MicroslopSam/docs/NOTES.md

# Notes

## Registry facts

From `spchapi/SPCHAPI.INF` (`%49000%` = `%25%\speech` = `C:\windows\speech`):

- `CLSID_TTSEnumerator` `{D67C0280-C743-11cd-80E5-00AA003E4B50}` -> `Speech.dll`
- "Audio Destination to File Object" `{D4623720-E4B9-11cf-8D56-00A0C9034A7E}` -> `Speech.dll`
- "Audio Destination Object" `{CB96B400-C743-11cd-80E5-00AA003E4B50}` -> `Speech.dll`
- Lexicon dialog `{92655FB1-ADF9-11d1-BEB9-006008317CE8}` -> `Speech.dll`
- Phoneme converter `{B9F11A96-90E3-11d0-8D77-00A0C9034A7E}` -> `Speech.dll`

From `msttsl/mstts.inf` (`[TTSReg]`):

- Engine `MSTTSSyn` = `{E0725551-286F-11d0-8E73-00A0C9083363}` -> `MSTTSSYN.dll`
- Engine install dir: `Common Files\microsoft shared\SpeechEngines\TTS`.
- Enumerates voices by scanning `*.vce` in its own directory.

## Flextalk registration (from the installer's `setup.ins`)

- Engine value: `HKLM\Software\Voice\TextToSpeech\Engine`
  `"AT&T FlexTalk" = "{FFE78E00-C69C-11CE-AEC1-0000C0E9CB87}"`
- Engine subkey (forces mode enumeration): `...\Engine\{FFE78E00-...}` with
  `Dirty = 1` (REG_DWORD). The runtime rewrites `Count` + `Mode0..N` and
  clears `Dirty` after enumerating.
- `HKCR\CLSID\{FFE78E00-C69C-11CE-AEC1-0000C0E9CB87}\InprocServer32` =
  `<root>\EngineDLLs\FlexTalk.dll`, `ThreadingModel = Apartment`.
  Also registered: `{778EAAE2-499F-11CF-B0BA-444553540000}` -> same DLL.
- Directories: `HKLM\Software\AT&T\Watson\2.0\FlexTalk\2.0\Directories`
  `RootDirectory = C:\FlexTalk\`.
- Dictionaries: `...\FlexTalk\2.0\Dictionaries`:
  `StandardDictionary="2 ft_sdict.dll"`, `ExceptionDictionary="0 ft_udict.dll"`,
  `HomographDictionary="1 ft_hgrap.dll"`, `ProperNameDictionary="3 ft_namsa.dll"`,
  `LetterToSoundDictionary="4 ft_l2s.dll"`.
- Options: `...\FlexTalk\2.0\Options` `Gender=Male`, `SampleRate=11025`.
- Speakers (per-voice settings; names come from these values). **All numeric
  values must be REG_DWORD**: with REG_SZ the engine reads garbage and
  synthesizes a monotone buzz (pitch fixed at max, pitch control dead).
  - `{11260610-C84A-11ce-AEC4-0000C0E9CB87}`: Mode Name / Speaker Name
    "Keith Bell", Pitch 27, Speaking Rate 33, Head Size 46, Expression 59,
    Lowering 40, Prominences 17, Age 30, Volume 100.
  - `{000031CB-3699-551F-AED0-6B4893382D57}`: "Kerry Watson", Pitch 65,
    Speaking Rate 32, Head Size 45, Expression 100, Lowering 40,
    Prominences 17, Age 30, Volume 100.
- Voice data: `EngineData\AE1M08LC.*`, `AE1M11LC.*` (male, 8/11 kHz) and
  `AE2F08LC.*`, `AE2F11LC.*` (female); `.INV` inventory, `.NDX` index,
  `.COD` codebook, plus shared `DICT_CI.*`, `phondef`, `AMPRULES`, `ENDINGS`.

Extraction: `unshield -O -d extracted x installer/flextalk/data1.cab`
(`-O` = old compression). The package is single-volume: the original 1997 upload log
(`WS_FTP.LOG`) lists `data1.cab` only, and `-O` extracts all 450 entries with exact
sizes. Without `-O`, unshield's new decompressor reports a phantom "volume 2" and fails
on `MSVCRT.DLL` / `WatsonUNInstall.dll`; with `-O` both are complete, valid PE files.

The `.ENV` files the engine probes (`ae1m11lc.ENV`) were never shipped; synthesis works
without them.

