Anira Documentation#

anira Logo#

build_test codecov build_benchmark build_examples build_sanitizer build_docs on_tag


Anira is a high-performance library designed to enable easy real-time safe integration of neural network inference within audio applications. Compatible with five inference engines, ONNX Runtime, LibTorch, LiteRT, TensorFlow Lite and ExecuTorch, and with engines of your own, anira bridges the gap between advanced neural network architectures and real-time audio processing. In the paper you can find more information about the architecture and the design decisions of anira, as well as extensive performance evaluations with the built-in benchmarking capabilities.

Features#

  • Real-time Safe Execution: Ensures deterministic runtimes suitable for real-time audio applications

  • Thread Pool Management: One inference thread pool per process, shared by every handler, avoids oversubscription and runs inferences in parallel

  • Minimal Latency: Designed to minimize latency while maintaining real-time safety

  • Built-in Benchmarking: Includes tools for evaluating the real-time performance of neural networks

  • Comprehensive Inference Engine Support: Integrates ONNX Runtime, LibTorch, LiteRT, TensorFlow Lite and ExecuTorch, and runs engines you register yourself

  • Flexible Neural Network Integration: Supports a variety of neural network models, including stateful and stateless models

  • Cross-Platform Compatibility: Works seamlessly on macOS, Linux, Windows, Android and iOS, and in the browser via WebAssembly

Usage#

anira takes its configuration from two JSON files that travel with the model: the model file names the model’s exports per inference engine (paths relative to the file) and describes the tensors, the contract file says how it runs: the inference budget and the warm-up. By default the warm-up pass at prepare measures the budget on the machine; "budget": { "ms": 5.33 } pins it for a reproducible latency. The host blocks are patched in at prepare.

{
  "models": [ { "engine": "onnxruntime", "path": "model.onnx" } ],
  "inputs":  [ { "name": "audio_in",  "axes": [["time", 256], ["channel", 1], ["any", 1]] } ],
  "outputs": [ { "name": "audio_out", "axes": [["time", 256], ["any", 1]] } ]
}
{ "hard": { "budget": "measured", "warmup": "until_stable" } }

The basic usage of anira is as follows:

#include <anira/anira.hpp>

// The configuration: the model file next to your model, and the contract file
anira::ModelConfig model_config = anira::ModelConfig::from_file("path/to/model.json");
anira::ContractHandle contract = anira::ContractHandle::from_file("path/to/contract.json");

// The context (the inference thread pool) and the handler over the model's inference stage,
// on every engine of this build the model file names
anira::Context context{anira::ContextConfig{}};
anira::Pipeline pipeline;
pipeline.add_inference(model_config);
anira::InferenceHandler inference_handler(context, pipeline);

// Complete the contract with the host blocks; prepare allocates, warms the model up, measures
// it and returns the plans, one per engine the model runs on here
contract.hard_host_blocks(buffer_size, buffer_size, sample_rate);
const anira::PlanReport report = inference_handler.prepare(contract);

// Whether the running plan keeps up on this machine (hold a NOT_REALTIME one and ask the user)
const anira_plan_info row = report.plans()[inference_handler.get_plan()];
const bool realtime = static_cast<anira_rt_verdict>(row.rt_verdict) != ANIRA_RT_NOT_REALTIME;

// Optionally select another plan (the default engine's plan runs first, plan 0 without one) and
// get the latency
inference_handler.set_plan(0);
uint32_t latency_in_samples = inference_handler.get_latency();

// The audio block as a tensor over the host's channel pointers, built once at prepare
std::array<float*, 2> channels{};
anira::Tensor audio = anira::Tensor::from_host_planar<float>(
    std::span<float* const>(channels), std::array<int64_t, 2>{2, buffer_size});

// Real-time safe audio processing in process callback of your application
void process(float* const* audio_data, size_t num_samples) noexcept {
    channels = {audio_data[0], audio_data[1]};
    audio.shape[1] = static_cast<int64_t>(num_samples);
    inference_handler.process(audio, 0, audio, 0);
}
// audio_data now contains the processed audio samples

Installation#

C++ Library#

Anira can be easily integrated into your CMake project. You can either add anira as a submodule, download the pre-built binaries from the releases page, or build from source.

Option 2: Use Pre-built Binaries#

Download pre-built binaries from the releases page.

