Rust examples: advanced users

This note is for users who want to control how the sherpa-onnx Rust crate finds and links its native libraries.

Most users do not need anything here. The default behavior is:

1. build normally
2. let the build script download the matching native libraries automatically
3. run the examples or your own Rust program

Use your own sherpa-onnx libraries

If you already have sherpa-onnx libraries on disk, set SHERPA_ONNX_LIB_DIR to the lib directory before building:

export SHERPA_ONNX_LIB_DIR=/path/to/sherpa-onnx/lib

Examples:

• source build output: /path/to/sherpa-onnx/build/install/lib
• manually extracted release archive: /path/to/sherpa-onnx-v1.13.7-linux-x64-static-lib/lib

If SHERPA_ONNX_LIB_DIR is set, the build script uses that directory and does not auto-download another archive.

Automatic download behavior

If SHERPA_ONNX_LIB_DIR is not set, sherpa-onnx-sys/build.rs downloads a matching prebuilt -lib archive from GitHub releases and uses its lib directory automatically.

The build script currently selects archives like this:

Default mode

Default mode uses the default crate feature set, which means static linking. Most users just get this behavior automatically.

Shared mode

If you enable the shared feature, the build script downloads these shared archives instead:

In practice, use the latest release tag instead of the example version above.

Configure the sherpa-onnx crate in Cargo.toml

Default configuration

This is enough for most users:

[dependencies]
sherpa-onnx = "1.13.7"

That uses the crate's default feature set.

Use shared libraries

To use shared libraries instead, disable default features and enable shared:

[dependencies]
sherpa-onnx = { version = "1.13.7", default-features = false, features = ["shared"] }

From the command line, the equivalent example command is:

cargo run --no-default-features --features shared --example version

Enable microphone examples

In rust-api-examples, microphone support is controlled by the mic feature:

cargo run --features mic --example streaming_zipformer_microphone -- --help

If you want both microphone support and shared libraries:

cargo run --no-default-features --features "shared,mic" \
  --example streaming_zipformer_microphone -- --help

Notes about runtime behavior

When shared libraries are used:

• Linux and macOS: the build script adds both absolute and relative runtime rpath entries automatically
• Linux and macOS: the build script also copies the required shared runtime libraries next to Cargo-generated binaries and examples
• Windows: the build script copies the required DLLs next to the generated binaries automatically
• iOS: the build script copies the required shared runtime libraries next to Cargo-generated binaries (same behavior as macOS)

When SHERPA_ONNX_LIB_DIR is set, the same behavior applies, but the files come from your directory instead of an auto-downloaded archive.

Cross-compiling for iOS

iOS supports real devices (aarch64-apple-ios) and simulators (aarch64-apple-ios-sim, x86_64-apple-ios). iOS only supports shared linking (the build script auto-selects shared mode for iOS, like Android).

Prerequisites

• macOS with Xcode installed
• Rust iOS target: rustup target add aarch64-apple-ios
• Tauri CLI: cargo install tauri-cli

Build with Tauri

The easiest way to build an iOS app with sherpa-onnx is through Tauri v2. See the hello_world example for a working setup.

cd tauri-examples/hello_world
npm install
cargo tauri ios init
cargo tauri ios build

The build script automatically downloads the iOS xcframework archive and creates symlinks so the linker finds the libraries under the expected names.

Build for iOS

cargo build --target aarch64-apple-ios --release

This downloads the sherpa-onnx-v{VERSION}-ios-shared-onnxruntime-static.xcframework.zip archive from the xcframework release tag. The shared dylib has onnxruntime statically linked in, so no separate onnxruntime framework is needed.

