Reference · Formats

Voice, streamed music and cutscene PCM

partialfrom#14#245#335

Three kinds of raw PCM, all signed 16-bit little-endian at 48 kHz, played by the IOP module SEWORDS.IRX straight into SPU2 streaming voices.

Voice lines: VOICE/ and VOICE_E/

file        headerless mono PCM; lines at 2048-byte boundaries, 0xff between
VOICE_DATA  8 bytes, in tables in main and gcmn
  +0x00  s32  ofs     byte offset of the line in its file; -1 = no line
  +0x04  s32  siz     bytes

VOICE/ is Japanese, VOICE_E/ English; the game uses VOICE_E when ccSaveData.voice (+0x842c) is set. evVoicePlay (0x0017eca0) sends the file name, ofs and siz with IOP command 0x80e0; wordPlay (SEWORDS.IRX:0x2c74) seeks to ofs and plays mono.

File names come from evVoiceFile (0x00307330) and evVoiceFileE (0x00307380). Five names there - EVVOL1P, EVVOL2-EVVOL4, BOSSTALK - have no file on this disc.

54 files, 8,208 lines: 4,117 Japanese (7,261.7 s), 4,091 English (6,840.7 s).

Event lines

Every message window (ccMessage::Change, ChangeInfo, Open) calls ccEvVoiceRequest(event, msg) (0x0017e810) as it opens, unless event is -1; the camera tutorial (teach_camera1-3) asks with event 3.

event           table (Japanese / English)                     voiceFile   file
0-49            evVoiceDataVol1M  / evVoiceDataVol1ME [event]      6       EVVOL1   / EVVOL1_E
50-99           evVoiceDataVol1S  / evVoiceDataVol1SE [event-50]   7       EVVOL1S  / EVVOL1SE
100-349         the volume 2-4 tables                            9-17      EVVOL2 ... (not on this disc)
below -1        ccVoiceRequest (0x0017eeb0), the field's voices
-1, 350 on      nothing

• Each table is 51 pointers: NULL (no voice for the event) or a VOICE_DATA array indexed by msg (evm1NNTbl, evs1NNTbl, and ...TblE). ofs -1 is no line. msg is not bounds-checked.
• saveData.voice (+0x842c) picks the English table at the request and the English file name (evVoiceFileE) when the line is sent.
• In Parody Mode (saveData.parodyFlag, +0x842b) events 0-49 have no voice; events 50-99 keep theirs.
• The row found is kept in one global, vdRequest (0x003789e0), with voiceFile (0x003789e4). ccMesVoicePlay (0x0017ec30) puts 0x80e0 in the first free of two slots (ccSnd +0x140); ccEvVoiceStop (0x0017ee40) puts 0x120. With both slots taken the call is lost.
• Once a frame the sound task's evVoicePlay (0x0017eca0) sends each taken slot in order: 0x80e0 as vBank {vdRequest->ofs, vdRequest->siz, 0, 0x6fff6fff, file name} through sewordCmd, 0x120 alone. Both slots send the same vdRequest, the last one asked for. SEWORDS (bgmFunc's 0x80e0, wordPlay 0x2c74) reads the offset (+0x00), the size (+0x04), the volume (+0x0c) and the name (+0x10). The word at +0x08 is never read.

Infection's four volume 1 tables hold 1,292 rows: 1,152 lines and 140 of -1, one row per message of their event (event 22 has one row more than messages). Every line is 2048-aligned and at least 40,756 bytes (0.42 s). Infection's volume 2-4 tables are empty externs or name files its disc does not have, so no event from 100 on is heard.

Streaming: SEWORDS.IRX channel 0

Voice lines and BGM.BIN share channel 0, core 0's sound-data input: the module refuses a new file (wordPlay, bgmPlay) while one is open, so a line asked for while the last still plays is not heard. At start the EE sends 0x130 (2, a DVD) and 0x140 (1, ring-buffer reads).

wordPlay(0, vBank)          SEWORDS 0x2c74
  refuse if gFd[0] >= 0; open, lseek(ofs)
  BgmInit(0, 0xc000)         buffer: SPU 2 x 16 KiB, then raw 2 x 8 KiB
  BgmSetMode(0, 0x10)        mono: gSPacketSize 16384, gRPacketSize 8192
  BgmPreLoad                 read 16 KiB into raw; convert both SPU halves;
                             if more than 16 KiB remain, read 16 KiB more
  BgmSetVolume(vol)          BVOLL = vol >> 16, BVOLR = vol & 0xffff
  BgmStart                   sceSdBlockTrans(0, loop, SPU buffer, 32 KiB)

SPU block                   1024 bytes: 256 left then 256 right samples;
                            mono copies each 512 raw bytes to both

on each transfer interrupt (half h played; _BgmPlay 0x1fb4):
  convert raw packet h into SPU half h
  end flagged:  let 2 interrupts pass, stop at the third:
                _BgmStop, BVOL 0, loopEnd (close, gFd[0] = -1)
  else if more than 8 KiB remain: read 8 KiB into raw h
  else:         read the rest up to the next 2 KiB sector, zero the
                packet's remainder, flag the end

A packet read at one interrupt starts to play at the third after it, and the third interrupt after the end is flagged stops the transfer - just as the last packet would start. A line plays 4096 * (ceil(siz / 8192) - 1) samples at 48 kHz; its last 1 to 4,096 samples are never heard (quiet in all but 16 of the 1,152 event lines: the median peak of the cut part is 98). A line of 16 to 32 KiB would play its first two packets twice (the preload leaves the raw buffer as it was); none is that short.

