Skip to content

BTAC1C

Jump to bottom Edit New page
Brendan G Bohannon edited this page Dec 18, 2016 · 1 revision

BTAC1C will be a hacked superset of IMA ADPCM.

If used, it will reuse the RIFF WAVE format:

  • If encoded without extensions, it may use the IMA ADPCM tag (wFormatTag=0x0011).
    • In this mode, only power-of-2 fudging is used.
    • The audio data will be approx 2% shorter if decoded as normal ADPCM.
  • With extensions, it will use a different tag (wFormatTag=0x7B1C).

Table of Contents

General Block Format

Simple Case:

  • Block coded as in IMA ADPCM (Mono)
    • init:WORD //Initial Predictor
    • step:BYTE //Initial Step
    • pfcn:BYTE //0 or Predictor Function
    • data:BYTE[] //4 bits/sample
  • Block coded as in IMA ADPCM (Stereo)
    • init_l:WORD //Left Initial Predictor
    • step_l:BYTE //Left Initial Step
    • pfcn_l:BYTE //0 or Predictor Function
    • init_r:WORD //Right Initial Predictor
    • step_r:BYTE //Right Initial Step
    • intr_r:BYTE //0 or Interpolation Data
    • data:BYTE[] //4 bits sample, interleaved left/right
  • Or, As extended Joint-Stereo (BTAC1C Stereo Only)
    • init_c:WORD //Center Initial Predictor
    • step_c:BYTE //Center Initial Step
    • pfcn_c:BYTE //0 or Center Predictor Function
    • init_s:WORD //Side Initial Predictor (or Offset)
    • step_s:BYTE //Side Initial Step (128..216)
    • pfcn_s:BYTE //0 or Side Predictor Function
    • data:BYTE[] //3 or 4 bits/sample
Step (Right/Side):
  • 0..88: Normal ADPCM Block (Split Stereo)
    • Samples encoded at 4 bit each.
    • Split stereo will have a separate left and right channel, with both in this range.
    • Split stereo will have 1/2 the effective sample rate as Mono or J-S encoding.
  • 89: Mono Channel (Stereo Ext)
    • Only left-channel predictor state is used.
    • Audio data encodes a mono channel, which is duplicated to both.
    • Right channel initial predictor interpreted as an L/R offset.
    • May encode an interleaved left/right channel offset towards center.
  • 90..127: Reserved
  • 128..216: Joint Stereo (Stereo Ext)
    • 16 bit sample blocks, 4x 3-bit center, 4 bit side.
      • 0..2=Center0, 3..5=Center1, 6..8=Center2, 9..11=Center3, 12..15=Side
    • Side channel will be stored at 1/4 the sample rate of the center channel.
    • The 3 bit samples will behave the same as a 4 bit sample with the LSB padded with a 0.
  • 217..255: Reserved
In mono samples, BTAC1C will use the same base format as IMA ADPCM.
  • Stereo blocks or stereo extensions may not be used in Mono mode.
BTAC1C will impose a constraint over normal ADPCM that both the storage block and logical sample block are to be power of 2 sizes.

The storage block will be a power-of-2 size, but will only encode 2^N-8 samples. Every 2^(N-3) samples, there will be a missing sample. During decoding, these missing samples are to be filled in with the average of the 2 adjacent samples (or a duplicate of the prior sample for the last sample).

The result will be that the recording will be slightly shorter if decoded as normal ADPCM than as BTAC1C.

Example, a block will be 512 bytes, and will encode 1016 samples. A filler sample will be inserted every 128 samples for this block size.

For split stereo blocks (in BTAC1C mode), similarly samples will be either doubled or averaged to make up for the 2x difference in effective sample rate.

In extended stereo mode, the block will encode 2^N-16 samples (rather than 2^N-8), as a result 16 logical pad samples are inserted per block.

For split stereo blocks (in ADPCM mode), then the stereo data is at the normal sample rate (and 8 logical samples are inserted).

BTAC1C WAVEFORMATEX

  • Constists of a normal WAVEFORMATEX
  • Followed by 2 magic bytes (0xF9, 0x03), inherited from MS IMA ADPCM.
    • Don't actually know their purpose.
  • Any additional byes will be interpreted as TLV data.
    • TWOCC('PF')
      • 4+ 9/18/27/36 bytes, interpreted as 1-4x 9-byte predictor functions.
      • Each consists of 8 weight bytes (signed), and a scale byte.
      • Internally, this is converted into a fixed-point FIR filter.
      • Weight bytes are nominally 4.4 fixed-point (-8.0 to 7.94).
      • Scale value is a 5.3 minifloat, which is multiplied with the weights.
  • TWOCC(T):
    • 8L-LL TT-TT
      • Max Length=4kB (4092 bytes of payload)
    • 9L-LL-LL-LL TT-TT
      • Max Length=256MB
  • FOURCC(T):
    • AL-LL TT-TT-TT-TT
      • Max Length=4kB (4092 bytes of payload)
    • BL-LL-LL-LL TT-TT-TT-TT
      • Max Length=256MB

