Linux: Enable Parakeet transcription backend

[nephalemsec]

Problem

Parakeet transcription models (via sherpa-rs/sherpa-onnx) are currently disabled on Linux because the sherpa-onnx pre-built archive only includes shared libraries (.so), not static libraries (.a). The build uses features = ["static", "download-binaries"] which fails at link time on Linux.

Current workarounds:

• CI uses --no-default-features to skip Parakeet on Linux
• Nix flake uses buildNoDefaultFeatures = true

Proposed Solutions

1. Dynamic linking (low effort, ~2-4h) — remove "static" from sherpa-rs features, use ["download-binaries"] only. Ship .so files with the app or require system library. Test if sherpa-rs handles the fallback correctly.

2. System package approach (medium effort, ~4-8h) — use system-provided libsherpa-onnx instead of download-binaries. Cleaner for Nix packaging. Requires packaging for major distros.

3. Upstream contribution (high effort) — request sherpa-onnx to include .a files in Linux releases. Long-term fix but requires coordination.

Code References

• Parakeet backend: src-tauri/src/transcription/parakeet.rs
• Cargo features: src-tauri/Cargo.toml (sherpa-rs dependency)
• CI exclusion: commit 5f02ab1
• Nix build: flake.nix (buildNoDefaultFeatures = true)

[poodle64]

Backlog decision 2026-05-30: deferred — Linux-specific (macOS operator). Note: PR #33 attempts this but as written it sets default = ['fluidaudio'], which would drop Parakeet from default features on ALL platforms incl. macOS; do not merge #33 as-is. Grouped with #28/#51/#52 for a future Linux pass.

[poodle64]

Triaged — keep open, needs a Linux build to validate.

Current state in main: parakeet.rs already gates CoreML to macOS and uses the CPU ONNX provider on non-macOS, so the code is Linux-ready. The blocker is the dependency: sherpa-rs = { features = ["static", "download-binaries"] } — upstream sherpa-onnx's Linux release archive ships only .so files, no .a, so the static feature fails to link on Linux. That's why the Linux release excludes Parakeet (--no-default-features).

Recommended approach (to validate on a Linux box): drop the static feature for the Linux target (features = ["download-binaries"] only, dynamic link), cargo build --features parakeet on Linux, confirm the model downloads and a transcription runs. If dynamic linking needs an rpath/patchelf fixup for the bundled .so, handle that in the Tauri bundle step.

Can't be exercised on the macOS dev machine (the failure is Linux-specific linking). Holding for a Linux session.

[poodle64]

Parakeet builds and links statically on Linux — the blocker is resolved

Verified on the x86_64 NixOS box (atlas):

• cargo build --no-default-features --features "vulkan,parakeet" compiles and links cleanly. The backend is the official k2-fsa sherpa-onnx 1.13.2 crate that replaced the deprecated sherpa-rs.
• It links statically: ldd on the resulting binary shows no dynamic sherpa/onnx dependencies and no missing libraries, and no onnxruntime/sherpa .so is left in the build tree. So there is nothing to bundle or patchelf — the original blocker (sherpa-rs shipped no Linux static archive) is fully resolved by the crate migration.

This means the year-old "can't link on Linux" problem is gone.

Why it is not yet enabled in the default Linux CI/release

Two honest reasons, both worth a deliberate decision rather than flipping silently:

1. Runtime is unverified on Linux. Linking ≠ working. A real Parakeet transcription on Linux (model auto-download + decode via the CPU ONNX provider) needs a live Linux box; atlas is headless. This runtime check is in the consolidated live-test issue (Linux desktop: final live-test of paste, tray, indicator, Wayland portal, and GPU/Parakeet runtime #81).
2. Build-time download. sherpa-onnx-sys downloads prebuilt binaries during the build. Adding parakeet to the Linux CI/release feature set introduces a network dependency that can make builds intermittently red — a reliability tradeoff to weigh first.

Remaining for this issue

• Runtime-verify a Parakeet transcription on a real Linux desktop (tracked in Linux desktop: final live-test of paste, tray, indicator, Wayland portal, and GPU/Parakeet runtime #81).
• Decide whether to add parakeet to the Linux release feature set (and CI), given the build-time download tradeoff.

docs/development/linux-setup.md has been updated: it no longer claims the sherpa-rs static-archive blocker, and documents that --features parakeet now builds on Linux. Keeping this issue open, narrowed to the two boxes above.

[poodle64]

Decision (v2026.6.2): keep Parakeet off the default Linux build; opt-in via --features parakeet.

Build/link blocker: resolved. The migration to the official k2-fsa sherpa-onnx 1.13.2 crate links statically on Linux — confirmed building on the x86_64 box. The old "sherpa-rs ships no Linux static archive" reason is gone; the stale ci.yaml/release.yaml comments are corrected in 53219cd.

Runtime: not clean — this is the new blocker. Added a headless smoke harness (src-tauri/examples/parakeet_smoke.rs) that drives the real transcription::parakeet path on a WAV. On the x86_64 NixOS box it crashes at recognizer init — free(): invalid pointer, immediately after "Using CPU provider" and before any audio is decoded. That is a heap/ABI fault, the classic signature of the prebuilt onnxruntime in sherpa-onnx-sys clashing with the Nix toolchain's libstdc++/allocator.

The shipped .deb/.AppImage build on ubuntu-22.04, not Nix, so this specific crash may not reproduce on a glibc desktop — but (a) the runtime is unverified on the actual release target, and (b) it definitely crashes under the Nix flake build the project uses on atlas. Shipping a backend that crashes on at least one supported Linux build path is not acceptable for a release, so Parakeet stays opt-in on Linux.

Remaining (narrowed)

• Run parakeet_smoke on a real Ubuntu/glibc desktop (cargo run --example parakeet_smoke --no-default-features --features vulkan,parakeet -- <model_dir> <audio.wav>). If it transcribes cleanly, add parakeet to the Linux CI + release feature set.
• Investigate the Nix free(): invalid pointer at model init — likely a libstdc++/onnxruntime pinning fix in the flake, or point the build at a Nix-built sherpa via SHERPA_ONNX_LIB_DIR.

Keeping this open, narrowed to those two boxes. The GUI side of a Linux Parakeet run is part of #81.