Speech

Technical appendix · Home · Français

The speech is concatenated recordings, not a synthesised signal: a single voice, cut into phoneme and transition fragments, is reassembled from phoneme strings. There is no text-to-phoneme code in the game: each spoken line was converted in advance and stored. tools/maup_speech.py ports the runtime, and adds a French front end, written for this project, so that any text can be spoken.

python3 -I tools/maup_speech.py "Je vous attendais." -o out.wav [--voice male]
python3 -I tools/maup_speech.py --phrase 1 -o out.wav
python3 -I tools/maup_speech.py --phonemes "ZetEEdAA" -o out.wav

Samples

Sound	Source
The game's line "J'étais dans ma chambre."	Phrase 1, stored in the game
The same line from the French rules	Text, through `text_to_phonemes`
A new sentence, faster voice	Text not in the game
The same sentence, slower voice	Text not in the game

How far each part is checked

Part	Status
Fragment script (`compile_words`)	Matches `maup.exe`'s code in an emulator on 449 of 450 phrases; the other is the quirk under "Driver"
Audio decoding (`render_script`)	Same samples as the game's Sound Blaster player in an emulator, on all 450 phrases
Sample rate	Read from the code (PIT divisor), not measured
Pause lengths	Estimated: the game uses CPU-dependent busy loops
French front end	Not from the game: written for this project. 95.2% of phonemes agree with the game's 450 phrases, and phrase lengths correlate at 0.997. Those phrases were also used to tune it, so new text will do worse
Real game output	Not compared with a recording of the game running

Run python3 -I tools/verify_speech.py to repeat the emulator checks (about 90 s, needs unicorn).

Data

Both records are in G4.CON.

Record 0x2C, phrases. 460 big-endian offsets relative to 0x398, then words. Phrase n is class 5 string n: the dialogue. 450 are used. Each word:

Bits	Meaning
15-12	duration, the only field the driver uses
11-6	two 3-bit fields the driver reads but never uses, possibly left over from another version
5-0	phoneme

Codes	Phonemes (tool's ASCII name)
`00`-`0D`	i, é `e`, è `E`, a, (`04` unused), open o `O`, o, ou `u`, u `y`, eu `2`, schwa `@`, in `I`, an `A`, on `U`
`16`-`25`	p b m f v t d n s z l, ch `S`, j `Z`, k g r
`2D`-`30`	yod `j`, `w`, u of "lui" `H`, gn `N`
`3C`, `3D`, `3E`	short pause, long pause, hiatus without a pause

Labels were read off the aligned text; the driver only knows the numbers. Codes 0E-15 have driver support but no phrase uses them.

Record 0x2B, samples. Big-endian word tables, then audio at 0x508:

Offset	Entries	Used by script op
`0x000`	22 vowel entries, 6 bytes each, with 4 boundaries sharing the next entry's first	0
`0x086`	consonant set 1	2
`0x15A`	consonant set 2	6
`0x22E`	transitions: `(vowel + consonant * 14)`, 2 bytes each	4

Driver

INT 0FEh service 0x1E = EA 0x166ED is called once, from MAUPITI1 EA 0x1191B, with the phrase, the character, and both records.

Compile, EA 0x167D8. A window over three phonemes (previous, current, next) produces a script of (op, a, b) triples: which fragment of which table. Phoneme class comes from the code and table DS:F72; DS:FA6 and DS:F8C pick variants by neighbouring vowel.
Read the script, EA 0x1605E, resolving each triple to a sample range.
Play. Timer 0 is reprogrammed to the sample rate, with one sample per interrupt. Three outputs: PC speaker PWM (EA 0x16245), Sound Blaster direct DAC, DSP command 0x10 (EA 0x163C6), and port 0x389 (EA 0x1654B).

Decoding. Each byte is two samples, high nibble first. A sample adds table DS:1212 to the previous one, modulo 256, starting from 0x80 in every fragment: -36 -28 -21 -15 -10 -6 -3 -1 0 1 3 6 10 15 21 28.

Voices. There is one set of recordings. The pitch comes only from the playback rate. The selector table at MAUPITI1 EA 0x11864 has nine entries, which appear to be the nine characters in their usual order (inferred):

Characters	Selector	PIT divisor	Rate
Maguy, Sue, Anita	above 0	`0x50`	14,915 Hz
Roy, Bruce, Anton, Chris, Bob, Juste	0 or below	`0x59`	13,407 Hz

With a Sound Blaster, both become 0x6E when the speed test near EA 0x1738C gives under 12 (inferred: a slow machine).

Quirk. The compile loop compares offsets as signed numbers, so phrase 368 ("Tu as besoin de moi, biquet?"), which crosses offset 0x8000, stops after its first phoneme in the emulated code. It presumably does in the game too. The tool plays the whole phrase. The port also keeps a fall-through at EA 0x16C74 that looks like a bug: it ignores the following vowel when choosing a consonant variant.

The French front end

text_to_phonemes applies an exception list and ordered letter rules, then liaison, schwa dropping, and the game's habits (je devoiced before a voiceless consonant, n + yod written as gn). phonemes_to_words adds durations: long vowel pairs before a pause, short vowels elsewhere, and the most common duration for each consonant.

Known limits: French spelling is irregular, so unusual words will be wrong; verb endings in -ent are only recognised after a subject pronoun; and there is no intonation, as in the game itself. For exact pronunciation, use --phonemes.