parakeet-server (examples/server, the OpenAI-compatible HTTP server)
already builds in the release workflow — PARAKEET_BUILD_SERVER defaults ON
and the configure step never disables it — but the Package steps only copy
parakeet-cli, so it never reaches the release assets.

Package it into the bin- bundles for every (platform, backend) pair, with a
--help usage smoke test, the same way llama.cpp / whisper.cpp ship
llama-server / whisper-server. BUILD_SHARED_LIBS=OFF keeps it a single
self-contained binary like the CLI, so no extra libraries are bundled.

The release workflow ships a vulkan bundle for linux x64 but linux arm64
only gets cpu, so arm64 hosts with a Vulkan GPU have nothing to pin.

Add a vulkan/arm64 entry to the build-linux matrix on ubuntu-24.04-arm.
LunarG publishes no arm64 SDK apt repo, so the LunarG install step is
gated to x64 and arm64 installs the toolchain from the Ubuntu archive:
libvulkan-dev (loader/headers), glslc (shaderc, to compile ggml's compute
shaders), and spirv-headers. glslangValidator is not required.

24.04 rather than 22.04 because glslc/shaderc is packaged on noble but not
jammy -- the same reason the cuda job already uses 24.04. The entry
produces parakeet-<ver>-bin-linux-vulkan-arm64.tar.gz like the others.
Like every CI-built Vulkan/CUDA bundle it is GPU-unvalidated (the runners
have no GPU); shader generation and linkage are exercised on the runner.

The decoder already computes a NeMo per-word confidence (pk::Word.conf, the
'min' aggregate of the word's per-token confidences) and parakeet-cli surfaces
it in --json as "conf", but the server's verbose_json words array dropped it,
emitting only word/start/end.

Add it under the same "conf" key (4 decimals) on each word object so clients can
flag low-confidence spans. The value is already on the Word, so this is a
formatting-only change; extend test_server_format.cpp to assert it.

Co-authored-by: kasf <[email protected]>

Assisted-by: Codex:gpt-5 [Codex]

Assisted-by: Codex:gpt-5 [Codex]

Assisted-by: Codex:gpt-5 [Codex]

)

Offline transcribe emitted raw special-token pieces (e.g. the <en-US>
language tag) at the end of the transcript because decode_enc_out(),
transcribe_16k_batch(), decode_enc_out_with_timestamps(), and
transcribe_16k_batch_with_timestamps() passed decoded ids straight to
detokenize() with no filtering. Add pk::strip_special_tokens() (drops
ids whose piece matches <...>/[...]) and apply it at all four call
sites, matching the filtering the streaming path already does for
<EOU>/<EOB>.

Assisted-by: Claude:claude-sonnet-5 [Claude Code]

Co-authored-by: Ettore Di Giacinto <[email protected]>