feat: whisper-cpp 1.8.5 + ggml-speech 2026-05-27 + tts-cpp/parakeet-cpp re-validation (QVAC-19236, QVAC-18992, QVAC-19071)

[Zbig9000]

Summary

Four-port bump that pairs with the transcription-whispercpp@0.9.0 addon release on tetherto/qvac#2270.

Port	From	To	Reason
whisper-cpp	1.8.4.3#1	1.8.5#0	Upstream sync + migration to system ggml-speech
ggml-speech	2026-04-09#4	2026-05-27#0	Pick up the merge of qvac-ext-ggml PR #13 (v0.10.2 sync + Mac M2 PAD fix + Apple-Metal typedef fix + GustavoA1604 Android per-arch CPU fallback); pin spirv-headers version>=1.4.341.0 (matches qvac-fabric convention)
tts-cpp	2026-05-20#2	2026-05-20#3	Re-validate against new ggml-speech (on top of #173's BLAS-only-on-mac refactor)
parakeet-cpp	2026-05-26#0	2026-05-26#1	Re-validate against new ggml-speech (on top of #171 REF bump to 9cf4571f)

All four port bumps belong together: the new whisper-cpp@1.8.5 depends on the system ggml-speech port (no more bundled ggml), whisper-cpp@1.8.5's vendored sources are tightly coupled to the ggml-org v0.10.2 baseline shipped in the new ggml-speech@2026-05-27, and the two sibling speech ports (tts-cpp, parakeet-cpp) get the new baseline ggml automatically — so they need a port-version bump declaring they have been re-validated against it (keeps the registry self-consistent and lets downstream consumers express "I want a tts-cpp/parakeet-cpp known to work against ggml-speech v0.10.2").

This branch carries a merge of the latest main (which includes #170, #171, #172 and #173) so the tts-cpp and parakeet-cpp bumps cleanly stack on Freddy's parakeet REF bump and Gianni's BLAS refactors. The ggml-speech / whisper-cpp / parakeet-cpp git-trees are unchanged by the merge — only tts-cpp was rebumped (#2 → #3) to sit on top of #173.

QVAC-18992 — Migrate whisper-cpp to consume the ggml-speech system port

whisper-cpp now consumes the system-installed ggml provided by the ggml-speech port (same shape as parakeet-cpp and tts-cpp) instead of building the bundled qvac-ext-lib-whisper.cpp/ggml/ subdirectory. Driven by -DWHISPER_USE_SYSTEM_GGML=ON; whisper.cpp's top-level CMake then calls find_package(ggml CONFIG REQUIRED).

All backend wiring is delegated to ggml-speech:

• GGML_METAL / GGML_VULKAN / GGML_OPENCL / GGML_CUDA
• GGML_BACKEND_DL=ON on Android (dynamic-backend mode)
• GGML_CPU_ALL_VARIANTS=ON / GGML_CPU_REPACK=ON on Android (per-arch CPU MODULE .so files)
• GGML_VULKAN_DISABLE_COOPMAT[2]=ON on Android (Adreno NDK)
• spirv-headers -isystem shim for Vulkan
• Android NDK Vulkan-Headers download
• libqvac-speech-ggml-* output filename prefix

This removes ~95 lines from the whisper-cpp portfile and one whole companion file (android-vulkan-version.cmake). whisper-cpp + parakeet-cpp + tts-cpp now share the same backend binary set on every triplet (was: whisper brought a separate libqvac-ggml-* set, parakeet/tts brought libqvac-speech-ggml-*).

whisper-cpp's version>=ggml-speech constraint is bumped to 2026-05-27 so consumers with a stale baseline cannot combine whisper-cpp@1.8.5 with the old ggml-speech.

QVAC-19236 — Expose backend selection as features on whisper-cpp

Adds the metal feature (was unconditionally ON for macOS / OFF for iOS in the portfile) so all 3 GPU backends are now declarative:

whisper-cpp[metal]  -> ggml-speech[metal]   (osx | ios)
whisper-cpp[vulkan] -> ggml-speech[vulkan]  (linux | windows | android)
whisper-cpp[opencl] -> ggml-speech[opencl]  (android)

Each whisper-cpp feature simply propagates into the matching ggml-speech feature (with the platform guard kept). Consumers can express the full GPU matrix declaratively in their own vcpkg.json (the transcription-whispercpp addon does exactly this in the companion PR — selecting [metal] for osx only).

