DokuWiki - atrac3p

atrac_spectrum_anatomy

Anonymous (anonymous@undisclosed.example.com) — 2010-11-16T18:37:34+00:00

ATRAC3+ anatomy ATRAC3+ spectrum layout The spectrum (0..22.05kHz) for 44.1kHz sample rate is divided into 16 bands of equal spectral width of 1378.125Hz. Some of these 16 bands are further divided into quantization units (all quantization units of the same bands have the same spectral), as at low frequencies the same spectral width occupies a higher relative spectrum range (i.e. a higher pitch interval). All quantization units' data has to be combined into the total band data before the IMDCT…

band_envelope_processing

Anonymous (anonymous@undisclosed.example.com) — 2009-11-14T19:03:41+00:00

Band envelope processing The band envelope is given as a pair of X/Y values to represent a step function. The X values are quantized down to 32 instead 128 possible values. The Y values are exponents for the base 2, the range is -6 to 9 (so 1/64..512). The envelope will be expanded with geometric interpolation (resulting in 3 interpolated samples) on the step. The left half of the IMDCT result will be scaled by the previous envelope multiplied by the first envelope Y value of the right half.

band_frequency_domain_dithering

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:32:11+00:00

Frequency domain dithering Frequency domain dithering adds white noise to the spectrum. The noise intensity is determined per quantization unit. The first two quantization units get no noise at all, the noise amplitude is determined by * The quantization units' quantization precision

band_imdct

Anonymous (anonymous@undisclosed.example.com) — 2010-10-27T08:54:22+00:00

Band IMDCT The band IMDCT is a standard 128-point (input) IMDCT, yielding 256 output points. The window for the IMDCT can be steep or soft, which can be chosen for the right half of the current block and is reused for the left half of the next block. As with ATRAC3, the coefficients of odd bands are stored in reverse order because of the spectral mirroring of odd-numbered bands by the QMF synthesis filter. Please note that the block overlapping is not performed yet, but after

band_joint_stereo_post_processing

Anonymous (anonymous@undisclosed.example.com) — 2009-11-14T18:39:07+00:00

Band level joint-stereo post processing For each band, after calculating the coefficients, the following operations can be performed * The two channels can be swapped * The second channel can be negated

band_tone_adding

Anonymous (anonymous@undisclosed.example.com) — 2009-11-14T19:21:28+00:00

Band tone adding Each band can have an associated number of tones. Each tone has one envelope and a collection of sine waves. The envelope is mainly a rectangular window that gets ramped in an out over four samples (start and end position of the ramp is specified in 4-sample-units). The ramp is sin^2-like over four (essentially three, as the first ramp point is zero) samples according to ramp[x] = sin^2(x*pi/8). The zero of the ramp is at the start position of a start ramp or directly before t…

basic_level_prediction_vectors

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:26:32+00:00

basic prediction vectors for quantization unit levels There exist only three prediction vectors, because the vector number 3 is reserved as redirection to the extended level prediction vectors. * 0: 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 (just zeroes) *

canonical_huffman_trees

Anonymous (anonymous@undisclosed.example.com) — 2009-11-16T22:35:47+00:00

Canonical Huffman Trees in ATRAC3+ Canonical huffman trees are uniquely defined just by the number of bits for each symbol. The codes of one bit length are ordered alphabetically by their corresponding symbols. In ATRAC3+, they are allocated with the shortest symbols at low codes (the alphabetically first shortest symbol consists only of zeroes) and the longest symbols at high codes (the alphabetically last longest symbol consists only of ones).