Use your own iOS xcframework

If you already have an iOS xcframework, set SHERPA_ONNX_LIB_DIR to a directory containing libsherpa-onnx-c-api.dylib:

export SHERPA_ONNX_LIB_DIR=/path/to/ios-dylib-dir
cargo build --target aarch64-apple-ios --release

Alternatively, set SHERPA_ONNX_ARCHIVE_DIR to a directory containing the .xcframework.zip archive, and the build script will use it instead of downloading:

export SHERPA_ONNX_ARCHIVE_DIR=/path/to/archives
# The build script looks for sherpa-onnx-v{VERSION}-ios-shared-onnxruntime-static.xcframework.zip
# in this directory.

iOS archive structure

The iOS xcframework archive is published on the xcframework release tag (not the versioned release tag):

sherpa-onnx-v{VERSION}-ios-shared-onnxruntime-static.xcframework.zip

After extraction, the xcframework contains:

SherpaOnnxC.xcframework/
  ios-arm64/
    SherpaOnnxC.framework/
      SherpaOnnxC          # shared dylib with onnxruntime statically linked
      Headers/
      Modules/
      Info.plist
  ios-arm64_x86_64-simulator/
    SherpaOnnxC.framework/
      ...

The build script finds ios-arm64/SherpaOnnxC.framework/SherpaOnnxC and creates a symlink libsherpa-onnx-c-api.dylib in a lib/ directory so the Rust linker finds the library under the expected name.

Cross-compiling for Android

Android support uses shared libraries only (static linking is not supported). The build script downloads a single archive containing all ABIs.

Supported ABIs

Prerequisites

• Android NDK (r27 or later recommended)
• Rust Android targets: rustup target add aarch64-linux-android armv7-linux-androideabi
• Set ANDROID_NDK_HOME or ANDROID_NDK_ROOT to your NDK installation
• Tauri CLI: cargo install tauri-cli

Build with Tauri

The easiest way to build an Android app with sherpa-onnx is through Tauri v2. See the hello_world example for a working setup.

cd tauri-examples/hello_world
npm install
cargo tauri android init
cargo tauri android build

The build script automatically downloads the Android archive and places the shared libraries in the correct jniLibs/{abi}/ directory.

Build a shared library for Android

# Example: arm64
cargo build --target aarch64-linux-android --release --no-default-features --features shared

# Example: armv7
cargo build --target armv7-linux-androideabi --release --no-default-features --features shared

This downloads sherpa-onnx-v{VERSION}-android.tar.bz2 from GitHub releases and extracts the shared libraries for the target ABI.

Android archive structure

The Android archive is a single .tar.bz2 file containing all ABIs:

sherpa-onnx-v{VERSION}-android/
  jniLibs/
    arm64-v8a/
      libsherpa-onnx-c-api.so
      libonnxruntime.so
    armeabi-v7a/
      libsherpa-onnx-c-api.so
      libonnxruntime.so
    x86/
      libsherpa-onnx-c-api.so
      libonnxruntime.so
    x86_64/
      libsherpa-onnx-c-api.so
      libonnxruntime.so

The build script selects the correct ABI directory based on the Rust target triple (e.g., aarch64-linux-android maps to arm64-v8a).

Use your own Android libraries

If you already have Android shared libraries, set SHERPA_ONNX_LIB_DIR to a directory containing the .so files for your target ABI:

export SHERPA_ONNX_LIB_DIR=/path/to/jniLibs/arm64-v8a
cargo build --target aarch64-linux-android --release --no-default-features --features shared

Alternatively, set SHERPA_ONNX_ARCHIVE_DIR to a directory containing the sherpa-onnx-v{VERSION}-android.tar.bz2 archive:

export SHERPA_ONNX_ARCHIVE_DIR=/path/to/archives
cargo build --target aarch64-linux-android --release --no-default-features --features shared

Notes

• Android only supports the shared feature. Do not use the default static feature for Android targets.
• The archive includes libonnxruntime.so alongside libsherpa-onnx-c-api.so. Both are required at runtime.
• The build script sets rpath to $ORIGIN so the shared libraries are found relative to the binary at runtime.

Building on Linux

Linux supports both static and shared linking on x86_64 and aarch64.

Supported architectures

Prerequisites

• GCC or Clang
• For aarch64 cross-compilation: rustup target add aarch64-unknown-linux-gnu and an aarch64 cross-compiler (e.g., gcc-aarch64-linux-gnu)

Build with Tauri

cd tauri-examples/hello_world
npm install
npm run build

Build with default (static) linking

cargo build --release

This downloads the matching static archive and produces a self-contained binary. The static build links against libstdc++, libm, libpthread, and libdl.