QVAC-19071 — ggml-speech bump to 2026-05-27#0 (qvac-ext-ggml PR #13 merged)

Picks up tetherto/qvac-ext-ggml@speech HEAD c9126afc — the merge commit of qvac-ext-ggml PR #13 ("QVAC-18992: merge ggml-org @ 19eac6f0 (v0.10.2) into speech"):

c9126afc  Merge pull request #13 from Zbig9000/QVAC-18992-merge-ggml-from-whisper-cpp
e31785e4  fix(metal): restore lost 'typedef struct {' before
          ggml_metal_kargs_supertonic_depthwise_1d (the ggml-org v0.10.2
          sync in 166c4e12 dropped the typedef header; caught by qvac CI
          Apple prebuild matrix via an overlay on tetherto/qvac#2270)
d39c0d29  metal: stride-aware src indexing in kernel_pad_f32 /
          kernel_pad_reflect_1d_f32 (fixes Mac M2 PAD test failure)
166c4e12  Merge ggml-org @ 19eac6f0 into speech (sync to v0.10.2)

PR #13 also brings in PR #11 (GustavoA1604 Android per-arch CPU dlopen fallback), already wired through port-version 4 via the GGML_CPU_ALL_VARIANTS=ON + GGML_CPU_REPACK=ON shape — no port-side change needed for that.

Build-side changes vs port-version 4:

• New patches/0001-ggml-vulkan-find-spirv-headers.patch wiring SPIRV-Headers into ggml-vulkan. The v0.10.2 sync added an unconditional #include <spirv/unified1/spirv.hpp> in src/ggml-vulkan/ggml-vulkan.cpp, but the upstream ggml-vulkan CMakeLists.txt never finds spirv-headers nor wires its include dir into the ggml-vulkan target. Patch does the find + link, gated on the vulkan feature.
• spirv-headers added as a dependency of the vulkan feature, pinned to version>=1.4.341.0 (matches the existing convention in ports/qvac-fabric/vcpkg.json so both speech-stack vulkan ports resolve from the same SPIRV-Headers version when they coexist in a consumer manifest).

tts-cpp + parakeet-cpp re-validation bumps

Pure metadata bumps — no portfile.cmake change, no REF change, no patch change. The ggml-speech baseline jump from 2026-04-09#4 to 2026-05-27#0 will automatically propagate into every consumer that has version>=ggml-speech (which both tts-cpp and parakeet-cpp do). Each port bump records that the combination "current tts-cpp/parakeet-cpp REF + new ggml-speech v0.10.2 baseline" has been validated.

Port	vcpkg.json change
tts-cpp	port-version: 2 → 3; version>=ggml-speech constraint bumped 2026-04-09#4 → 2026-05-27 (sits on top of #173's BLAS-only-on-mac refactor at 2026-05-20#2)
parakeet-cpp	port-version: 0 → 1; version>=ggml-speech constraint bumped 2026-04-09#3 → 2026-05-27 (sits on top of #171's REF bump to 9cf4571f, which introduced version-date 2026-05-26#0)

Local validation on linux-x64 (vulkan feature) against this registry tree (the merged state — i.e. with #171's new parakeet REF, #172's initial BLAS-off, and #173's BLAS-only-on-mac refactor on top):

• tts-ggml addon (consumes tts-cpp): bare-make generate resolves ggml-speech[core,vulkan]@2026-05-27 + tts-cpp[core,vulkan]@2026-05-20#3 + spirv-headers@1.4.341.0 from the registry tree, fetches qvac-ext-ggml@c9126afc, builds clean (BLAS-off path honored on linux; macOS would now keep BLAS on, per tts-cpp: enable BLAS only on macOS #173). npm run test:cpp — 35/35 C++ unit tests pass. npm run test:unit — 61/61 JS unit tests pass.
• transcription-parakeet addon (consumes parakeet-cpp): bare-make generate resolves ggml-speech[core,vulkan]@2026-05-27 + parakeet-cpp[core,vulkan]@2026-05-26#1, fetches tetherto/qvac-ext-lib-whisper.cpp@9cf4571f (the new parakeet REF from fix: update parakeet-cpp to port version 3 and adjust dependencies #171), builds clean against the new ggml-speech. npm run test:cpp — 32/32 C++ unit tests pass. npm run test:unit — 17/17 JS unit tests pass.

whisper-cpp REF

Bumped to v1.8.5 = 738d2e9e (tetherto/qvac-ext-lib-whisper.cpp@master), which is the merge of whisper.cpp #33 ("update to v1.8.5 and sync vendored whisper.cpp + ggml to ggml-org upstream"). Pure upstream sync — no tetherto-specific behavior change at the consumer level beyond the ggml-source switch above.