envelope_x_trees

Anonymous (anonymous@undisclosed.example.com) — 2010-10-23T22:03:06+00:00

envelop point Y coordinate trees These trees are built as canonical huffman trees. Remember that negative values are 2's complement encoded and thus appear after the positive values in the trees. delta x for decreasing y tree * 0: not possible * 1: 2 bits * 2, 3: 3 bits

envelope_y_trees

Anonymous (anonymous@undisclosed.example.com) — 2010-10-21T22:02:48+00:00

envelop point Y coordinate trees These trees are built as canonical huffman trees. Remember that negative values are 2's complement encoded and thus appear after the positive values in the trees. envelope y tree * 0-3: 7 bit * 4: 5 bits * 5: 3 bits * 6: 4 bits *

extended_level_prediction_vectors

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:28:40+00:00

Extended prediction vectors for quantization unit levels The tables in this list consist of only 9 entries. Together with the starting value given separately, you yield ten numbers, used for a three quantization units each. The last triplet is extended to a quintet to fill all 32 quantization units. The table below gives the level differences for every triplet of quantization units but the first relative to the first triplet, with positive numbers meaning lower level of the later band.

final_band_combination_imdct_fir_filtering

Anonymous (anonymous@undisclosed.example.com) — 2009-11-14T23:03:54+00:00

Final band combination and FIR filtering Samples from the 16 bands are transformed via a standard type-IV-DCT, and the resulting 16 samples are grouped into 2 8-sample blocks. The FIR filter operates on the current and 11 most recent sets of results for the first block and on the 12 most recent sets of results for the second block.

level_delta_encoding_huffman_tree

Anonymous (anonymous@undisclosed.example.com) — 2009-11-21T20:56:00+00:00

Level difference encoding Huffman trees There are two sets of trees. The 6-bit-trees and the 4-bit-trees. The four-bit trees encode values betwenn -7 and 7 in 4-bit-twos-complement representation (so 0..7 and 9..15), while the 6 bit trees store values between 0 and 63 and are used modulo 64 in most contexts anyway, so signed/unsigned doesn't matter. These trees are built as

main_trees_list

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:41:38+00:00

mode_2_precision_prediction_vectors

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:20:03+00:00

Precision Prediction Vectors for Precision Encoding Mode 2 on Master Channel These are the 128 base tables available on mode 2 encoding. You may note that they are ordered by the first precision value and there are 16 tables for each initial precision value.

muting

Anonymous (anonymous@undisclosed.example.com) — 2009-11-14T18:41:25+00:00

Muting Optionally, after frequency-domain-processing, the coefficents can be zeroed out. Most probably for getting the decoder into a well-known state before unmuting.

point_count_trees

Anonymous (anonymous@undisclosed.example.com) — 2010-03-20T15:12:18+00:00

Point count trees These trees are built as canonical huffman trees. Remember that negative values are 2's complement encoded and thus appear after the positive values in the trees. point count encoding tree * 0: 1 bit * 1: 2 bits * 2: 3 bits * 3: 4 bits * 4: 5 bits

polyphase_coefficents

Anonymous (anonymous@undisclosed.example.com) — 2009-11-14T23:15:23+00:00

The coefficients for the polyphase filterbank are the following. They are arranged as a [2*12][2*8] array. The first index (row) is the tap number and incremented by 12 for processing the second 8-sample-block, the second index is the number of the sample in the IDCT16 result, and incremented by 8 if generating the second 8 output samples.

precision_delta_encoding_huffman_tree

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:55:06+00:00

Quantization precision difference encoding trees values encoded as three-bit signed values, so the negative values get codes after the positive values. The trees are built as canonical huffman trees Tree 0 * -1: 2 bit * 0: 1 bit * +1: 2 bit Tree 1 * -2: 3 bit

precision_prediction_vectors

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:15:59+00:00

Precision prediction vectors (dots are only for readability) * 0 -> 0 0 0 0 . 0 0 0 0 . 0 0 0 0 . 0 0 0 0 .. 0 0 0 0 . 0 0 0 0 . 0 0 0 0 . 0 0 0 0 * 1 -> 5 5 4 4 . 3 3 2 2 . 1 1 0 0 . 0 0 0 0 .. 0 0 0 0 . 0 0 0 0 . 0 0 0 0 . 0 0 0 0 * 2 -> 5 5 5 4 . 4 4 3 3 . 3 2 2 2 . 1 1 1 0 .. 0 0 0 0 . 0 0 0 0 . 0 0 0 0 . 0 0 0 0

