The RF64 format chunk

The fmt chunk says how the samples in the data chunk are encoded. RF64 keeps the WAV format chunk unchanged: it has grown over the years, each form adding fields at the end, and its first field decides how the rest of the file is read. It is small, so its size is always a real 32-bit size, and it must come before the data chunk. In an RF64 file it usually follows the ds64 chunk, at offset 48.

Fields

Offsets are from the start of the chunk’s data, after the eight-byte chunk header. All numbers are little-endian.

Offset	Size	Type	Name	Description
0	2	u16	`wFormatTag`	the format code: 1 for PCM, 3 for float, `0xFFFE` for extensible
2	2	u16	`nChannels`	the number of channels
4	4	u32	`nSamplesPerSec`	sample frames per second, such as 44,100 or 48,000
8	4	u32	`nAvgBytesPerSec`	bytes per second, for buffering: `nSamplesPerSec` × `nBlockAlign` for PCM
12	2	u16	`nBlockAlign`	bytes per frame for PCM (`nChannels` × bytes per sample), or the size of one block for ADPCM and other codecs
14	2	u16	`wBitsPerSample`	bits per sample: the container size for PCM (8, 16, 24, 32), 32 or 64 for float, 4 for both ADPCMs, 0 for some codecs such as GSM and MP3
16	2	u16	`cbSize`	the number of bytes of format-specific data that follow (`WAVEFORMATEX`)
18	`cbSize`	bytes	extension	depends on the format code: see extensible, Microsoft ADPCM, IMA ADPCM and MPEG

The explorer checks the block align and byte rate of PCM, float and G.711 files against the other fields, and reports a channel count or sample rate of 0.

Sizes

The chunk size tells which form a file uses.

Size	Structure	Used for
14	`WAVEFORMAT`	the 1991 form for formats other than PCM, without `wBitsPerSample`; rare
16	`PCMWAVEFORMAT`	plain PCM: the six fields up to `wBitsPerSample`
18 + `cbSize`	`WAVEFORMATEX`	adds `cbSize`, and codecs put their own data after it
40	`WAVEFORMATEXTENSIBLE`	`wFormatTag` is `0xFFFE` and `cbSize` is 22

Microsoft’s 1994 revision of the WAVE specification says every format other than PCM must use the extended chunk, with cbSize present even when it is 0, and must have a fact chunk. Readers should accept every size, read cbSize only when the chunk is long enough to hold it, and not trust a cbSize larger than the chunk; the explorer reports one.

Format codes

Microsoft registered hundreds of codes, listed in its mmreg.h header and, as of 1998, in RFC 2361. These are the ones commonly met in WAV files.

Code	Name	Audio
0x0000	`WAVE_FORMAT_UNKNOWN`	unknown
0x0001 (1)	`WAVE_FORMAT_PCM`	integer PCM
0x0002 (2)	`WAVE_FORMAT_ADPCM`	Microsoft ADPCM
0x0003 (3)	`WAVE_FORMAT_IEEE_FLOAT`	32- or 64-bit floating point, full scale 1.0
0x0006 (6)	`WAVE_FORMAT_ALAW`	G.711 A-law, 8-bit codes
0x0007 (7)	`WAVE_FORMAT_MULAW`	G.711 µ-law, 8-bit codes
0x0011 (17)	`WAVE_FORMAT_DVI_ADPCM`	IMA ADPCM (Intel DVI), 4 bits per sample
0x0031 (49)	`WAVE_FORMAT_GSM610`	GSM 6.10 speech
0x0050 (80)	`WAVE_FORMAT_MPEG`	MPEG-1 audio layer 1 or 2
0x0055 (85)	`WAVE_FORMAT_MPEGLAYER3`	MP3
0x0092	`WAVE_FORMAT_DOLBY_AC3_SPDIF`	AC-3 packed for S/PDIF
0x00FF	`WAVE_FORMAT_RAW_AAC1`	raw AAC
0x0160 to 0x0163	`WAVE_FORMAT_MSAUDIO1`, `WMAUDIO2`, `WMAUDIO3`, `WMAUDIO_LOSSLESS`	Windows Media Audio
0x1600	`WAVE_FORMAT_MPEG_ADTS_AAC`	AAC in ADTS frames
0x1610	`WAVE_FORMAT_MPEG_HEAAC`	HE-AAC
0x674F to 0x6751	`WAVE_FORMAT_OGG_VORBIS_MODE_1`, `_2`, `_3`	Ogg Vorbis, the whole Ogg stream in the data chunk
0x676F to 0x6771	`WAVE_FORMAT_OGG_VORBIS_MODE_1_PLUS`, `_2_PLUS`, `_3_PLUS`	Ogg Vorbis
0xF1AC	`WAVE_FORMAT_FLAC`	FLAC
0xFFFE (65534)	`WAVE_FORMAT_EXTENSIBLE`	see `SubFormat`
0xFFFF	`WAVE_FORMAT_DEVELOPMENT`	for codecs in development