Build with shared linking

cargo build --release --no-default-features --features shared

This downloads the matching shared archive. The build script:

• sets rpath to $ORIGIN so libsherpa-onnx-c-api.so and libonnxruntime.so are found next to the binary at runtime
• copies the .so files next to the generated binaries and examples

Cross-compile for aarch64

# Static
cargo build --target aarch64-unknown-linux-gnu --release

# Shared
cargo build --target aarch64-unknown-linux-gnu --release \
  --no-default-features --features shared

Build from source

If you prefer building sherpa-onnx from source instead of downloading prebuilt archives:

mkdir build && cd build
cmake \
  -DBUILD_SHARED_LIBS=ON \
  -DSHERPA_ONNX_ENABLE_PORTAUDIO=OFF \
  -DSHERPA_ONNX_ENABLE_WEBSOCKET=OFF \
  -DBUILD_ESPEAK_NG_EXE=OFF \
  -DSHERPA_ONNX_ENABLE_BINARY=OFF \
  -DCMAKE_INSTALL_PREFIX=./install \
  ..
cmake --build . --target install --config Release

export SHERPA_ONNX_LIB_DIR=$PWD/install/lib
cargo build --release --no-default-features --features shared

Use -DBUILD_SHARED_LIBS=OFF for static linking instead.

Use your own libraries

export SHERPA_ONNX_LIB_DIR=/path/to/lib
cargo build --release

Or point to a directory containing the release archive:

export SHERPA_ONNX_ARCHIVE_DIR=/path/to/archives
cargo build --release

Notes

• The shared archives include both libsherpa-onnx-c-api.so and libonnxruntime.so. Both are required at runtime.
• The aarch64 shared archive has a -cpu suffix in its name (linux-aarch64-shared-cpu-lib) to distinguish it from GPU variants.
• For static builds, the binary is self-contained with no runtime .so dependencies (aside from system libraries).

Building on Windows

Windows supports both static and shared linking on x86_64 and arm64. All Windows builds use the MSVC toolchain with static CRT (/MT).

Supported architectures

Prerequisites

• Visual Studio 2022 (or later) with the "Desktop development with C++" workload
• Rust MSVC toolchain: rustup target add x86_64-pc-windows-msvc
• For arm64: rustup target add aarch64-pc-windows-msvc

Important: select the correct MSVC linker

Cargo on Windows may find Git's link.exe before MSVC's. You must explicitly point Cargo at the MSVC linker by setting the appropriate environment variable:

For x86_64:

$env:CARGO_TARGET_X86_64_PC_WINDOWS_MSVC_LINKER = "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Tools\MSVC\<version>\bin\HostX64\x64\link.exe"

For arm64:

$env:CARGO_TARGET_AARCH64_PC_WINDOWS_MSVC_LINKER = "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Tools\MSVC\<version>\bin\HostX64\ARM64\link.exe"

Alternatively, use ilammy/msvc-dev-cmd@v1 in CI and set the linker from $env:VCToolsInstallDir:

$linker = Join-Path $env:VCToolsInstallDir 'bin\HostX64\x64\link.exe'
$env:CARGO_TARGET_X86_64_PC_WINDOWS_MSVC_LINKER = $linker

Build with Tauri

cd tauri-examples\hello_world
npm install
npm run build

Build with default (static) linking

cargo build --release

This downloads the matching static archive and produces a self-contained binary. The static CRT (/MT) is used automatically.

Build with shared linking

cargo build --release --no-default-features --features shared

This downloads the matching shared archive. The build script copies the required DLLs next to the generated binaries automatically:

• sherpa-onnx-c-api.dll
• onnxruntime.dll
• onnxruntime_providers_shared.dll

All three DLLs are required at runtime.

