Vendor SAPI4 runtime/voices, Flextalk cabinet, NetTTS reference (standalone, no online deps)

third_party/README.md

@@ -1,5 +1,36 @@
 # third_party
 
-- `speech.h` — Microsoft Speech API 4.0 SDK header (`speech.h`), the same file
-  referenced from the SAPI4 resources in `docs/NOTES.md`. It is vendored here so the
-  host builds without any files outside the repository.
+Vendored dependencies. Everything the tool needs to run is kept here so the project
+does not depend on external downloads that can disappear. Only toolchain packages
+(Wine, Winelib, ffmpeg, python3, unshield, mingw) are installed by the system.
+
+## speech.h
+
+`speech.h` — Microsoft Speech API 4.0 SDK header. Used by `host/sapi4host.c`. Copied
+from the public SAPI4 header published in the community gist referenced by the asset
+repository (see `docs/NOTES.md`).
+
+## sapi4/
+
+The SAPI4 runtime and the MSTTS (Sam/Mary/Mike) voices, extracted from the original
+Microsoft installers. Private archival copies; do not redistribute.
+
+- `spchapi/SPEECH.DLL`, `spchapi/SPCHAPI.INF` — SAPI4 runtime and its registration INF.
+- `msttsl/msttssyn.dll`, `msttsl/*.vce`, `msttsl/*.cfg`, `msttsl/mstts.inf` — MSTTS
+  engine, voice data and registration.
+- `installers/spchapi.exe`, `installers/msttsl.exe` — original installers, kept for
+  completeness (the extracted files above are what `setup-prefix.sh` uses).
+
+## flextalk/
+
+- `data1.cab` — InstallShield cabinet with the AT&T Flextalk 2.1 engine and the
+  "Keith Bell" / "Kerry Watson" voices (the archive.org package).
+  `setup/install.sh` extracts it with `unshield -O` into `flextalk/extracted/`
+  (gitignored).
+
+## nettts-reference/
+
+- `vox_parser.cpp`, `vox_parser.hpp` — the VOX prosody encoder from
+  [h4rm0n1c/NetTTS](https://github.com/h4rm0n1c/NetTTS), kept as the reference for
+  validating `setup/vox_parser.py`. Upstream declares no license; these files are
+  reference-only and are not linked into anything we ship.