commit 1602db544b49aec77b5ce455bf6a56f083975303
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Sep 11 00:50:28 2026 +0200

sci-ml/sherpa-onnx: trim 1.13.8 comments

commit 4e2143029add616d681824148b89557b6eae302e
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 02:02:51 2026 +0200

sci-ml/sherpa-onnx: trim 1.13.6 comment prose

commit 87c0f4e9036a0a021c0d381875178c3fcd899ba1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 02:02:23 2026 +0200

sci-ml/sherpa-onnx: trim 1.13.7 comment prose

commit 32cad369428b83ef00ab392d651f3df3595d9c62
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Thu Sep 10 22:34:30 2026 +0200

sci-ml/sherpa-onnx: add 1.13.8

commit dcdbddeffa15b40827d7b9f3e076eb4701c62c47
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Sep 2 10:56:50 2026 +0200

sci-ml/sherpa-onnx: document why kissfft stays a .zip

pkgcheck reports TarballAvailable on both versions, suggesting the
kissfft archive be switched from GitHub's .zip to the .tar.gz. Acting on
that would break the build, so record why rather than leave the finding
to be "fixed" by a later pass.

The archive is not ours to choose. kissfft is not fetched by
sherpa-onnx's own cmake but by kaldi-native-fbank, which vendors its own
cmake/kissfft.cmake two levels down. That module searches
possible_file_locations for the exact filename kissfft-<sha>.zip --
$/ being one of them, which is precisely where this
ebuild's src_unpack stages it -- and then verifies
SHA256=497103e664168ebe39580b757adbe616f6cf85a16572af581ca7bc42d0ab13fd,
which is the hash of the .zip.

So a rename breaks the filename match, and even if the name were patched
the hash check would still reject a .tar.gz of the same commit. Read out
of kaldi-native-fbank-1.22.3, the version both ebuilds pin.

Comment only; no functional change.

commit fa5a878245f7a53bf11719801362db73f8e96249
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Sep 1 11:27:49 2026 +0200

sci-ml/sherpa-onnx: drop 1.13.5

Retention: keep the last two (1.13.6, 1.13.7). Nothing outside the
package depends on it -- the only atoms naming sherpa-onnx are the
mutual blockers it and sci-ml/sherpa-onnx-bin carry for each other.

commit 19b9dd4882bf03ce94f333f134caf02c2c4ee004
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Sep 1 11:25:11 2026 +0200

sci-ml/sherpa-onnx: add 1.13.7

Build interface re-checked against the 1.13.6 tree rather than assumed:
diff -rq over cmake/ shows the ONLY changed modules are the six macOS
onnxruntime ones (osx-arm64, osx-universal, osx-x86_64, each in a static
and a shared variant). So every vendored FetchContent pin this ebuild
pre-stages is unchanged -- eigen 5.0.1, cargs 1.0.3, hclust-cpp, json
3.12.0, kaldi-decoder 0.3.0, kaldi-native-fbank 1.22.3, openfst,
simple-sentencepiece 0.7, kissfft, kaldifst 1.8.0, portaudio, pybind11
3.0.0, espeak-ng, piper-phonemize, asio and websocketpp -- and the
compressed-tarballs patch applies against a byte-identical
cmake/eigen.cmake. The top-level option() surface is identical too.

Built with default USE (portaudio python tts websocket): the CLI tools,
libsherpa-onnx-c-api.so, the site-packages _sherpa_onnx extension and
the env.d file are all in the image.

commit 64364c3d903fef3f50c274f5fe8fe2cd6e10ac94
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Aug 18 22:58:47 2026 +0200

sci-ml/sherpa-onnx: drop 1.13.4

commit feaa3c39e4367c70874cea584752d0b3bbabfe6a
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Aug 18 22:58:47 2026 +0200

sci-ml/sherpa-onnx: add 1.13.6

commit 78eb27ba44e325a5024f0d158c34bf8bbab356c3
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Fri Aug 14 00:40:51 2026 +0200

sci-ml/sherpa-onnx: propagate CUDA and prefer tarballs

SHERPA_ONNX_ENABLE_GPU selects the ONNX Runtime CUDA execution provider, so require a
matching backend instead of allowing CPU-only resolution. Use better-compressed
upstream archives for the direct FetchContent dependencies while retaining the nested
kissfft ZIP.

commit d32719384132b8ed3601a88d4b6a25e6b43e0c14
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Aug 12 12:29:19 2026 +0200

sci-ml/sherpa-onnx: remove host-specific CUDA wording

Public metadata should describe the compiler constraint portably rather than tying it
to one machine. nvcc still requires a supported host compiler, so retain that
requirement without naming a local pin.

commit 4da8d44e477dc49690eec69e8eb8df86fe5ab328
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Aug 11 23:28:33 2026 +0200

sci-ml/sherpa-onnx: drop 1.13.3

commit 3a53e4b568cfb5af6416f2c52c38707308a382b0
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Tue Aug 11 23:28:30 2026 +0200

sci-ml/sherpa-onnx: add 1.13.5

commit 7a4b0eb1eea6acc7143c95cc8f28595a928c6bbc
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat Jul 11 11:19:48 2026 +0200

sci-ml/sherpa-onnx: fix 1.13.4 vendored espeak-ng and piper-phonemize pins

v1.13.4 moved its FetchContent commit pins for espeak-ng and piper-phonemize, but the ebuild kept the 1.13.3 revisions. The prestaged tarball names no longer matched what cmake/espeak-ng-for-piper.cmake and cmake/piper-phonemize.cmake search for, so FetchContent fell back to a network download that the build sandbox blocks. Re-pin espeak-ng to ed530aa1 and piper-phonemize to f3ff95af to match v1.13.4; kissfft and websocketpp are unchanged.

