DV (raw .dv, DV in AVI)

How RecursiveAI.OneMedia.Dv, Codecs.Dv and the AVI reader handle DV. Documents: IEC 61834-2 and SMPTE 314M (DIF structure, packs) as publicly described, Microsoft's note on DV data in AVI files (type 1 and type 2), and ExifTool's DV tag documentation for the names of the pack fields. The field positions below were checked against files written by independent tools (stated where used).

Frame layout

A DV frame is a set of DIF sequences of 150 blocks of 80 bytes (12,000 bytes). Each block starts with a 3-byte identifier (section type in the top three bits, then sequence and channel numbers, then the block number).

A sequence is a header block, two subcode blocks, three VAUX blocks, then nine groups of one audio block and 15 video blocks. The stream parameters come from the first sequence:

• System (525 or 625 lines): the DSF bit, bit 7 of byte 3 of the header block.
• Class: the low five bits of the video source pack's last data byte (0x60): 0 is 25 Mbit/s, 4 is 50 Mbit/s, anything else the four-channel (100 Mbit/s) systems. Observed: DVCPRO HD 1080i/50 is 0x14; 1080i/60 and 720p are not verified.
• Picture: 720×480 (525) or 720×576 (625); DVCPRO50 the same; DVCPRO HD 1080i/50 is stored as 1440×1080. Frame rate 25, or 30000/1001. Display aspect (4:3 or 16:9) from the video source control pack (0x61); pixel aspect follows from it.
• Audio parameters from the audio source pack (0x50): sampling rate (bits 5–3 of the last data byte: 48, 44.1, 32 kHz) and quantisation (bits 2–0: 16-bit linear, 12-bit non-linear), channel count from the class and mode.
• Recording date and time from the VAUX packs 0x62/0x63 (BCD; the camera's local clock). Zone unknown. Dates before 1995 and unset clocks (all 0xFF) are not recording times. The time code (0x13 in the subcode blocks) is kept as a native tag.

Everything above is read from the first frame; the other frames are only framed.

Raw DV

.dv and .dif streams are a sequence of whole frames: the format sniffer accepts a header block followed by the subcode and VAUX blocks in their places (certain) or a header block alone (possible). The track is one dvvideo video track with the frame rate, size and aspect above; each packet is one frame (size = frame size), always a key frame, pts = dts = frame number at the frame rate. A file may start at any frame boundary but must start with a frame: bytes before the first header block are not skipped (a file that does not begin at a frame is not recognised). A partial last frame is delivered with the Corrupt flag and a dv.truncated-frame defect; a file shorter than one sequence raises dv.truncated-first-frame.

DV in AVI

• Type 1: one stream of type iavs whose chunks (00dc) are whole DV frames, picture and sound together. It is exposed as one dvvideo video track (size, rate and aspect from the first frame) with the stream's dvinfo (DVINFO) as configuration. The DV packs of the first frame are native tags (system dv).
• Type 2: an ordinary vids stream with fourcc dvsd/dvhd/dv25/… holding DV frames without audio, and the sound in a separate PCM auds stream. Both are ordinary tracks.

Audio in DV

The sound inside a DV frame is stored in the audio blocks shuffled across the frame's DIF sequences and, at 12 bits, non-linear. Returning it as PCM or as an audio packet needs a de-shuffling step that goes beyond framing and header parsing. The reader reports the audio parameters as native tags and does not expose an audio track for type 1 AVI and raw DV. ffprobe lists the synthesised audio streams, so each file with embedded audio shows one deviation class in the differential, dv-embedded-audio (avi.md). A de-shuffler is a follow-up.

Defects

Deviations from ffprobe 7.0.2

• dv-embedded-audio: ffprobe adds one pcm_s16le stream, with one packet per frame, for each audio channel pair it finds in the frame, so it lists 1 + N streams and N + 1 packets per frame where we list 1 and 1 (N is 1 or 2 on the corpus: ffprobe shows 2 or 3 streams). The harness accepts a stream-count or packet-count difference only when every extra stream is pcm_s16le and the first track is dvvideo; all other fields of the video track still have to match. De-shuffling the sound is not implemented and is a documented gap.
• truncated-last-frame: ffprobe lists an empty packet, we return the bytes present with Corrupt.
• dv-frame-start: a raw file that does not start at a frame boundary is opened by ffprobe (it searches for the next header); sniffing here looks at the head of the file only, so the file is not recognised.

Everything else (frame sizes, positions, payload hashes, times, stream parameters) agrees on the corpus.