Speech files¶

The radio's lines, one file each in ms_speech\msspeech.hog (.HOG), kept without the .ut extension the missions' scripts name them by: ms1_ban_001.ut is the member MS1_BAN_001; Enriquez's words in the briefing, from the same archive: her last word before each mission, enrbr_tag01 to enrbr_tag28, all but 12, and her speech at the campaign's end, enddebriefing (Briefing); and the scenes, the .box files of the discs' archives, Enriquez's in the Reliant's rooms and the news reports (The Reliant's rooms). Each is a stream of the game's own speech codec, scrambled. engine/game/cbox.zig reads a file and engine/game/voice.zig decodes its stream.

sltool speech decode <file> <out.wav>       # one file to 22,050 Hz 16-bit mono WAV
sltool speech extract <archive> <out-dir>   # every line of an archive
sltool speech encode <in.wav> <out.ut>      # a WAV file to a line

sltool speech encode (speech_encode.zig) mixes a WAV file to mono at 22,050 Hz and writes it with the header every shipped file has: pulses at a step of two, the variable code below the first coefficient's level 24, a first gain of 64 and a ratio of 1.068. It takes each frame's reflection coefficients from the frame (Levinson-Durbin) at their nearest levels, and chooses each subframe's pitch, gain and pulses by analysis by synthesis against the decoder's own filter, ending each frame as the decoder ends it (voice.finish). It codes for accuracy rather than size, so its files are larger than the game's.

Layout¶

The game plays a file whatever its tag. The stream is XORed with the key AB 2D 9A AA, repeating from its first byte, up to 8 bytes before the file's end; the last 8 bytes are plain, in every shipped file. The game unscrambles a file in place as it loads it (speech_unscramble, 0x00462000), writes man over the length as its mark of having done so, and leaves a file so marked, or one beginning CB, as it is.

Fix: the game unscrambles length - 0x18 bytes, 8 short of what is scrambled, so it reads the stream's last 8 bytes scrambled, which end the last frame in noise; OpenReliant unscrambles them. Where the stream runs out before the samples do, as it does by a few bytes in the shipped files, the game reads whatever follows the file in memory; OpenReliant reads zeros.

The codec¶

The stream is read a bit at a time, least significant first (0x004A7360, 0x004A7850), and decodes to 22,050 Hz mono in frames of 432 samples, four subframes of 108 each, by linear prediction: twelve reflection coefficients, a pitch predictor over the past excitation, and a coded pulse excitation. voice.State holds a stream's state as the game does at +0x18 of its streams (speech_streams, 0x00539AD8, 0xD64 bytes each). Unverified: the source file's name; its code lies between timer.cpp's and winmain.cpp's, and OpenReliant calls it voice.cpp.

The header, once (0x004A7290):

Each frame (0x004A73B0):

The pulses come in the plain code, 0 for nothing, 01 for -2 and 11 for 2, or the variable code: two contexts of 256 lookups by the next eight bits into 29 symbols (tables.symbols, tables.records, each the next context, a length in bits and a value), a pulse of -6 to 6, a run of 7 to 70 zeros (six bits more), or an escape: 7, one more for each leading 1 bit, then a sign bit.

The subframe's excitation is the pulses times the gain plus the past excitation at the lag times the pitch gain. The coefficients step a quarter of the way to the frame's over its first four blocks of 12 samples, and the all-pole synthesis filter (0x004A78D0), its weights from the reflection coefficients (0x004A8050), makes the samples, which the player rounds to 16 bits, halves to even, and holds at full scale (speech_fetch, 0x00462290; speech_timer).

Playing¶

speech_start (0x00461EB0) plays a file through one sample of Miles's, 16-bit stereo at 22,050 Hz with both channels the same, at the speech volume times the master volume's share; a timer (speech_timer, 0x00462190) decodes 2048 frames at a time into the sample's double buffer. OpenReliant decodes a line whole and plays it on the sample (Radio, Sound).