Sample Data (Mono and Split Stereo)

Encoded with 4 bits per sample:

  • Bits 0..3=First Sample
  • Bits 4..7=Second Sample
For split stereo, audio is encoded in blocks of 32 bits:
  • 32 bits, left channel
  • 32 bits, right channel

Joint Stereo

Definitions:

  • Audio data consists of a series of 16-bit sample words
    • Each sample word is little endian.
    • 16 bit sample blocks, 4x 3-bit center, 4 bit side.
      • 0..2=Center0, 3..5=Center1, 6..8=Center2, 9..11=Center3, 12..15=Side
  • Center=(Left+Right)/2
  • Side=Left-Center
  • Otherwise it is functionally similar to other ADPCM blocks.
In joint-stereo mode, if both predictor bytes are 0, then the LSB of the center values is padded with 0. If either predictor byte is non-zero, then the LSB is padded with 1 if the other two low bits are 1.

Extended Predictors

The low 4 bits of the reserved/pfcn byte are used as a predictor:

  • 0=A (Last Sample)
  • 1=2*A-B
  • 2=(3*A-B)/2
  • 3=(5*A-B)/4
  • 4=(A+B)-(C+D)/2
  • 5=(3*(A+B)-(C+D))/4
  • 6=(5*(A+B)-(C+D))/8
  • 7=(18*A-4*B+3*C-2*D+1*E)/16
  • 8=(72*A-16*B+12*C-8*D+5*E-3*F+3*G-1*H)/64
  • 9=(76*A-17*B+10*C-7*D+5*E-4*F+4*G-3*H)/64
  • 10=(5*(A+B+C+D)-(E+F+G+H))/16
  • 11=(A+B+C+D+E+F+G+H)/8
  • 12-15=One of 4 filters given in the WAVEFORMATEX.
The high 4 bits are currently reserved.

For joint-stereo resv_c and resv_s will be used.

  • The joint-stereo case will thus have two different predictor functions.
For stereo, pfcn_l will be used.
  • The same predictor will apply to both channels.
For split-stereo and mono-stereo intr_r byte will give an interpolation hint:
  • 0=Average (average of samples)
  • 1=Even (encodes even sample values)
  • 2=Odd (encodes odd sample values)
  • 3=Even/Odd (encodes Even for Left, and Odd for Right)
  • For Split Stereo:
    • Average: Samples were averaged temporally.
    • Others: These values are stored in the respective channels.
  • For Mono Stereo:
    • Average: Samples are the average of left and right channel.
      • Offset is used to approximate L/R split.
    • Even: Even samples encoded in series (L,R).
    • Odd: Odd samples encoded in series (L,R).
    • Even/Odd: Even/Odd samples in series (L,R).
      • Offset was used to pull channels together.
While split-stereo and mono-stereo may encode similar information, split-stereo will favor disjoint stereo (at a lower sample rate), and mono-stereo will favor higher frequency (at the expense of L/R fidelity).

Full-Rate (Stereo)

The default format for stereo in BTAC1C will be Half-Rate Stereo, where the stereo is encoded at the same bitrate as mono (or 1/2 the bitrate of normal stereo ADPCM).

Alternatively, stereo may be encoded in Full-Rate mode, where it will use split-stereo blocks at the full sample rate.

The primary indicator of which mode is used will be the nAvgBytesPerSec field in the WAVEFORMATEX structure in relation to nSamplesPerSec: nAvgBytesPerSec==(nSamplesPerSec/2): Half-Rate nAvgBytesPerSec==nSamplesPerSec: Full-Rate

(Drop) Possible Ext: Sub-Range Interpolation (Mono and Mono as Stereo)

( Tested: Generally failed, average worse quality than baseline ADPCM )

Extension of BTAC1C Mono case.

  • Will be encoded using the step of 128..216.
Sample data is split into 32 bit sub-blocks.

Sub-Block Base:

  • Encoded as a Float8
  • S.E4.F3
  • Covers the signed 16-bit integer range.
Sub-Block Range:
  • E4.F2
  • Defines a range relative to the base (interpreted as a center point).
  • Defined as high-low, rather than as high-center.
Variant 1:
  • 8 bits: Sub-Block Base (relative to last sub-block)
  • 24 bits: Samples (8x 3-bit ADPCM)
  • First Sample is 0 | 4
    • First sample uses the starting value.
    • If first sample is 0:
      • ADPCM signal is decoded relative to the base value.
    • If first sample is 4:
      • ADPCM signal is relative to a linearly interpolated value.
      • The value is interpolated between the last and current base value.
      • Effectively, the base value gives the ending value.
Variant 2:
  • 8 bits: Sub-Block Base
  • 2 bits: Tag (1|2)
  • 6 bits: Sub-Block Range
  • 16 bits: 2 bits/sample.
    • Sample interpolates between the low and high value for the range.
    • Tag=1: Range is fixed.
    • Tag=2: Range interpolates between prior and next range.
Add a custom footer
Add a custom sidebar
Clone this wiki locally