# Copyright 1999-2026 Gentoo Authors # Distributed under the terms of the GNU General Public License v2 EAPI=8 PYTHON_COMPAT=( python3_{12..14} ) inherit cmake cuda flag-o-matic python-single-r1 toolchain-funcs DESCRIPTION="Speech-to-text, TTS, speaker diarization etc. using onnxruntime" HOMEPAGE=" https://k2-fsa.github.io/sherpa/onnx/ https://github.com/k2-fsa/sherpa-onnx " # THE OVERLAY'S REFERENCE ANSWER TO CMake FetchContent. # # Upstream's build declares every vendored dependency through FetchContent, # which fetches at *configure* time — inside the sandbox, with no network. The # defeat is not a patch: each of upstream's cmake/*.cmake modules checks a # `possible_file_locations` list before reaching for the network, and that list # includes ${CMAKE_SOURCE_DIR}. So the archives are declared here in SRC_URI, # left archived, and copied into ${S} by src_unpack; configure then finds them # locally and never opens a socket. # # Three properties make this work, and all three are load-bearing: # - The filenames must match what the cmake modules look for, which is why # every entry carries an explicit `->` rename. A renamed-wrong distfile is # not an error, it is a silent fallback to the network. # - .zip stays .zip. Upstream ships kissfft, websocketpp, piper-phonemize and # espeak-ng as .zip and its modules name them that way; converting to # .tar.gz would miss the fallback. This is what BDEPEND=app-arch/unzip is # for — CMake extracts them at configure time. # # pkgcheck reports TarballAvailable on those four, suggesting the .tar.gz # equivalents. DO NOT act on it: it is a false positive here, and acting on # it silently restores the network fetch this whole block exists to remove. # Verified against the 1.13.5 sources — cmake/websocketpp.cmake, # cmake/piper-phonemize.cmake and cmake/espeak-ng-for-piper.cmake each build # a `possible_file_locations` list containing the literal .zip filename # under ${CMAKE_SOURCE_DIR} and fall through to `_URL` (the network) when no # entry EXISTS. The name is matched verbatim; a .tar.gz sitting beside it # does not satisfy the check. kissfft has no module of its own — it is # pulled by cmake/kaldi-native-fbank.cmake through the same mechanism. # - Sub-dependencies resolve through the same global ${CMAKE_SOURCE_DIR} # fallback (kissfft is pulled by kaldi-native-fbank, kaldifst by # kaldi-decoder), so one staging directory covers the whole tree. # # Every entry is pinned by tag or by full commit SHA. Do not relax one to a # branch: FetchContent would then resolve to whatever HEAD happens to be. SRC_URI=" https://github.com/k2-fsa/sherpa-onnx/archive/refs/tags/v${PV}.tar.gz -> ${P}.gh.tar.gz https://gitlab.com/libeigen/eigen/-/archive/5.0.1/eigen-5.0.1.tar.gz https://github.com/likle/cargs/archive/refs/tags/v1.0.3.tar.gz -> cargs-1.0.3.tar.gz https://github.com/csukuangfj/hclust-cpp/archive/refs/tags/2026-02-25.tar.gz -> hclust-cpp-2026-02-25.tar.gz https://github.com/nlohmann/json/archive/refs/tags/v3.12.0.tar.gz -> json-3.12.0.tar.gz https://github.com/k2-fsa/kaldi-decoder/archive/refs/tags/v0.3.0.tar.gz -> kaldi-decoder-0.3.0.tar.gz https://github.com/csukuangfj/kaldi-native-fbank/archive/refs/tags/v1.22.3.tar.gz -> kaldi-native-fbank-1.22.3.tar.gz https://github.com/csukuangfj/openfst/archive/refs/tags/v1.8.5-2026-07-09.tar.gz -> openfst-1.8.5-2026-07-09.tar.gz https://github.com/pkufool/simple-sentencepiece/archive/refs/tags/v0.7.tar.gz -> simple-sentencepiece-0.7.tar.gz https://github.com/mborgerding/kissfft/archive/febd4caeed32e33ad8b2e0bb5ea77542c40f18ec.zip -> kissfft-febd4caeed32e33ad8b2e0bb5ea77542c40f18ec.zip https://github.com/k2-fsa/kaldifst/archive/refs/tags/v1.8.0.tar.gz -> kaldifst-1.8.0.tar.gz portaudio? ( https://files.portaudio.com/archives/pa_stable_v190700_20210406.tgz ) python? ( https://github.com/pybind/pybind11/archive/refs/tags/v3.0.0.tar.gz -> pybind11-3.0.0.tar.gz ) tts? ( https://github.com/csukuangfj/espeak-ng/archive/ed530aa113046142eb5115cf2fc9157854d0ffe1.zip -> espeak-ng-ed530aa113046142eb5115cf2fc9157854d0ffe1.zip https://github.com/csukuangfj/piper-phonemize/archive/f3ff95afc03640bc1399e113e83361192a2fafb4.zip -> piper-phonemize-f3ff95afc03640bc1399e113e83361192a2fafb4.zip ) websocket? ( https://github.com/chriskohlhoff/asio/archive/refs/tags/asio-1-24-0.tar.gz -> asio-asio-1-24-0.tar.gz https://github.com/zaphoyd/websocketpp/archive/b9aeec6eaf3d5610503439b4fae3581d9aff08e8.zip -> websocketpp-b9aeec6eaf3d5610503439b4fae3581d9aff08e8.zip ) " LICENSE="Apache-2.0" SLOT="0" # arm64 is deliberately not keyworded, and the omission is not about upstream. # # Upstream does publish aarch64 and armv7l wheels, which is real evidence and # would normally be enough under this overlay's arm64 policy. But the ebuild's # own build path has never been exercised on arm64: it compiles a large C++ # tree plus eleven vendored dependencies, and it links against # sci-libs/onnxruntime, which is itself ~amd64 here. Keywording ~arm64 would # assert an arm64 build nobody has run, against a dependency that has no arm64 # keyword to satisfy it. # # This matches the decision taken across the ggml family on 2026-08-08 and # 2026-08-16: sci-ml/{llama-cpp,whisper-cpp,stable-diffusion-cpp,ik_llama-cpp, # koboldcpp} are all ~amd64 for the same reason. # # Restore ~arm64 when sci-libs/onnxruntime carries it AND a recorded arm64 # build of this package exists. Upstream wheels alone are not the missing # piece; they were never in doubt. KEYWORDS="~amd64" IUSE="cuda +portaudio +python +tts +websocket" REQUIRED_USE="python? ( ${PYTHON_REQUIRED_USE} )" # sherpa-onnx vendors and statically links portaudio (PA_BUILD_STATIC=ON in # cmake/portaudio.cmake), so USE=portaudio adds no system dependency — it only # toggles whether the mic-recording demo CLIs get built. media-libs/alsa-lib # *is* needed regardless: several *-alsa demo binaries link -lasound # unconditionally. # # The blocker against -bin is not currently reciprocated, because bentoo does # not carry sci-ml/sherpa-onnx-bin (see the note above src_install). Should one # ever be added, it MUST declare the mirror blocker in the same commit: the # sci-libs/onnxruntime{,-bin} pair shipped without one, shared 26 installed # paths, and only failed at the end of a 266 MB build. RDEPEND=" !sci-ml/sherpa-onnx-bin sci-libs/onnxruntime:= media-libs/alsa-lib cuda? ( dev-util/nvidia-cuda-toolkit:= ) python? ( ${PYTHON_DEPS} ) " DEPEND="${RDEPEND}" # app-arch/unzip is required at build time: CMake extracts the .zip-archived # vendored deps (kissfft, websocketpp, piper-phonemize, espeak-ng) during # configure. See the SRC_URI comment for why they stay .zip. BDEPEND=" app-arch/unzip python? ( ${PYTHON_DEPS} $(python_gen_cond_dep ' dev-python/pybind11[${PYTHON_USEDEP}] ') ) " # NO QA_PREBUILT HERE, DELIBERATELY. # # The adopted ebuild set