоростью | |
| EP1692689B1 (fr) | Procede de codage multiple optimise | |
| WO2009027606A1 (fr) | Codage/decodage par plans de symboles, avec calcul dynamique de tables de probabilites | |
| US20070094035A1 (en) | Audio coding | |
| EP2727107B1 (fr) | Fenêtres de pondération en codage/décodage par transformée avec recouvrement, optimisées en retard | |
| WO2007107659A2 (fr) | Quantification vectorielle contrainte | |
| WO2009010674A1 (fr) | Codage hierarchique de signaux audionumeriques | |
| EP0801790A1 (fr) | Procede de codage de parole a analyse par synthese | |
| EP2289171B1 (fr) | Procède de traitement de donnees numeriques | |
| EP1037196B1 (fr) | Procédé de codage, de décodage et de transcodage audio | |
| EP1692687A1 (fr) | Transcodage entre indices de dictionnaires multi-impulsionnels utilises en codage en compression de signaux numeriques | |
| EP1197952B1 (fr) | Procédé de codage de la prosodie pour un codeur de parole à très bas débit | |
| EP2372699B1 (fr) | Codage d'échantillons audio ou vidéo utilisant des quantificateurs multiples | |
| EP1525663B1 (fr) | Compression de donnees numeriques robuste au bruit de transmission | |
| WO2024213553A1 (fr) | Quantification optimisée d'un espace latent en codage audio neuronal | |
| WO2008104725A1 (fr) | Procede de codage et decodage algebrique optimise, modules et programmes d'ordinateur associes | |
| EP4268024A1 (fr) | Procédé et dispositif de transmission de données représentatives d'un hologramme numérique, procédé et dispositif de construction d'un hologramme numérique et système de transmission et de construction d'un hologramme numérique | |
| EP1836699A1 (fr) | Procede et dispositif de codage optimise entre deux modeles de prediction a long terme | |
| EP0812070B1 (fr) | Procédé et dispositif de codage en compression d'un signal numérique | |
| EP3828886B1 (fr) | Procede et systeme pour separer dans un flux audio la composante voix et la composante bruit | |
| EP1756806A1 (fr) | Procede de quantification d'un codeur de parole a tres bas debit | |
| Deshpande et al. | Audio Spectral Enhancement: Leveraging Autoencoders for Low Latency Reconstruction of Long, Lossy Audio Sequences | |
| CN121725796A (zh) | 一种基于矢量量化变分自编码器的水声信号压缩方法 |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| WWE | Wipo information: entry into national phase |
Ref document number: 200780009915.7 Country of ref document: CN |
|
| REEP | Request for entry into the european phase |
Ref document number: 2007731724 Country of ref document: EP |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 2007731724 Country of ref document: EP |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 12225312 Country of ref document: US |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 2009500891 Country of ref document: JP Ref document number: 3832/KOLNP/2008 Country of ref document: IN |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| WWE | Wipo information: entry into national phase |
Ref document number: 1020087025757 Country of ref document: KR |