## Latency and output shape (measured 2026-09-16)

- Warm wrapper time: ~410-425 ms per short line (engine call 190-220 ms, ffmpeg measure
  ~110 ms, apply ~90 ms, profile write ~57 ms when it runs). First (cold) call ~730 ms.
- The Flextalk profile write is cached per voice in the Wine prefix
  (`~/.wine-microslopsam/.microslopsam-profile<GUID>.reg`): unchanged profiles skip
  `wine reg import`. Recreating the prefix invalidates the cache.
- Fades: the 8 ms fade-in/out is applied to the trimmed file whose 25-30 ms padding
  silence is retained, so it shapes silence, not speech. Measured head: first sample
  >30 at 22.6 ms, speech (>500) at 29.6 ms, natural ~15 ms attack; tail decays over
  ~18 ms. No click, no attenuation of the first word.
- Leading hum (default on): 80 ms of 75 Hz at -48 dBFS peak before the engine padding,
  to prime Discord's pipeline. Verified 75.0 Hz / -48.0 dBFS, speech onset ~110 ms,
  final -18.6 LUFS / -3.0 dBFS, apply pass +~10 ms. `VESKSLOP_TTS_HUM_MS=0` disables and
  reproduces the previous output byte-for-byte. Tuning: `_HUM_HZ` (30-200), `_HUM_DB`
  (-80..-6; raise toward -36 if a client still gates the start).
  The hum bypasses the high-pass/trim chain, so `VESKSLOP_TTS_HIGHPASS` (up to 400 Hz)
  cannot remove it (verified at 200 and 400 Hz).

## SAPI4 runtime engine discovery (`speech.dll`)

- `FUN_635747a4` reads `HKLM\Software\Voice\TextToSpeech\Engine` values
  (value name = engine name, data = CLSID string), validates the GUID with
  `FUN_635712e4`, then opens a *subkey named by the CLSID string*.
- `FUN_635758b1` (TTS mode-list ctor) reads `Dirty`; when non-zero/missing it
  calls `CoCreateInstance(CLSID, ..., IID @635791c0)` and pulls modes via
  vtable+0xC into 0xaf0-byte records, writes `Mode%d` (REG_BINARY), `Count`,
  then clears `Dirty`. When `Dirty=0` it reads the cached `Mode%d` blobs.
- Practical rule: to add an engine, register the `Engine` value + the CLSID
  `InprocServer32` + an engine subkey with `Dirty=1` (a missing subkey also
  triggers enumeration).

## Environment gotchas

- Wine needs a working X display (the runtime creates windows). `xvfb-run -a`
  proved flaky here; `setup/run.sh` starts a persistent `Xvfb :77` instead.
- Wine reports `err:ole:apartment_add_dll couldn't load in-process dll` when
  the `InprocServer32` path does not exist. Check the file copy first
  (this bit us: a bad `unshield` run created empty directories).
- Generate registry `.reg` files with `printf`, never `echo`: dash's `echo`
  interprets `\2` in `...\Watson\2.0\...` as 0x02, corrupting the key so that
  `wine reg import` is a silent no-op (it still returns 0). The wrapper rejects
  a generated profile containing control characters.
- 32-bit engines require a 32-bit prefix (`WINEARCH=win32`).
- `sudo` on this machine requires a password; installs are manual.

## Engine units (measured)

- speed: MSTTS range 30-450, Flextalk 50-390; the natural rate is about 150 for every
  voice (duration at 100/200/300 fits a natural value near 140-150).
- pitch: engine units, roughly linear in F0 near the natural point (~1.1 Hz per unit).
  Natural values: about 100 (MSTTS), 106 (Keith Bell), 186 (Kerry Watson); Flextalk max 250.
  Verified with the corrected DWORD registry: Kerry 0 st = 209 Hz, -6 st = 148 Hz,
  +6 st = 250 Hz (clamped at max); Keith 0 st = 119 Hz, +6 st = 170 Hz.
- The wrapper maps semitones as `natural * 2^(st/12)` and clamps to the range.
- volume: available through `ITTSAttributes` (host flag `-V`). Flextalk is linear
  (0-65535, default 65535; half is about -6 dB). MSTTS is non-linear (0 = silence,
  4096 about -24 dB, 65535+ full scale); treat it as on/off.
- Flextalk per-speaker profile values are applied at voice load; `Expression` was
  verified audible (0/59/100 produce different waveforms at equal duration). Changing
  them requires a registry edit and re-selecting the voice.

## VOX parser

- `setup/vox_parser.py` reimplements the NetTTS VOX prosody encoder (vendor tags
  `\!wH1/\!wH0`, `\!br`, `\!sfNN`); the host already passes `TTSDATAFLAG_TAGGED`, so
  the engine honors them.
- Validation: the reference `vox_parser.cpp` from a local clone of
  `h4rm0n1c/NetTTS` was compiled with g++ in a throwaway harness and diffed against
  the Python port; outputs are identical on the 31-line corpus. Re-check after any
  parser change while the clone is available.
- VOX mode is Flextalk-only. The wrapper enables it with `VESKSLOP_TTS_VOX=1`
  (`VESKSLOP_TTS_VOX_CLEAN=1` drops the stutter wrapper) and it takes precedence
  over `VESKSLOP_TTS_WORDS`.
- Line-leading `/rate N` and `/pitch N` controls are consumed by the wrapper
  (percent, 100 = normal; `<= 2.0` means multiplier). They override the env
  speed/pitch and are not spoken; a controls-only line yields 100 ms of silence.

## Decompiled corpora (Ghidra)

- `/mnt/ghidra/decomp/msttssyn.dll/` (786 functions, engine)
- `/mnt/ghidra/decomp/speech.dll/` (1030 functions, SAPI4 runtime)