speech addons: enable per-arch ggml CPU variants on desktop linux-arm64 (whispercpp, bci, tts-ggml)

[ogad-tether]

Extends the transcription-parakeet linux-arm64 CPU enablement (#3263) to bci-whispercpp, transcription-whispercpp and tts-ggml.

Problem

Desktop linux-arm64 prebuilds compiled ggml at the plain armv8-a baseline — the ARM dotprod/fp16/i8mm kernels were compiled out, leaving f16 and quantized models on the slowest generic paths (Android already builds GGML_CPU_ALL_VARIANTS).

Changes

The build-flags change lives in the ggml-speech vcpkg port (2026-07-14, qvac-registry-vcpkg#249) — per review guidance, a change for every consumer of the port belongs in the port, not in triplets. The port builds per-arch runtime-scored CPU MODULE variants (SIGILL-safe), statically links libc++ into the modules, stages them flat in lib/ (where the addons' existing staging glob already looks), and excludes the SVE tiers (2–4× slower than NEON on VL=128 cores — A/B linked in the port PR).

This PR:

1. Bumps the -cpp constraints — whisper-cpp override 1.9.1#1 → 1.9.1#3 (transcription-whispercpp, bci-whispercpp) and tts-cpp >= 2026-07-13#2 (tts-ggml) — whose ports now require ggml-speech >= 2026-07-14 (per review: no direct ggml-speech dependency in the packages).
2. transcription-whispercpp + bci-whispercpp: the backend-loading gate (ensureBackendsLoadedAndroid → ensureBackendsLoaded) widens from __ANDROID__ to __ANDROID__ || (__linux__ && __aarch64__) — with GGML_BACKEND_DL=ON no backend is statically registered, and the addon must call ggml_backend_load_all_from_path(backendsDir/BACKENDS_SUBDIR) before whisper_init. Gated to linux-arm64 only per review (linux-x64 stays on static backends). tts-ggml needs no addon change — it already threads backends_dir unconditionally through tts-cpp EngineOptions.
3. CHANGELOG entries under [Unreleased] for all three packages.

Verification (ubuntu-24.04-arm, mean RTF, lower is better)

End-to-end benchmark runs with this change set (port content via overlay):

• whisper (main baseline → branch): base f16 0.3316 → 0.0970 (3.4×), base q8_0 0.1270 → 0.0629 (2.0×), base q5_1 0.1774 → 0.1303, small f16 1.3450 → 0.3426 (3.9×), small q8_0 0.4561 → 0.2047 (2.2×)
• tts-ggml (main baseline → branch): chatterbox q4 2.70 → 2.28 (−15%), supertonic ~flat
• bci-whispercpp (branch): linux-arm64 leg green, CPU modules load and benchmark completes

All other platforms and Android unchanged/green in the same runs.

Merge order: qvac-registry-vcpkg#249 first (version resolution needs ggml-speech 2026-07-14 in the registry). Independent of #3263.

🤖 Generated with Claude Code

[github-actions]

Review Status

Current Status: ✅ APPROVED
Approvals so far: Team Lead: 1, Member: 1

[GustavoA1604]

Also put concise entry in CHANGELOG.md under [unreleased] for the packages

[ogad-tether]

Reworked per review guidance:

• Build flags moved out of the package triplets into the ggml-speech port (ggml-speech 2026-07-14: per-arch CPU variants on desktop aarch64 Linux qvac-registry-vcpkg#249) — the triplet approach was also setting the GGML flags for every vcpkg dependency in the manifests; the port is the right scope. The three packages now just consume ggml-speech 2026-07-14.
• @GustavoA1604 the whisper/bci backend-loading gate is now __ANDROID__ || (__linux__ && __aarch64__) — linux-x64 stays on static backends.
• CHANGELOG entries added under [Unreleased] for all three packages.

[ogad-tether]

Per review: consumption now flows through the -cpp ports — tetherto/qvac-registry-vcpkg#249 bumps parakeet-cpp/whisper-cpp/tts-cpp to require ggml-speech 2026-07-14, and this PR just bumps the whisper-cpp override (1.9.1#3) / tts-cpp constraint (2026-07-13#2). Direct ggml-speech deps removed. Also rebased onto main (CHANGELOG conflict with #3265 resolved).

[github-actions]

QVAC E2E — android — running

View run
Config: suite=smoke · filter=(none) · exclude=(none)