Gentoo Packages

# 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 python-single-r1

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
"

# FetchContent searches CMAKE_SOURCE_DIR for exact archive names, including
# transitive kissfft and kaldifst; prefetch and stage them to keep builds offline.
# Keep kissfft as ZIP despite TarballAvailable: kaldi-native-fbank verifies that
# filename and archive hash, so GitHub's tarball fails. verified 2026-09-02
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.bz2
	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.tar.gz
			-> espeak-ng-ed530aa113046142eb5115cf2fc9157854d0ffe1.tar.gz
		https://github.com/csukuangfj/piper-phonemize/archive/f3ff95afc03640bc1399e113e83361192a2fafb4.tar.gz
			-> piper-phonemize-f3ff95afc03640bc1399e113e83361192a2fafb4.tar.gz
	)
	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.tar.gz
			-> websocketpp-b9aeec6eaf3d5610503439b4fae3581d9aff08e8.tar.gz
	)
"

LICENSE="Apache-2.0"
SLOT="0"
KEYWORDS="~amd64 ~arm64"
IUSE="cuda +portaudio +python +tts +websocket"
REQUIRED_USE="python? ( ${PYTHON_REQUIRED_USE} )"

# portaudio is vendored and static; its flag only controls microphone demo CLIs.
# ALSA demos always link libasound. Block -bin because both install sherpa_onnx.
RDEPEND="
	!sci-ml/sherpa-onnx-bin
	sci-libs/onnxruntime:=[cuda?]
	media-libs/alsa-lib
	cuda? ( dev-util/nvidia-cuda-toolkit:= )
	python? (
		${PYTHON_DEPS}
	)
"
DEPEND="${RDEPEND}"
BDEPEND="
	app-arch/unzip
	python? (
		${PYTHON_DEPS}
		$(python_gen_cond_dep '
			dev-python/pybind11[${PYTHON_USEDEP}]
		')
	)
"

# Private vendored libraries under /opt are outside the system ABI.
QA_PREBUILT="opt/sherpa-onnx/lib/*"

src_unpack() {
	# Leave dependency archives intact for FetchContent's source-directory lookup.
	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
	done
}

PATCHES=(
	"${FILESDIR}/${PN}-1.13.4-compressed-tarballs.patch"
)

src_configure() {
	use python && python_setup

	local mycmakeargs=(
		# Isolate the many CLI tools and vendored libraries under /opt.
		-DCMAKE_INSTALL_PREFIX="${EPREFIX}/opt/sherpa-onnx"
		-DBUILD_SHARED_LIBS=ON
		-DSHERPA_ONNX_USE_PRE_INSTALLED_ONNXRUNTIME_IF_AVAILABLE=ON
		-DSHERPA_ONNX_ENABLE_C_API=ON
		-DSHERPA_ONNX_ENABLE_BINARY=ON
		-DSHERPA_ONNX_ENABLE_SPEAKER_DIARIZATION=ON
		-DSHERPA_ONNX_LINK_LIBSTDCPP_STATICALLY=OFF
		-DSHERPA_ONNX_ENABLE_TESTS=OFF
		-DSHERPA_ONNX_ENABLE_PYTHON=$(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
		# CUDA 13 nvcc rejects GCC >15.
		export CUDAHOSTCXX="/usr/bin/g++-15"
	fi

	cmake_src_configure
}

src_install() {
	cmake_src_install

	# Expose the isolated commands and private libraries system-wide.
	newenvd - 99sherpa-onnx <<-EOF
		PATH="${EPREFIX}/opt/sherpa-onnx/bin"
		LDPATH="${EPREFIX}/opt/sherpa-onnx/lib"
	EOF

	if use python; then
		# Put bindings in site-packages; their private libraries resolve via LDPATH.
		local opt_pylib="${ED}/opt/sherpa-onnx/lib"
		local pylib_dst_rel="$(python_get_sitedir)/sherpa_onnx/lib"

		dodir "${pylib_dst_rel}"
		local so
		for so in "${opt_pylib}"/_sherpa_onnx*.so; do
			[[ -e ${so} ]] || continue
			mv "${so}" "${ED}/${pylib_dst_rel}/" || die
		done

		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 "the binaries 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 + 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 ""
}