A line under 16 KiB. After the preload size - pos is negative, so at the first interrupt _BgmPlay takes the last-packet branch and asks read(fd, raw half, size - 16384), rounded to a sector: -16,384 for a line of size 0. By the code (not run on a console), that read never ends:

• IOMAN's read (0x053c) hands the size to the device unchanged.
• SEWORDS opens its files with mode 0x40000000 while ring-buffer reads are on (0x140, which the EE sends at start), and CDVDMAN's open then marks the file for stream reads (flag 8). The cdrom device's read (0x1414) sends those to a stream read of size >> 11 sectors (0x1378); a normal read would return -22 for a negative size (0xde0).
• The stream read (0x62d8) loops, blocking, until the bytes copied equal the size asked. Its copy (the dispatcher's mode 2, 0x65b0) compares sizes unsigned: -16,384 is 0xffffc000, so it copies every sector the stream has buffered, 2 KiB at a time, on past the 8 KiB raw half. The total can never equal -16,384, so it waits for more sectors and copies those too.

Channel 0 would stop for good, and the IOP memory after SEWORDS's buffer would be overwritten with the disc's following sectors.

Only one kind of row asks for such a line: every skill-word table (voiceKiteTbl ... voiceHerubaTbl, both languages, all four discs) ends with a row {ofs 0, siz 0}, and skillVoicePlay (0x0017e350) checks only ofs == -1 before sending a row. The row is reached when a skill's id less the character's base is that last index. For Kite (row 172 of 173) that is skill 295, Para Repth (bit 0 of its +0x2c clear: base 123), the skill of the item Recovery Drink (itemTblR 23, restores 800 HP). Kite using one in a field or dungeon would ask for it, through _ccSkillRequest (stype 2) and ccWordsPlay. Only Quarantine hands Recovery Drinks out: five in Kite's new-game items, and a trade. Quarantine's skillVoicePlay, ccWordsPlay (one more check, a byte at ccSnd +0x139) and SEWORDS.IRX do the same.

The port's seword::Word returns nothing for a read of 0 or fewer bytes, so such a line plays its preloaded 16 KiB and ends. evVoicePlay's 0x6fff is the only volume: no option scales it, only core 1's master volume (saveData.mainVol). Nothing lowers the music while a line plays, and the EE never asks whether one does (voiceStFlag is only ever cleared).

0x120 (ccEvVoiceStop) is BgmStop, BgmClose, BgmQuit and BVOL 0 on channel 0: it cuts a line, or a BGM.BIN track, at once. The game sends it when the confirm button is pressed in a message window (while the text types and to close it), when the tutorial's prompt closes, from the event op sound 7 (ccSndEvRequest), and from ccAllSoundOff at the start of every mode's set-up.

Streamed music: VOICE/BGM.BIN

Interleaved stereo (L, R, L, R ...). bgmWavTbl (INF gcmn.prg:0x006317d0) gives each track's byte range; wavPlay (0x0017e6e0) sends it with command 0x80f0.

The last 732 bytes of the file are 0xff. All other music is sequenced, in the sound banks.

Cutscene audio: CCSF Pcm chunks

In the STREAM/ CCSF files:

0x2200 Pcm (setup)     CCSTRM_PCM, 16 bytes, then dataNum * dataSize words
  +0x00  u32  myID
  +0x04  u8   type          0
  +0x05  u8   bitNum        16
  +0x06  u8   trackType     1
  +0x07  u8   lang          Outbreak on: one Pcm chunk per voice language,
                            the save's voice byte (1 English, 0 Japanese);
                            Infection's and Mutation's is 0, one chunk
  +0x08  u32  dataNum       blocks that follow
  +0x0c  u32  dataSize      words per block, 256

0x2201 F_Pcm (per frame)   CCSTRM_FSET_PCM, 12 bytes, then blocks
  +0x00  u32  myID
  +0x04  u16  dataNum       2 to 7 per frame
  +0x06  u16  padding       0x0012 in every chunk
  +0x08  u32  dataSize      256

block        1024 bytes: 256 left samples, then 256 right samples

ccPcmSound passes the blocks to the IOP untouched. Decode_Pcm (0x0014ddb0) reads the 16-byte CCSTRM_PCM onto its stack and uses only dataNum and dataSize; type, bitNum and trackType are read by no code. 132 cutscenes carry 3,000.2 s. A frame carries 1,600 samples (48,000 / 30), except str6100 at 800.

Notes

OPENING.PSS carries its audio in an MPEG private stream with an SShd header: format 1, 48,000 Hz, 2 channels, interleave 0x200.

crates/piney-data holds the tables (tables::voice, generated by piney-gen from each volume's code, with the file names as paths on the disc) and reads the lines (sound::voice); crates/piney-audio runs the requests (driver) and channel 0 (seword). Checked: the requests against the game's own ccEvVoiceRequest, ccEvVoiceStop and evVoicePlay run in tools/eemu.py for every row in both languages and modes (tools/sound_ee.py voice-fixture); the stream against SEWORDS.IRX itself run in tools/eemu.py over 115 lines, event 1's six among them (tools/iopemu.py voice-fixture); the decoded lines against tools/sound.py (tools/test_sound_rs.py).

Unknown

• The 48 kHz rate is the SPU2's fixed streaming rate plus measurements (pitch, spectrum, frame timing); no code states it.
• The transfer interrupt's timing is libsd's documented behaviour (at the end of each half, sceSdBlockTransStatus bit 24 naming the half transferring); the cut tail rests on it and was not heard on a console.
• How long the disc takes before a line's first sample.
• Whether a Recovery Drink used by Kite on Quarantine does stop channel 0 as the code says (above); not run on a console or an emulator.