commit 637ab2cbf802c6d08e45f4cafc1e2539aa5902e1
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jul 8 09:57:11 2026 +0200

sci-ml/sherpa-onnx: add 1.13.4

commit 9837bda6646e68a1e707f1428824b4f6560fa2ed
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Wed Jun 17 18:01:58 2026 +0200

sci-ml/sherpa-onnx: Keyword 1.13.3 for ~arm64

Speech toolkit on top of the onnxruntime C++ API; built with the
default USE (portaudio, python, tts). aarch64-portable.

commit beabe0da9bc74059db54c0ee3fdbf32be5e0e27b
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jun 15 23:53:50 2026 +0200

sci-ml/sherpa-onnx: generalize CUDA host-compiler comment

Drop the maintainer-private notes reference and host-specific wording
from the CUDAHOSTCXX rationale comment; the gcc>15 pin reason is kept.

commit 901ee16e37f059155ac82ec7146361557116ea61
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jun 15 15:40:44 2026 +0200

sci-ml/sherpa-onnx: drop 1.13.2

Superseded by 1.13.3. 1.13.2 staged its vendored FetchContent deps in
formats the bundled cmake possible_file_locations never match (kissfft/
espeak-ng/piper-phonemize/websocketpp as .tar.gz, eigen as .tar.bz2), so
it could only configure with network access -- broken under the default
network sandbox. 1.13.3 carries the corrected staging.

commit 201f6cdf8a49c4ee12af1b7f5b91ab2202077743
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon Jun 15 15:34:48 2026 +0200

sci-ml/sherpa-onnx: add 1.13.3

Stage the vendored FetchContent deps with the archive formats the
bundled cmake modules' possible_file_locations actually look for:
kissfft, espeak-ng, piper-phonemize and websocketpp as .zip (not
.tar.gz), and eigen as .tar.gz (not .tar.bz2). The earlier renames were
never found under $, so a network-sandboxed configure
fell through to a FetchContent download and died. The app-arch/unzip
BDEPEND already anticipated the .zip inputs.

commit faafbeab78d7ddff0599a7caedf18f30b8724cce
Author: Raukaan Cogbrother <cogbrother@raukaan.local>
Date: Sun May 31 03:41:08 2026 +0200

sci-ml/sherpa-onnx: switch four github ZIPs and eigen to better compressions

GitHub serves each archive ref as either .zip or .tar.gz; the tarball
form ships about half the bytes for the same extracted tree. Eigen's
GitLab archive endpoint offers .tar.bz2 with a similar size win.
Swap the four commit-pinned GitHub archive URLs (kissfft, espeak-ng,
piper-phonemize, websocketpp) to .tar.gz and the eigen 5.0.1 archive
to .tar.bz2 to clear pkgcheck TarballAvailable and BetterCompressionUri.
Extracted contents are identical; just the distfile compression
changed.

commit de9fa32587a3d1678eded88c22ee659769c18d51
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Mon May 18 19:48:26 2026 +0200

sci-ml/sherpa-onnx: switch portaudio SRC_URI to https

files.portaudio.com serves the same pa_stable_v190700_20210406.tgz
over HTTPS; switch the SRC_URI to https://. Manifest unchanged
(same file content).

The remaining pkgcheck advisories on this ebuild — TarballAvailable
on 4 vendored zips (kissfft, websocketpp, piper-phonemize,
espeak-ng) and BetterCompressionUri on eigen — are intentional
per the BDEPEND comment: CMake's possible_file_locations fallback
expects the specific .zip filenames, so switching extensions
would break the build.

commit c2c8fcf614864689333caa08a596a79cebda7b57
Author: Ivan S. Titov <iohann.s.titov@gmail.com>
Date: Sat May 16 15:17:25 2026 +0200

sci-ml/sherpa-onnx: new package, add 1.13.2

Source build of the k2-fsa speech stack: ASR, TTS, speaker
diarization, VAD, source separation, KWS. Sibling of the prebuilt
sci-ml/sherpa-onnx-bin, depends on the freshly-landed
sci-libs/onnxruntime for inference.

Installs self-contained under /opt/sherpa-onnx (23 CLI binaries +
2 public .so libs + private vendored .so files for cargs/kaldi/
kissfft/fst). /etc/env.d/99sherpa-onnx exposes PATH + LDPATH so
`sherpa-onnx --help` works post-emerge. Python bindings (USE=python)
land in standard site-packages with the dlopen-able libs resolving
via LDPATH.

USE flags: cuda, portaudio, python, tts, websocket — all default off
on first land. cuda path pins CUDAHOSTCXX=g++-15 unconditionally per
the internal CUDA-13 host-compiler note.

RESTRICT=network-sandbox because the CMake build pulls 7+ small
vendored deps (eigen, asio, cargs, json, hclust-cpp, kaldi-decoder,
kaldi-native-fbank, optionally espeak-ng) via FetchContent. Hashing
each one into SRC_URI is friction-without-benefit for a first cut.

Build + smoke-tested with all USE off; binary runs, links against
system sci-libs/onnxruntime.