quantization_precision_table

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:31:23+00:00

Quantization Precision Table * 0 - no data (1 level, just 0) * 1 - -1..1 (3 levels) * 2 - -2..2 (5 levels) * 3 - -3..3 (7 levels) * 4 - -5..5 (11 levels) * 5 - -7..7 (15 levels) * 6 - -15..15 (31 levels) * 7 - -31..31 (63 levels)

quantization_unit_dequantization

Anonymous (anonymous@undisclosed.example.com) — 2010-10-30T08:25:54+00:00

Dequantization The level of a bandlet chooses the available range. The maximum level, 63, will choose -65536..65536; lowering the level by one will decrease the range by a factor of 2^(1/3), so decreasing the level by 3 will shrink the range by a factor of 2. As an example, the level 15 will choose the range -1..1. The range is then divided into equally sized intervals according to the

quantization_unit_joint_stereo_coefficient_cloning

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:42:17+00:00

Per-quantization unit joint stereo coefficient cloning If a quantization unit only has coefficients in it's primary channel but not in the secondary one, the quantized coefficients along with the quantization precision of the primary channel are cloned to the secondary channel. The level information is not cloned.

quantization_unit_tree_id_trees

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T16:56:58+00:00

Trees for quantization unit tree ID storage restricted tree ID This tree is used to encode the restricted tree numbers as well as differences between them * 0: 1 Bit * 1: 2 Bit * 2,3: 3 Bit full tree ID This tree is used to encode the Tree ID for a quantization unit, but not for differences.

serialized_band_envelopes

Anonymous (anonymous@undisclosed.example.com) — 2010-10-21T21:52:41+00:00

Serialized Band Envelopes Band envelopes are optional, their presence is indicated by a flag bit. If that bit is clear, the channel has no envelope info, otherwise, the number of stored band envelopes is stored. These envelopes are assigned to the bands starting at band 0. Another optional feature is envelope replication. If it is enabled, the total number of enveloped bands (should be bigger than the number of envelopes) is stored separately and the last stored envelope is repeated for all ban…

serialized_band_joint_stereo_info

Anonymous (anonymous@undisclosed.example.com) — 2010-03-20T13:05:20+00:00

Serialized Band Joint Stereo Info The data for the bands (master/slave) assigned to left/right in an arbitrary way, and the coefficients can be negated afterwards. Encoding favours the same settings for all channels. Encoding * One bit: 0 = master is left for all bands, otherwise

serialized_envelope_point_counts

Anonymous (anonymous@undisclosed.example.com) — 2010-03-20T15:05:26+00:00

Serialized envelope point counts The point counts for the envelopes may be stored in four different modes, some of them different for the master and slave channel. Encoding modes 0: Direct encoding Plain encoding as 3-bit numbers. 1: Variable length encoding

serialized_envelope_x_coordinates

Anonymous (anonymous@undisclosed.example.com) — 2010-10-25T11:32:57+00:00

Serialized Envelope X Coordinates The envelope X coordinates are values between 0 and 31. They are in units of 4 samples in the subband time domain buffers, so they represent sample numbers between 0 and 124. As the points are stored in ascending X order, some encoding method exploit that the X coordinates are monotonically increasing.

serialized_envelope_y_coordinates

Anonymous (anonymous@undisclosed.example.com) — 2010-10-23T17:44:19+00:00

Serialized Envelope Y Coordinates The envelope Y coordinates are values between 0 and 15, with 0 -> 2-6, 6 -> 1 and 15 -> 29. The Y coordinates can be stored in different modes. A mode once chosen is used for all stored envelopes. Bands with 0 envelope points don't have any envelopes and thus no data stored.

serialized_quantization_unit_levels

Anonymous (anonymous@undisclosed.example.com) — 2010-10-25T13:01:20+00:00

Serialized quantization unit levels The quantization unit levels are logarithmically scaled scale factor for the band coefficients. An increase of three in the quantization unit level value gives rise to an factor of two in the scaling value. The range of the level values is between 1 and 63. There are four level encoding modes, and all but mode 0 are different for the master and the slave channel.

