Installation#
The ABR SDK is delivered in two parts: the programming interface for your language, a Python package you install with pip or a C binding you compile into your program, and an application package, which contains the neural network and compiled inference backend for your target device. You need both to run the SDK.
Prerequisites#
Python 3.10 or later is required.
A C99 compiler. The binding has no dependencies beyond libc.
Install the SDK#
# With pip
pip install abr-sdk
# Or add it to a uv project
uv add abr-sdk
The C binding is a source tarball you download from the ABR developer portal after logging in. Extract it and add it to your project:
tar xzf abr-sdk-c.tar.gz
Put abr-sdk-c/abr_sdk on your include path and compile the .c files in it along with your own
sources. The tarball’s README.md has the compiler and CMake invocations, examples/docs the C
examples this documentation shows, and LICENSE.md the terms that cover it. The binding has no
link-time dependency on an application package: abr_lib_open loads the package’s shared library
with dlopen when your program runs, and refuses one built for another ABI version (see
Compatibility).
Get an application package#
The neural network and the compiled inference backend are delivered in an application package, a
.tar.gz archive you download from the
ABR developer portal after logging in.
Each package is built for a specific combination of functionality (ASR or TTS), language, model variant, and hardware platform. Packaging them separately keeps the Python install small and lets you choose the package that fits your device.
A package filename encodes those dimensions. For example, niagara-38m-live.en-linux-x86_64.tar.gz
is an English ASR package for x86-64 Linux. Download the package that matches your target device and
use case. See Package naming for how to read each segment of a package name.
Extract the package#
Extract the package into a directory on the device that will run the SDK:
mkdir -p ~/abr-packages
tar xzf niagara-38m-live.en-linux-x86_64.tar.gz -C ~/abr-packages
The extracted directory contains a shared library, neural network weight files, and supporting configuration. Keep all of these files together in the same directory, the SDK loads the library and weights as a unit at runtime. Moving the shared library file out of the directory will break loading.
A package works with any SDK release built for the same ABI version; Compatibility says what that means and what to update when either side changes.
Verify#
Confirm the Python package installed correctly:
import abr_sdk
print(abr_sdk.__version__)
You should see the SDK version number. This checks only the Python package; it does not load the application package.
Build the asr example that ships in the tarball:
cd abr-sdk-c
cc -std=c99 -Iabr_sdk abr_sdk/*.c examples/asr.c examples/common/*.c -o asr
A clean compile checks only the binding; it does not load the application package.
Next steps
License activation: activate this device so the SDK can load your package, then ASR quickstart to produce your first transcript.