[Zbig9000]

Why

All four speech addons override ggml-speech at 2026-10-07#1. In that version the default Adreno q8_0 prefill GEMM keeps its sums in fp16, so a dot product past 65504 comes back as inf. TranscribePsy's first feed-forward block reaches 66k on some 60 s windows; the inf becomes NaN and attention spreads it over the window. On a Galaxy S25 GPU, long recordings lost most of their text after the first window: 54 % WER on a 374 s English recording where the CPU is at 0 %.

The fix is tetherto/qvac-ext-ggml#129, published by tetherto/qvac-registry-vcpkg#446 as ggml-speech 2026-10-23. With it, the S25's OpenCL transcripts equal its CPU transcripts on English and French recordings, and differ by two words at one language switch on the mixed ones, at 1.5 % more encoder time. The measurements are in tetherto/qvac-fabric-speech.cpp#365.

speech-cpp 2026-10-23#2 (tetherto/qvac-registry-vcpkg#447) requires ggml-speech >= 2026-10-23, the same version, so the addons move to it in the same step.

Changes

• asr-ggml, audiogen-ggml, bci-whispercpp, tts-ggml:
  • ggml-speech override 2026-10-07#1 → 2026-10-23, and qvac["ggml-speech"] to match, in lockstep.
  • speech-cpp override 2026-10-20#1 → 2026-10-23#2, in lockstep.
• Other ggml-speech versions the jump includes:
  • 2026-10-19: the Adreno KQ/KQV image kernels no longer miscompute attention over 32 or more query tokens with unpadded lengths.
  • 2026-10-20: the WebGPU backend, an opt-in feature these addons do not enable.
  • 2026-10-07#2 only kept Hexagon out of the build.
• What speech-cpp 2026-10-21 to 2026-10-23#2 adds to the prebuilds (qvac-fabric-speech.cpp 303198cf..faed3c81):
  • asr-ggml: Core ML diarizer sidecars for Nemotron 3 Diarization and Sortformer v2.1 on macOS/iOS (speech QVAC-13245 fix[mod]: correct model licenses and remove deprecated piper TTS entries #353), Mosaic and Mosaic Compact loading (chore[notask]: backmerge release-qvac-registry-schema-0.1.1 into main #337), and Nemotron 3 Diarization loading again (Improve integration and mobile tests #352).
  • tts-ggml: the Parler-TTS Core ML DAC sidecar on macOS/iOS (qvac-lib-registry-server/data/models.prod.json: added "link" attribute to all vision-related models #348), and an error instead of a crash on non-finite Parler logits (fix(nmtcpp): add missing native dependency attributions to NOTICE #356).
  • asr-ggml, bci-whispercpp: whisper's fit sizes integrated and unified-memory GPUs that report no capacity against host RAM, and returns memory-capacity-unavailable when none is known (fix: add default registry core key and fix CLI install docs #360).
  • audiogen-ggml: nothing.
• Not built into the prebuilds:
  • The Hexagon work (speech chore[notask]: Make CLI self-contained with own bare-pack dependency #355, fix[notask]: add missing sdk rewrite for withDeviceInfo in publish workflow #359, fix: add default registry core key and fix CLI install docs #360, QVAC-12216 feat[api]: add parakeet transcription plugin #363): the port refuses a Hexagon ggml-speech.
  • WebGPU (fix(nmtcpp): add missing native dependency attributions to NOTICE #356): an opt-in feature.
  • The MiniMax Core ML conditioner and replay (fix(ocr-onnx): populate NOTICE with native dependency attributions #357, QVAC-12216 feat[api]: add parakeet transcription plugin #364): audiogen-ggml does not enable coreml.
  • The opt-in Parler Core ML decoder, the per-instance Parler Core ML policy (chore[skiplog]: Release @qvac/registry-client v0.1.5 #358) and Audio8 incremental PCM streaming (Deprecate Qwen2.5-Omni models and remove qwen-research license #361): new engine API that tts-ggml does not call yet.
• CHANGELOG: Fixed, Added and Changed entries under [Unreleased] in all four. No version bump; that goes in its own PR.

Verification

• node .github/scripts/check-speech-ggml-lockstep.mjs: passes. All four pin ggml-speech 2026-10-23 and speech-cpp 2026-10-23#2, and every prebuild triplet plans the same ggml-speech:
  • arm64-android: [core,opencl,vulkan]
  • Apple: [core,metal]
  • Linux and Windows: [core,vulkan]
• node --test .github/scripts/test/speech-ggml.test.mjs .../check-speech-ggml-builds.test.mjs: 26/26.
• build-backends tests: asr-ggml 12/12, audiogen-ggml 5/5, tts-ggml 2/2. The bci-whispercpp one only asserts the manifest shape (one override, no floors).
• Real vcpkg install --triplet arm64-android of asr-ggml's manifest (NDK 29, binary cache off) built speech-cpp[whisper,parakeet,vulkan,opencl] against ggml-speech[core,opencl,vulkan] 2026-10-23. That build was at speech-cpp 2026-10-20#1; for 2026-10-23#2 this PR's prebuild CI is the build check.

Interplay with open PRs

• QVAC-27211 chore: move the speech addons to speech-cpp 2026-10-21#1 (WebGPU engine support) #5299 (speech-cpp 2026-10-21#1) is now included here, its Parler entry too, and can be closed.
• QVAC-26890 feat[asr-ggml]: report Core ML diarizer memory from assessFit #5250 (speech-cpp 2026-10-21) conflicts on the speech-cpp lines and the asr-ggml changelog. After this PR lands, it keeps only its asr-ggml JS wiring (assessFit reports the diarizer bundles as coremlBytes) and drops its pin move and the entries this PR now carries.
• QVAC-27419 chore: release asr-ggml 0.11.1, audiogen-ggml 0.9.0, bci-whispercpp 0.14.0 and tts-ggml 0.14.0 #5327 (audiogen-ggml 0.9.0, bci-whispercpp 0.14.0, tts-ggml 0.14.0) conflicts in those three changelogs. If QVAC-27419 chore: release asr-ggml 0.11.1, audiogen-ggml 0.9.0, bci-whispercpp 0.14.0 and tts-ggml 0.14.0 #5327 lands first, this PR's entries move under a new ## [Unreleased]. If this PR lands first, QVAC-27419 chore: release asr-ggml 0.11.1, audiogen-ggml 0.9.0, bci-whispercpp 0.14.0 and tts-ggml 0.14.0 #5327 picks them up and releases them.

🤖 Generated with Claude Code

[github-actions]

Review Status

Current Status: ❌ PENDING
Approvals so far: none

Pending reviews: Needs 1 Management or Team Lead, and 1 more from Management, Team Lead, or Member.

[github-actions]

License compliance — clean

No new dependency license findings in this PR.

Warn-only (shadow) mode — this check does not block merges yet.

Updated automatically by the canonical license compliance workflow.

NOTICE presence (advisory)

Missing NOTICE (advisory, does not block):

• ./.github/actions/release-merge-guard
• ./packages/embed-llamacpp/benchmarks/server
• ./packages/embed-llamacpp/benchmarks/performance
• ./packages/fabric/test/integration
• ./packages/asr-ggml/benchmarks/server
• ./packages/inference-addon-cpp/mobile
• ./packages/llm-llamacpp/benchmarks/server
• ./packages/llm-llamacpp/benchmarks/performance
• ./packages/vla-ggml/sim/server
• ./packages/sdk/e2e
• ./docs/website