Cross-compile for arm64 from x64

On an x64 Windows host, you can cross-compile for arm64:

rustup target add aarch64-pc-windows-msvc

# Set the cross-compilation linker
$env:CARGO_TARGET_AARCH64_PC_WINDOWS_MSVC_LINKER = "C:\Program Files\Microsoft Visual Studio\2022\Enterprise\VC\Tools\MSVC\<version>\bin\HostX64\ARM64\link.exe"

# Static
cargo build --target aarch64-pc-windows-msvc --release

# Shared
cargo build --target aarch64-pc-windows-msvc --release --no-default-features --features shared

Build from source

mkdir build
cd build
cmake -A x64 -DBUILD_SHARED_LIBS=ON -DSHERPA_ONNX_ENABLE_PORTAUDIO=OFF -DCMAKE_INSTALL_PREFIX=./install ..
cmake --build . --config Release
cmake --build . --config Release --target install

$env:SHERPA_ONNX_LIB_DIR = "$PWD\install\lib"
cd ..
cargo build --release --no-default-features --features shared

Use -DBUILD_SHARED_LIBS=OFF for static linking instead. For arm64, replace -A x64 with -A ARM64 and run from a Developer Command Prompt with the arm64 toolchain.

Use your own libraries

$env:SHERPA_ONNX_LIB_DIR = "C:\path\to\lib"
cargo build --release

Or point to a directory containing the release archive:

$env:SHERPA_ONNX_ARCHIVE_DIR = "C:\path\to\archives"
cargo build --release

Notes

• All Windows archives use static CRT (/MT). Dynamic CRT (/MD) variants are available on the release page but are not used by the Rust build script.
• The shared archives include sherpa-onnx-c-api.dll, onnxruntime.dll, and onnxruntime_providers_shared.dll. All three are required at runtime.
• For static builds, the binary is self-contained with no runtime DLL dependencies (aside from system DLLs).
• If you see link.exe errors, make sure the MSVC linker is used instead of Git's link.exe. See the "select the correct MSVC linker" section above.

Building on macOS

macOS supports both static and shared linking on arm64 (Apple Silicon) and x86_64 (Intel).

Supported architectures

Prerequisites

• Xcode Command Line Tools: xcode-select --install
• For x86_64 cross-compilation on arm64: rustup target add x86_64-apple-darwin

Build with Tauri

cd tauri-examples/hello_world
npm install
npm run build

Build with default (static) linking

cargo build --release

This downloads the matching static archive and produces a self-contained binary. The static build links against libc++ and the Foundation framework.

Build with shared linking

cargo build --release --no-default-features --features shared

This downloads the matching shared archive. The build script:

• sets rpath to @loader_path so libsherpa-onnx-c-api.dylib and libonnxruntime.dylib are found next to the binary at runtime
• copies the .dylib files next to the generated binaries and examples

Cross-compile for x86_64 from arm64

On an Apple Silicon Mac, you can cross-compile for Intel:

rustup target add x86_64-apple-darwin

# Static
cargo build --target x86_64-apple-darwin --release

# Shared
cargo build --target x86_64-apple-darwin --release \
  --no-default-features --features shared

Build from source

mkdir build && cd build
cmake \
  -DCMAKE_OSX_ARCHITECTURES="arm64" \
  -DBUILD_SHARED_LIBS=ON \
  -DSHERPA_ONNX_ENABLE_PORTAUDIO=OFF \
  -DSHERPA_ONNX_ENABLE_WEBSOCKET=OFF \
  -DBUILD_ESPEAK_NG_EXE=OFF \
  -DSHERPA_ONNX_ENABLE_BINARY=OFF \
  -DCMAKE_INSTALL_PREFIX=./install \
  ..
cmake --build . --target install --config Release

export SHERPA_ONNX_LIB_DIR=$PWD/install/lib
cargo build --release --no-default-features --features shared

Use -DBUILD_SHARED_LIBS=OFF for static linking instead. For a universal (arm64 + x86_64) build, use -DCMAKE_OSX_ARCHITECTURES="arm64;x86_64".

Use your own libraries

export SHERPA_ONNX_LIB_DIR=/path/to/lib
cargo build --release

Or point to a directory containing the release archive:

export SHERPA_ONNX_ARCHIVE_DIR=/path/to/archives
cargo build --release

Notes

• The shared archives include both libsherpa-onnx-c-api.dylib and libonnxruntime.dylib. Both are required at runtime.
• For static builds, the binary is self-contained with no runtime .dylib dependencies (aside from system frameworks).
• The @loader_path rpath ensures shared libraries are found relative to the binary, so no DYLD_LIBRARY_PATH setup is needed in most cases.