The values of the Ogg Vorbis modes 1 and 3 plus are confirmed in Microsoft’s headers as published in the winapi bindings; the others follow in sequence.

The other registered codes

RFC 2361 lists these as well. Each name starts WAVE_FORMAT_. Few of them are met today.

Code	Name	Code	Name
0x0004	`VSELP` (Compaq)	0x0005	`IBM_CVSD`
0x0010	`OKI_ADPCM`	0x0012	`MEDIASPACE_ADPCM`
0x0013	`SIERRA_ADPCM`	0x0014	`G723_ADPCM` (Antex)
0x0015	`DIGISTD`	0x0016	`DIGIFIX`
0x0017	`DIALOGIC_OKI_ADPCM`	0x0018	`MEDIAVISION_ADPCM`
0x0019	`CU_CODEC` (HP)	0x0020	`YAMAHA_ADPCM`
0x0021	`SONARC`	0x0022	`DSPGROUP_TRUESPEECH`
0x0023	`ECHOSC1`	0x0024	`AUDIOFILE_AF36`
0x0025	`APTX`	0x0026	`AUDIOFILE_AF10`
0x0027	`PROSODY_1612`	0x0028	`LRC`
0x0030	`DOLBY_AC2`	0x0032	`MSNAUDIO`
0x0033	`ANTEX_ADPCME`	0x0034	`CONTROL_RES_VQLPC`
0x0035	`DIGIREAL`	0x0036	`DIGIADPCM`
0x0037	`CONTROL_RES_CR10`	0x0038	`NMS_VBXADPCM`
0x0039	`ROLAND_RDAC`	0x003A	`ECHOSC3`
0x003B	`ROCKWELL_ADPCM`	0x003C	`ROCKWELL_DIGITALK`
0x003D	`XEBEC`	0x0040	`G721_ADPCM`
0x0041	`G728_CELP`	0x0042	`MSG723`
0x0052	`RT24`	0x0053	`PAC`
0x0059	`LUCENT_G723`	0x0060	`CIRRUS`
0x0061	`ESPCM`	0x0062	`VOXWARE`
0x0063	`CANOPUS_ATRAC`	0x0064	`G726_ADPCM`
0x0065	`G722_ADPCM`	0x0066	`DSAT`
0x0067	`DSAT_DISPLAY`	0x0069 to 0x0081	mostly Voxware codecs
0x0080	`SOFTSOUND`	0x0082	`MSRT24`
0x0083	`G729A`	0x0084	`MVI_MV12`
0x0085	`DF_G726`	0x0086	`DF_GSM610`
0x0088	`ISIAUDIO`	0x0089	`ONLIVE`
0x0091	`SBC24`	0x0097	`ZYXEL_ADPCM`
0x0098	`PHILIPS_LPCBB`	0x0099	`PACKED`
0x0100	`RHETOREX_ADPCM`	0x0101	`IRAT`
0x0112	`VIVO_SIREN`	0x0123	`DIGITAL_G723`
0x0200	`CREATIVE_ADPCM`	0x0202	`CREATIVE_FASTSPEECH8`
0x0203	`CREATIVE_FASTSPEECH10`	0x0220	`QUARTERDECK`
0x0300	`FM_TOWNS_SND`	0x0400	`BTV_DIGITAL`
0x0680	`VME_VMPCM`	0x1000 to 0x1004	Olivetti codecs
0x1100	`LH_CODEC`	0x1400	`NORRIS`
0x1401	`ISIAUDIO` (AT&T)	0x1500	`SOUNDSPACE_MUSICOMPRESS`
0x2000	`DVM`

WAVE_FORMAT_EXTENSIBLE

With the format code 0xFFFE, cbSize is 22 and three fields follow. The real format is then the SubFormat GUID.

