Application packages

Application packages#

An application package is the .tar.gz archive you download from the ABR developer portal for your target device. It contains everything the SDK needs to run inference: the compiled neural network library, the model weights, and supporting configuration files. You extract it once onto the device and point the SDK at it when creating an Asr or Tts instance.

Package contents#

Every application package contains three things:

  • A compiled shared library (.so). The hardware-specific inference backend is already compiled into this library.

  • Neural network weight files. Binary data encoding the trained model parameters.

  • Supporting configuration files. Metadata the library reads during initialization.

The library, weights, and configuration files are a unit. After you extract the package, keep all of them in the same directory. The shared library expects to find the weight files beside it; moving the library file out of the directory will break initialization.

Platform specificity#

Each package targets a specific combination of capability, language, model variant, and hardware platform. For example, niagara-38m-live.en-linux-x86_64.tar.gz is an English-language ASR package for Linux on x86-64.

Available targets are Linux x86-64, Linux ARM64, and Android ARM64. The package you download determines which inference backend runs on your device. You do not configure or select backends in code.

Extracting a package#

Extract the package into a directory on the device where the SDK will run:

mkdir -p ~/abr-packages
tar xzf niagara-38m-live.en-linux-x86_64.tar.gz -C ~/abr-packages

The resulting directory contains the shared library and all supporting files. Do not move the shared library out of this directory after extraction.

For the full download-and-extract walkthrough, see Installation.

Pointing the SDK at the package#

Pass the path to the shared library as the first argument when you create an Asr or Tts instance:

from abr_sdk.asr import Asr

LIBRARY_PATH = "/home/user/abr-packages/niagara-38m-live.en-linux-x86_64/libniagara_38m_live.so"

with Asr(LIBRARY_PATH) as asr:
    ...

from abr_sdk.tts import Tts

# female voice
LIBRARY_PATH = "/home/user/abr-packages/nith-5m-live.en-f-quill-linux-x86_64/libnith_5m_live.so"

with Tts(LIBRARY_PATH) as tts:
    ...

An absolute path is the most predictable option. When you pass a relative path, the SDK searches for it in this order:

  1. Any directories you pass in lib_search_paths.

  2. The current working directory.

  3. Directories listed in LD_LIBRARY_PATH.

To skip the environment-variable directories and rely only on paths you supply, set use_default_lib_search_paths=False:

with Asr(
    "libniagara_38m_live.so",
    lib_search_paths=["/home/user/abr-packages/niagara-38m-live.en-linux-x86_64"],
    use_default_lib_search_paths=False,
) as asr:
    ...

The same lib_search_paths and use_default_lib_search_paths options apply to Tts.

Weights directory#

The SDK locates the weight files using the directory that contains the shared library. This default works when a package is extracted in its original layout. If you store the weights in a separate directory, pass its path with resources_dir:

with Asr(
    "/path/to/libniagara_38m_live.so",
    resources_dir="/path/to/weights",
) as asr:
    ...

with Tts(
    "/path/to/libnith_5m_live.so",
    resources_dir="/path/to/weights",
) as tts:
    ...

Next steps

  • Streaming and chunks: how Buffers, Events, and EventSets work underneath the high-level API.

  • SDK lifecycle: how object lifetimes work and when to call close().

  • Asr: full constructor parameter list.

  • Tts: full constructor parameter list.