whisper-cpp GNUInstallDirs bug-fix patch

patches/0001-move-gnuinstalldirs-before-add-subdirectory-src.patch fixes an install-tree bug in upstream whisper.cpp (present since at least 1.8.x, also in ggml-org/whisper.cpp@master): src/CMakeLists.txt uses $<INSTALL_INTERFACE:${CMAKE_INSTALL_INCLUDEDIR}/whisper> to populate the imported target's include dirs, but include(GNUInstallDirs) runs after add_subdirectory(src). The variable is empty at the point the generator-expression is captured, so whisper-targets.cmake ends up with INTERFACE_INCLUDE_DIRECTORIES "/whisper" and downstream find_package(whisper) consumers fail with:

CMake Error in CMakeLists.txt:
Imported target "whisper::whisper" includes non-existent path "/whisper" in its INTERFACE_INCLUDE_DIRECTORIES.

This never bit anything before because WHISPER_USE_SYSTEM_GGML=OFF consumers never exercise the install-tree branch. Now that this port flips to system-ggml mode, the bug surfaces immediately. The patch moves include(GNUInstallDirs) above add_subdirectory(src). A follow-up will submit the same fix upstream to tetherto/qvac-ext-lib-whisper.cpp so we can drop the vcpkg-side patch on the next REF bump.

Result

File	Change
ports/whisper-cpp/portfile.cmake	150 → 65 lines (-85)
ports/whisper-cpp/vcpkg.json	+40 lines (3 features now, was 2; ggml-speech dep added; version>=ggml-speech: 2026-05-27)
ports/whisper-cpp/android-vulkan-version.cmake	deleted (-36)
ports/whisper-cpp/patches/0001-…patch	new (+50)
ports/ggml-speech/portfile.cmake	header/REF/SHA512 bump (-9, +7)
ports/ggml-speech/vcpkg.json	version + description update + spirv-headers vulkan dep pinned to version>=1.4.341.0 (+8, -2)
ports/ggml-speech/patches/0001-ggml-vulkan-find-spirv-headers.patch	new (+18)
ports/tts-cpp/vcpkg.json	port-version: 2 → 3 + version>=ggml-speech: 2026-05-27 (+2, -2)
ports/parakeet-cpp/vcpkg.json	port-version: 0 → 1 + version>=ggml-speech: 2026-05-27 (+2, -2)
versions/baseline.json	4-port baseline update
versions/w-/whisper-cpp.json	new 1.8.5#0 entry on top
versions/g-/ggml-speech.json	new 2026-05-27#0 entry on top
versions/t-/tts-cpp.json	new 2026-05-20#3 entry on top
versions/p-/parakeet-cpp.json	new 2026-05-26#1 entry on top

Test plan

Validated end-to-end on linux-x64 and the full qvac CI matrix via the companion qvac monorepo branch QVAC-19236-QVAC-18992-transcription-whispercpp-0.9.0:

• Overlay-port phase (proof both whisper-cpp + ggml-speech work BEFORE registry merge) — qvac PR #2270 with vcpkg-overlays/{whisper-cpp,ggml-speech}/ containing this PR's exact port contents (overlays kept in lock-step with the registry tree).
• Registry-resolved phase (proof the addon works WITHOUT overlays when ports come straight from this registry branch) — local file:// simulation:
  • bare-make generate on the transcription-whispercpp addon resolves whisper-cpp[core,vulkan]:x64-linux@1.8.5 + ggml-speech[core,vulkan]:x64-linux@2026-05-27 + spirv-headers@1.4.341.0 from the registry tree, downloads tetherto/qvac-ext-ggml@c9126afc (the new ggml-speech REF), builds clean
  • npm run test:cpp — 107/107 pass
  • npm run test:unit + test:integration + test:integration:accuracy + test:integration:chunking — all pass
• tts-cpp re-validation against new ggml-speech (on top of tts-cpp: enable BLAS only on macOS #173 BLAS-only-on-mac) — local file:// simulation on the tts-ggml addon: resolves tts-cpp[core,vulkan]@2026-05-20#3 + ggml-speech[core,vulkan]@2026-05-27, builds clean. test:cpp 35/35 pass, test:unit 61/61 pass.
• parakeet-cpp re-validation against new ggml-speech (on top of fix: update parakeet-cpp to port version 3 and adjust dependencies #171 REF bump) — local file:// simulation on the transcription-parakeet addon: resolves parakeet-cpp[core,vulkan]@2026-05-26#1 + ggml-speech[core,vulkan]@2026-05-27, fetches new REF 9cf4571f, builds clean. test:cpp 32/32 pass, test:unit 17/17 pass.

The qvac PR will get a follow-up commit removing both overlays + bumping its default-registry.baseline to this PR's merge SHA, then one more full CI run, then merge.

Made with Cursor