Offset	Size	Type	Name	Description
18	2	u16	`wValidBitsPerSample`	how many bits of each sample carry the signal, such as 20 in a 24-bit container; for a compressed subformat the same field is `wSamplesPerBlock`
20	4	u32	`dwChannelMask`	which speaker each channel feeds, see channel mask
24	16	GUID	`SubFormat`	the real format, see SubFormat

McGill’s notes give the rules: wBitsPerSample must be the container size, 8 × nBlockAlign / nChannels, and a multiple of 8; wValidBitsPerSample is for information, from 1 to the container size. The extensible form should be used for PCM above 16 bits, for more than two channels, for a container larger than the bit depth, and whenever the speakers need naming.

The format chunk of an RF64 test file from SciPy, one channel of 32-bit PCM at 44,100 Hz:

00000000  66 6D 74 20 28 00 00 00 FE FF 01 00 44 AC 00 00  fmt (.......D...
00000010  10 B1 02 00 04 00 20 00 16 00 20 00 00 00 00 00  ...... ... .....
00000020  01 00 00 00 00 00 10 00 80 00 00 AA 00 38 9B 71  .............8.q

A 40-byte extensible format chunk: PCM, 4 bytes a frame, 32 valid bits, channel mask 0.

The size is 40, the format code 0xFFFE, 1 channel, 44,100 Hz, 176,400 bytes a second, 4 bytes a frame and 32 bits; cbSize is 22, wValidBitsPerSample 32, the channel mask 0, so no speaker is named, and the subformat is PCM.

Channel mask

Each set bit of dwChannelMask names a speaker. The channels in each frame appear in the order of the set bits, lowest first. A mask of 0 means the channels have no particular speaker. The explorer reports a mask whose number of set bits differs from the channel count.

Bit	Speaker	Bit	Speaker
`0x1`	front left (L)	`0x200`	side left (Sl)
`0x2`	front right (R)	`0x400`	side right (Sr)
`0x4`	front centre (C)	`0x800`	top centre (Tc)
`0x8`	low frequency (LFE)	`0x1000`	top front left (Tfl)
`0x10`	back left (Ls)	`0x2000`	top front centre (Tfc)
`0x20`	back right (Rs)	`0x4000`	top front right (Tfr)
`0x40`	front left of centre (Lc)	`0x8000`	top back left (Trl)
`0x80`	front right of centre (Rc)	`0x10000`	top back centre (Trc)
`0x100`	back centre (Cs)	`0x20000`	top back right (Trr)

The short names are libsndfile’s. Common masks are 0x4 for mono, 0x3 for stereo, 0x3F for 5.1 (front left, front right, centre, LFE, back left, back right) and 0x63F for 7.1 with side speakers.

SubFormat

An extensible chunk carries its real format as a GUID. For the classic formats the first field is the old format code and the rest is fixed: 0000xxxx-0000-0010-8000-00AA00389B71. Microsoft names them KSDATAFORMAT_SUBTYPE_PCM (1), _IEEE_FLOAT (3), _ALAW (6), _MULAW (7) and _ADPCM (2).

A GUID is stored with its first three groups little-endian, so on disk PCM is 01 00 00 00 00 00 10 00 80 00 00 AA 00 38 9B 71 and float starts 03 00 instead. Ambisonic B-format, as libsndfile writes it, uses 00000001-0721-11D3-8644-C8C1CA000000 for PCM and 00000003-0721-11D3-8644-C8C1CA000000 for float.

Microsoft ADPCM

Format code 2. cbSize is 32 with the usual seven coefficient pairs.