serialized_quantization_unit_quantization_precisions

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T18:08:37+00:00

Serialized quantization unit quantization precisions The quantization precision to load are numbers between 0 and 7, where 7 indicates highest precision and 1 means lowest precision and zero means no data at all (see quantization precision table). There are 4 different encoding modes to choose from for each channel. The modes 1 and 2 mean different things on the master and the slave channel. Many encoding modes work with the combination of a prediction vector (selected from fixed built-in pred…

serialized_quantization_unit_spectral_data_and_dither_noise_levels

Anonymous (anonymous@undisclosed.example.com) — 2010-07-23T17:37:13+00:00

Quantization unit spectral data and band dither noise level For each quantization unit with non-zero quantization precision, the quantized spectral data is stored using one out of many many many prefix code trees, see Main Trees List. There is no data stored if the precision is zero. The format of the spectral data is also described in the main tree list page.

serialized_quantization_unit_tree_choice_info

Anonymous (anonymous@undisclosed.example.com) — 2010-10-16T22:36:47+00:00

Serialized quantization unit tree IDs The frequency space data of the quantization units is generally Huffman encoded. There are two tree sets, which will be called tree set A and tree set B. Each of the tree sets consists of 8 trees for each quantization precision (except precision 0 which means that no data is present). Some of the trees are duplicate.

serialized_tone_data

Anonymous (anonymous@undisclosed.example.com) — 2010-10-25T22:28:48+00:00

Serialized tone data Tonal components are sine-shaped signals that get added after IMDCT on decoding. Header info Tonal components have their own stereo processing which is not connected to the residual spectrum stereo processing, this is the cause that the serialized tone data has a header that is not replicated per channel. The first bit in the tone info header defines the dynamic range. In the high-dynamic-range mode, the levels for each tone are chosen on a 64 step exponential amplitude s…

serialized_window_shape_info

Anonymous (anonymous@undisclosed.example.com) — 2010-07-21T21:27:45+00:00

Serialized window shape info For each band, the shape of the right half of the MDCT window is saved, with 0 = soft window, 1 = steep window. The left half of the window has to be the same as the right half of the previous frame. As the first frame will not produce any sensible data, the left half of the first frame does not need to be defined.

start

Anonymous (anonymous@undisclosed.example.com) — 2010-07-24T08:44:18+00:00

ATRAC3+ format documentation This is not the only ATRAC3+ documentation that floats around in the internet. A mostly independently derived specification can be found at . intro ATRAC3+ is a hybrid subband codec like MP3. The signal is split into 16 bands, each band is then divided into two parts: A superposition of sinusodial tones and the spectral residue that is encoded in the frequency domain. For the frequency-domain encoding, some of …

substream_data

Anonymous (anonymous@undisclosed.example.com) — 2010-08-01T10:05:27+00:00

Atrac3+ substream bit stream The bytes of the ATRAC3+ frame make up a bitstream. The most significant bit of the input bytes is consumed first. In multi-bit-fields, also the first of the bits that make up the field is the most significant one. The substream data first contains the number of

time_domain_dithering

Anonymous (anonymous@undisclosed.example.com) — 2009-11-14T19:26:22+00:00

Time domain dithering All samples in all subbands optionally have white noise of one common level added. The noise level is specified as a power of two.

tone_info_trees

Anonymous (anonymous@undisclosed.example.com) — 2010-10-26T16:45:13+00:00

Trees for decoding tone info Bands with tones tree * 1: 1 bit * 2: 3 bit * 3,4: 4 bit * 5..8: 5 bit * 9..16: 6 bit Tone per band tree * 0: 1 bit * 1: 2 bit * 2: 3 bit * 3: 4 bit * 4: 5 bit * 5: 6 bit * 6: 7 bit * 7: 7 bit