Source of the native libraries (sherpa-onnx 1.13.8)

The source of audio-chat-box's native libraries

audio-chat-box links sherpa-onnx's prebuilt static library, which includes
eSpeak NG (GPL-3.0-or-later), so the program as built
is GPL-3.0-or-later. Whoever has a copy of it may have its Corresponding Source: the
audio-chat-box repository at the commit it was built
from, and the source in this release of every library in that static library the program
links, as the builds that made the library fetched them. Each archive keeps its own licence.

This release is https://github.com/cjus/audio-chat-box/releases/tag/sources-v1: every
archive named below can be downloaded from it, with SHA256SUMS.

It covers every copy of audio-chat-box built with sherpa-onnx 1.13.8 (Cargo.toml) for
Linux (x86-64 and arm64) and macOS (Apple silicon and Intel), whose libraries are:

Platform	sherpa-onnx's static library	SHA-256
Linux x86-64	sherpa-onnx-v1.13.8-linux-x64-static-lib.tar.bz2	e1fdc5b67530e15741ef897fa5ffff297056f3bf0c6d829a27af9225a4c4b5a6
Linux arm64	sherpa-onnx-v1.13.8-linux-aarch64-static-lib.tar.bz2	77983e3cf29aa60f2e531d249dbd01d15596530550c8db2e9e02fc6a655da6bb
macOS arm64	sherpa-onnx-v1.13.8-osx-arm64-static-lib.tar.bz2	9091bf160dc7fdacedbc906b212badf53c2993f4e5277a0e03998e96c31d60da
macOS x86-64	sherpa-onnx-v1.13.8-osx-x64-static-lib.tar.bz2	a3f88da3e54c850a12d61431e73f8affcd1f13738b75b768847dd79541835b4b

That includes the Docker image pnxstudios/audio-chat-box-demo from 0.1.0 (commit 4190c28)
on: its org.opencontainers.image.revision label names the commit of each.

What is in it

Library in the static library	Source here
libsherpa-onnx-c-api.a, libsherpa-onnx-core.a	sherpa-onnx-1.13.8.tar.gz, with nlohmann-json-3.12.0.tar.gz and hclust-cpp-2026-02-25.tar.gz compiled in
libkaldi-decoder-core.a	kaldi-decoder-0.3.0.tar.gz, with eigen-5.0.1.tar.gz (header-only)
libsherpa-onnx-kaldifst-core.a	kaldifst-1.8.0.tar.gz
libsherpa-onnx-fst.a, libsherpa-onnx-fstfar.a	openfst-1.8.5-2026-07-09.tar.gz
libkaldi-native-fbank-core.a	kaldi-native-fbank-1.22.3.tar.gz
libkissfft-float.a	kissfft-febd4caeed32.zip
libpiper_phonemize.a	piper-phonemize-f3ff95afc036.zip (with uni-algo in it)
libespeak-ng.a, libucd.a	espeak-ng-ed530aa11304.zip
libssentencepiece_core.a	simple-sentencepiece-0.7.tar.gz
libonnxruntime.a	onnxruntime-1.28.2.tar.gz and its three onnxruntime-1.28.2-submodule-* archives, the onnxruntime-dep-* archives its build fetched (all twelve on Linux; on macOS all but cpuinfo), and onnxruntime-build-6314c7b.tar.gz, which built it

The Linux and macOS libraries hold the same fifteen static libraries. They also hold
libsherpa-onnx-portaudio_static.a and libsherpa-onnx-cxx-api.a, which audio-chat-box
doesn't link (sherpa-onnx-sys names the libraries it links, and neither is among them), so
PortAudio's source isn't here. sherpa-onnx-1.13.8.tar.gz has the C++ API's anyway.

How each was built

Every one is built from sherpa-onnx's CMake files (cmake/*.cmake) at v1.13.8, which fetch
the archives above, each checked by the SHA-256 they give it. kaldi-decoder's own cmake/
would fetch Eigen 3.4.0 and kaldifst's OpenFst 2026-04-10, but sherpa-onnx's cmake/ comes
first in CMake's module path, so its eigen.cmake (5.0.1) and openfst.cmake (2026-07-09)
are the ones used (each build's log shows it).

• sherpa-onnx's static library for Linux x86-64: its .github/workflows/linux.yaml
  (GitHub Actions run 34486043450, job "Release shared-OFF tts-ON"), in the
  quay.io/pypa/manylinux2014_x86_64 image, with shared libraries off.

• sherpa-onnx's static library for Linux arm64: its
  .github/workflows/aarch64-linux-gnu-static.yaml (run 34486043530, job "aarch64 static lib
  test (ubuntu-22.04-arm)"), built on an arm64 Ubuntu 22.04 runner.