Offset	Size	Type	Name	Description
18	2	u16	`wSamplesPerBlock`	frames each block decodes to: 2 × (`nBlockAlign` − 7 × `nChannels`) / `nChannels` + 2
20	2	u16	`wNumCoef`	the number of coefficient pairs, 7 as a rule
22	4 each	i16, i16	`aCoef`	the pairs (`iCoef1`, `iCoef2`), in 256ths

The standard pairs are (256, 0), (512, −256), (0, 0), (192, 64), (240, 0), (460, −208) and (392, −232). How a block uses them is on Sample data. The chunk is the same as in an ordinary WAV file; this one, for mono 8,000 Hz audio with 256-byte blocks, is from a WAV file in an FL Studio sample pack:

00000000  66 6D 74 20 32 00 00 00 02 00 01 00 40 1F 00 00  fmt 2.......@...
00000010  00 10 00 00 00 01 04 00 20 00 F4 01 07 00 00 01  ........ .......
00000020  00 00 00 02 00 FF 00 00 00 00 C0 00 40 00 F0 00  ............@...
00000030  00 00 CC 01 30 FF 88 01 18 FF                    ....0.....

A 50-byte Microsoft ADPCM format chunk: 500 frames a block, then 7 pairs.

IMA ADPCM

Format code 0x11. cbSize is 2, and the one field is wSamplesPerBlock at offset 18: the frames each block decodes to, which must be 2 × (nBlockAlign − 4 × nChannels) / nChannels + 1. libsndfile refuses a file where it is not. A 256-byte mono block gives 505 frames, a 512-byte mono block 1,017 and a 1,024-byte stereo block 1,017. wBitsPerSample is 4. The blocks are described on Sample data.

MPEG

For MP3, format code 0x55, Microsoft’s MPEGLAYER3WAVEFORMAT has cbSize 12 or more:

Offset	Size	Type	Name	Description
18	2	u16	`wID`	`MPEGLAYER3_ID_MPEG`, 1
20	4	u32	`fdwFlags`	padding: 0 as ISO decides, 1 always, 2 never
24	2	u16	`nBlockSize`	the block size in bytes: the frame length × `nFramesPerBlock`
26	2	u16	`nFramesPerBlock`	MP3 frames a block, usually 1
28	2	u16	`nCodecDelay`	the encoder delay in samples, or 0 when unknown

For MPEG-1 layers 1 and 2, format code 0x50, MPEG1WAVEFORMAT has cbSize 22; wBitsPerSample is 0:

Offset	Size	Type	Name	Description
18	2	u16	`fwHeadLayer`	the layer, as flags `ACM_MPEG_LAYER1`, `LAYER2` and `LAYER3` (1, 2 and 4), combined when a stream mixes layers
20	4	u32	`dwHeadBitrate`	the bit rate in bits a second, or 0 when it varies
24	2	u16	`fwHeadMode`	flags for stereo, joint stereo, dual channel and single channel
26	2	u16	`fwHeadModeExt`	the joint stereo mode extensions allowed, 0x01 to 0x08; 0x0F for all
28	2	u16	`wHeadEmphasis`	de-emphasis: 1 none, 2 50/15 ms, 3 reserved, 4 CCITT J.17
30	2	u16	`fwHeadFlags`	private, copyright, original, protection and MPEG-1 ID bits
32	4	u32	`dwPTSLow`	the low 32 bits of the first frame’s presentation time stamp
36	4	u32	`dwPTSHigh`	its top bit

A multichannel example

The format chunk of a 24-bit, 48 kHz, 5.1 surround file.

00000000  66 6D 74 20 28 00 00 00 FE FF 06 00 80 BB 00 00  fmt (...........
00000010  00 2F 0D 00 12 00 18 00 16 00 18 00 3F 00 00 00  ./..........?...
00000020  01 00 00 00 00 00 10 00 80 00 00 AA 00 38 9B 71  .............8.q

An extensible format chunk, with offsets from the start of the chunk.

Field	Bytes	Value
`ckSize`	`28 00 00 00`	40
`wFormatTag`	`FE FF`	`0xFFFE`, extensible
`nChannels`	`06 00`	6
`nSamplesPerSec`	`80 BB 00 00`	48,000
`nAvgBytesPerSec`	`00 2F 0D 00`	864,000
`nBlockAlign`	`12 00`	18 bytes per frame: 6 channels × 3 bytes
`wBitsPerSample`	`18 00`	24
`cbSize`	`16 00`	22
`wValidBitsPerSample`	`18 00`	24
`dwChannelMask`	`3F 00 00 00`	5.1
`SubFormat`	16 bytes	PCM

At 864,000 bytes a second, such a file passes 4 GB after about 83 minutes, which is where RF64 takes over.

.WAV File Format – Format chunk