QA_PREBUILT="opt/sherpa-onnx/lib/*" and justified it # as covering "prebuilt" libraries. That description is wrong: everything under # /opt/sherpa-onnx/lib is compiled by this ebuild from the sources staged in # src_unpack — the vendored .a and private .so files are build products, not # vendor blobs. QA_PREBUILT suppresses exactly the checks that should run on # freshly built code (stripping, RPATH, unresolved symbols), so setting it # would hide real defects in output this ebuild is responsible for. # # If a future version genuinely ships a prebuilt blob, scope QA_PREBUILT to # that one path and say which file and why — never to the whole lib directory. src_unpack() { # Only the main tarball is unpacked. The vendored dependency archives stay # archived and are copied into ${S}, where the cmake modules' # possible_file_locations check finds them. See the SRC_URI comment. unpack "${P}.gh.tar.gz" # ${A} is a space-separated string of distfile names, not an array. local f for f in ${A}; do [[ ${f} == "${P}.gh.tar.gz" ]] && continue cp -- "${DISTDIR}/${f}" "${S}/" || die "failed staging ${f} into ${S}" done } src_prepare() { # MUST be defined, even though it looks like boilerplate. # # cuda.eclass does `EXPORT_FUNCTIONS src_prepare`, and its cuda_src_prepare # calls cuda_sanitize unconditionally -- which resolves cuda_gccdir, which # dies with "cuda-config not found" when the CUDA toolkit is absent. With # no src_prepare of its own, this ebuild inherited that one and failed in # the prepare phase for EVERY user without CUDA installed, including # USE=-cuda. Caught by the merge gate on 2026-08-16; pkgcheck cannot see it. # # cmake_src_prepare, not `default`: the cmake eclass tracks state here that # cmake_src_configure later depends on. cmake_src_prepare use cuda && cuda_src_prepare } src_configure() { use python && python_setup # Keep the build directory out of the installed binaries. # # Upstream puts __FILE__ in its message macros, so the sandbox path is not # merely debug metadata — it is printed AT THE USER. `sherpa-onnx --help` # opened with # /var/tmp/portage/sci-ml/sherpa-onnx-1.13.5/work/.../parse-options.cc:415 # before this, and 113 such paths were baked into the installed files. # # Map ${WORKDIR}, not ${S}: the vendored dependencies are configured under # ${S}_build/_deps (kaldi_decoder-src and friends) and leak their own paths, # so remapping only the source tree would clear about half of them. # # -ffile-prefix-map covers __FILE__, __BASE_FILE__ and the debug records in # one flag; -fdebug-prefix-map would leave the user-visible strings alone, # which are the ones that matter here. The target path follows Gentoo's # split-debug convention so a debugger still resolves sources when # FEATURES=splitdebug installs them. append-flags -ffile-prefix-map="${WORKDIR}=/usr/src/debug/${CATEGORY}/${PF}" local mycmakeargs=( # Self-contained install under /opt keeps the ~23 CLI tools and the # vendored helper libraries (cargs etc.) out of /usr/{bin,lib}. -DCMAKE_INSTALL_PREFIX="${EPREFIX}/opt/sherpa-onnx" -DBUILD_SHARED_LIBS=ON # This is what makes the package link bentoo's sci-libs/onnxruntime # instead of downloading and bundling its own copy. It is the reason # story 002 had to land onnxruntime before this package could exist. -DSHERPA_ONNX_USE_PRE_INSTALLED_ONNXRUNTIME_IF_AVAILABLE=ON -DSHERPA_ONNX_ENABLE_C_API=ON -DSHERPA_ONNX_ENABLE_BINARY=ON -DSHERPA_ONNX_ENABLE_SPEAKER_DIARIZATION=ON -DSHERPA_ONNX_LINK_LIBSTDCPP_STATICALLY=OFF -DSHERPA_ONNX_ENABLE_TESTS=OFF -DSHERPA_ONNX_ENABLE_PYTHON=$(usex python) -DSHERPA_ONNX_ENABLE_PORTAUDIO=$(usex portaudio) -DSHERPA_ONNX_ENABLE_WEBSOCKET=$(usex websocket) -DSHERPA_ONNX_ENABLE_TTS=$(usex tts) -DSHERPA_ONNX_ENABLE_GPU=$(usex cuda) -DSHERPA_ONNX_ENABLE_JNI=OFF -DSHERPA_ONNX_ENABLE_WASM=OFF ) if use cuda; then # nvcc refuses a host compiler newer than the toolkit supports. The # adopted ebuild hardcoded /usr/bin/g++-15, which is wrong twice: it # names a bare path that does not exist on a ${CHOST}-prefixed # toolchain, and it pins a version that goes stale at the next toolkit # bump. cuda_gccdir asks the installed toolkit which GCC slots it # accepts and returns the newest acceptable one. export CUDAHOSTCXX="$(cuda_gccdir)/${CHOST}-g++" [[ -x ${CUDAHOSTCXX} ]] || die "CUDAHOSTCXX=${CUDAHOSTCXX} is not executable" cuda_add_sandbox fi cmake_src_configure } src_install() { cmake_src_install # Put /opt/sherpa-onnx/bin on PATH and its lib on LDPATH so the unprefixed # `sherpa-onnx` command works and the libraries resolve transparently. newenvd - 99sherpa-onnx <<-EOF PATH="${EPREFIX}/opt/sherpa-onnx/bin" LDPATH="${EPREFIX}/opt/sherpa-onnx/lib" EOF if use python; then # The bindings must land in site-packages so `import sherpa_onnx` # works without PYTHONPATH gymnastics. The pybind11 module dlopens # libsherpa-onnx-c-api.so and friends, which the env.d LDPATH above # makes reachable. local opt_pylib="${ED}/opt/sherpa-onnx/lib" # python_get_sitedir already carries ${EPREFIX}: it is computed from # base="${EPREFIX}/usr". It must therefore be joined with ${D}, never # with ${ED} (which is itself ${D}${EPREFIX}) and never passed to # dodir, which prepends ${ED}. The adopted ebuild did both, yielding # ${D}${EPREFIX}${EPREFIX}/usr/... — a doubled prefix that is invisible # on a normal install, where EPREFIX is empty, and misplaces every # binding under Prefix. local pylib_dst="${D}$(python_get_sitedir)/sherpa_onnx/lib" mkdir -p "${pylib_dst}" || die local so found=0 for so in "${opt_pylib}"/_sherpa_onnx*.so; do [[ -e ${so} ]] || continue mv "${so}" "${pylib_dst}/" || die found=1 done # Fail loudly rather than shipping bindings that cannot import: an # upstream rename of the extension module would otherwise pass here # and only surface as ImportError on the user's machine. [[ ${found} == 1 ]] || die "no _sherpa_onnx*.so found in ${opt_pylib}; upstream may have renamed the extension module" python_moduleinto sherpa_onnx python_domodule "${S}/sherpa-onnx/python/sherpa_onnx"/*.py fi } pkg_postinst() { elog "" elog "sherpa-onnx ${PV} installed to /opt/sherpa-onnx." elog "After re-sourcing /etc/profile (new shell or 'source /etc/profile')" elog "these are on PATH:" elog " sherpa-onnx --help # generic ASR/runtime entry" elog " sherpa-onnx-offline-speaker-diarization --help" elog " sherpa-onnx-vad --help # voice activity detection" elog " ... (~23 task-specific tools under /opt/sherpa-onnx/bin/)" elog "" elog "Model files are not bundled — see" elog " https://k2-fsa.github.io/sherpa/onnx/pretrained_models/" elog "For speaker diarization specifically, the ONNX-converted pyannote" elog "models and 3D-Speaker embeddings live at" elog " https://huggingface.co/csukuangfj/sherpa-onnx-pyannote-segmentation-3-0" elog "(ungated, no HuggingFace token required, unlike sci-ml/pyannote-audio)." elog "" }