• sherpa-onnx's static libraries for macOS: its .github/workflows/macos.yaml (run
  34486043393, job "Release static tts-ON macos-latest arm64;x86_64"), which builds one
  universal library for both architectures, then cuts the arm64 and x86-64 libraries out of it
  with lipo -thin. Being universal, it links the universal2 ONNX Runtime
  (cmake/onnxruntime-osx-universal-static.cmake).

• ONNX Runtime 1.28.2, which sherpa-onnx links prebuilt from csukuangfj/onnxruntime-libs:
  csukuangfj/onnxruntime-build at commit 6314c7b built each from Microsoft's tag v1.28.2
  (commit 33ca962) with its submodules, after applying that repository's patch to ONNX
  Runtime's CMake files; protoc 21.12 ran as a build tool. Each build fetched the
  onnxruntime-dep-* archives ("-- Fetch ... from" in its log); the rest of ONNX Runtime's
  cmake/deps.txt (GPU, mobile, web and test dependencies) wasn't fetched.

  Platform	ONNX Runtime (onnxruntime-libs v1.28.2)	Built by	Fetched
  Linux x86-64	onnxruntime-linux-x64-static_lib-1.28.2-glibc2_17.zip	linux-217.yaml, run 33741118168, in quay.io/pypa/manylinux2014_x86_64	all twelve
  Linux arm64	onnxruntime-linux-aarch64-static_lib-1.28.2-glibc2_17.zip	aarch64-217.yaml, run 33741118071	all twelve
  macOS	onnxruntime-osx-universal2-static_lib-1.28.2.zip	macos-universal2.yaml, run 33741118112	all but cpuinfo

  Each library's libonnxruntime.a is that build's, byte for byte (on macOS, its arm64 slice
  is the arm64 library's).

The logs of all these runs are in the audio-chat-box repository, in
tools/sources/build-logs/,
as GitHub keeps a run's log for only 90 days.

eSpeak NG's pronunciation data

Kokoro's model archive carries eSpeak NG's compiled pronunciation data (espeak-ng-data/,
GPL-3.0-or-later), which sherpa-onnx copied in from the espeak-ng-data.tar.bz2 of its
tts-models release (2023-11-29). Neither records what compiled it, so this was found by
compiling candidates and comparing them with the data in
kokoro-multi-lang-v1_0.tar.bz2 (the one in this repository's models-v1 release):

• Its source is espeak-ng-8593723f10cf.zip: sherpa-onnx's fork of eSpeak NG at
  8593723 (2023-11-27, the head of the fork and of rhasspy's when the data was published),
  built with CMake. Of the 355 files, 343 are byte for byte the same as that build's.
• The other 12 (the bn, ja, shn, my, kok, sd, chr, ko, vi, as, da
  and th dictionaries) are the same size and hold exactly the same bytes in another order,
  near the end of each, where the compiler sorts its rules. Builds elsewhere differ the same
  way (rhasspy's piper-phonemize 2023.11.14-4 data does too), while two builds in one
  environment are identical: the C library's sorting orders equal entries its own way.
• Built with autotools, the same source gives a different fa_dict (293,235 bytes against
  292,423), so the CMake build is the one that made it.
• espeak-ng-ed530aa11304.zip, which the program links, is 9 commits on from 8593723,
  none of them to the data's sources (dictsource/, phsource/): built the same way in the
  same environment, the two give byte-identical data.

To compile it: cmake -S . -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build in the
unpacked archive puts it in build/espeak-ng-data.

Checking it

SHA256SUMS has the SHA-256 of every file here. Each archive is also the one upstream pins,
by the hash in tools/sources/sources.txt:
sherpa-onnx's, kaldi-decoder's and kaldi-native-fbank's CMake files (SHA-256) and ONNX
Runtime's cmake/deps.txt (SHA-1); audio-chat-box pinned the GitHub archives of
sherpa-onnx, ONNX Runtime and its submodules, and onnxruntime-build itself.
tools/sources.sh downloads them all again and checks them.

The GNU General Public License, version 3, is LICENSE-GPL-3.0 in the audio-chat-box
repository.

Models, v1

The models audio-chat-box downloads by default (src/models.rs, MODELS_URL), byte for byte the archives it checks by size and SHA-256:

• Kokoro 82M v1.0 (Apache-2.0): kokoro-multi-lang-v1_0.tar.bz2, and kokoro-int8-multi-lang-v1_0.tar.bz2 (8-bit).
• Parakeet TDT 0.6B v2, 8-bit (CC-BY-4.0, by NVIDIA): sherpa-onnx-nemo-parakeet-tdt-0.6b-v2-int8.tar.bz2.
• Silero VAD v5 (MIT): silero-vad-v5.tar.bz2.

The Kokoro and Parakeet archives are sherpa-onnx's own, from its tts-models and asr-models releases, copied unchanged when they were chosen. Silero VAD, which comes as a bare file, is packed by tools/pack-silero-vad.sh. NOTICE.txt gives each part's licence, and SHA256SUMS their checksums. A product may host the same files itself and pass --models-url.