In your CMakeLists.txt:

# Setup your project and target
project(your_project)
add_executable(your_target main.cpp ...)

# Add the path to the anira library as cmake prefix path and find the package
list(APPEND CMAKE_PREFIX_PATH "path/to/anira")
find_package(anira REQUIRED)

# Link your target to the anira library
target_link_libraries(your_target anira::anira)

# Only if your own code calls an engine: link its anira::<engine> target too
target_link_libraries(your_target anira::onnxruntime)

anira::anira carries anira’s headers but no engine header: the engines are linked privately. Inside anira’s build tree it also carries the USE_<ENGINE> definitions the bundled examples select on; the installed anira::anira does not, and code that needs to know which engines a build compiled in asks anira_enabled_engines (anira::enabled_engines()). A target that calls an engine itself links the matching engine target — anira::onnxruntime, anira::tflite, anira::litert, anira::libtorch or anira::executorch — which exists both in the build tree and in the installed package, carries the engine’s headers, and is the very file anira links, so the process never holds a second copy of the engine.

Option 3: Build from Source#

git clone https://github.com/anira-project/anira.git
cd anira
cmake . -B build -DCMAKE_BUILD_TYPE=Release -DANIRA_WITH_INSTALL=ON
cmake --build build --config Release --target anira
cmake --install build --prefix /path/to/install/directory

Or with the bundled presets (desktop-install-release / desktop-install-debug):

cmake --preset desktop-install-release
cmake --build --preset desktop-install-release
cmake --install build/desktop/Install/Release --prefix /path/to/install/directory

Without --prefix, the tree is installed to <build dir>/anira-<version>.

The Apache-2.0 core component of tanh-lib that backs anira’s buffers is built into the library and installs nothing of its own: find_package(anira) looks up Threads and the engines alone, so consumers only add the anira prefix to CMAKE_PREFIX_PATH.

C++ Build Options#

By default, LibTorch, ONNXRuntime, LiteRT and ExecuTorch are enabled. You can disable specific engines as needed:

  • LibTorch: -DANIRA_WITH_LIBTORCH=OFF

  • OnnxRuntime: -DANIRA_WITH_ONNXRUNTIME=OFF

  • LiteRT (LiteRt* C API): -DANIRA_WITH_LITERT=OFF — runs .tflite models through LiteRT’s native CompiledModel runtime. Enabled by default; it is the modern TensorFlow-Lite-family engine.

  • TensorFlow Lite (legacy TfLite* C API): -DANIRA_WITH_TFLITE=ON — the same runtime as LiteRT exposed through the older C API, so the two are mutually exclusive. To use it, disable LiteRT: -DANIRA_WITH_LITERT=OFF -DANIRA_WITH_TFLITE=ON.

  • ExecuTorch: -DANIRA_WITH_EXECUTORCH=OFF — runs .pte programs exported ahead-of-time with torch.export; PyTorch’s edge/mobile inference stack (CPU execution via XNNPACK). Enabled by default; static-only.

Platform / engine support#

anira builds on the targets below; the pre-built engines it downloads ship per target as shared and/or static. Engine linkage follows BUILD_SHARED_LIBS: a shared anira links shared engines, a static anira links static engines, and an engine that does not ship the required linkage is disabled with a warning:

Target

LibTorch

ONNXRuntime

LiteRT

TFLite (legacy)

ExecuTorch

macOS x86_64

shared

shared · static

shared · static

shared · static

static

macOS arm64

shared

shared · static

shared · static

shared · static

static

macOS universal

shared

shared · static

shared · static

shared · static

static

Linux x86_64

shared

shared · static

shared · static

shared · static

static

Linux aarch64

shared

shared · static

shared · static

shared · static

static

Windows x86_64

shared

shared · static

shared · static

shared · static

static

Windows arm64

shared

shared · static

shared · static

shared · static

static

WASM (Emscripten)

—

static

—

—

—

LibTorch is shared-only (disabled in a static anira build); ExecuTorch is static-only (disabled in a shared anira build). iOS and WebAssembly are static-only altogether (-DBUILD_SHARED_LIBS=OFF). LiteRT and TFLite are the same runtime via two C APIs and are mutually exclusive (LiteRT is the default). In a fully static anira build ExecuTorch cannot be combined with LiteRT or TFLite (each bundles its own copy of XNNPACK, whose symbols collide in one static image) — ExecuTorch is then auto-disabled; disable LiteRT/TFLite to build a static anira with ExecuTorch. On WebAssembly only ONNX Runtime is supported. Backends for Android and iOS are also published in the anira-project/backends release for cross-builds. — = not provided.

Pre-built engine binaries are downloaded at configure time from the anira-project/backends release pinned by ANIRA_BACKENDS_VERSION. Integrity is checked live: when GitHub is reachable, anira fetches each asset’s published SHA256 and re-downloads any engine whose archive changed upstream or downloaded incompletely (the download is verified against that hash). Nothing is pinned in-repo. The source is configurable (the linkage is not: it follows BUILD_SHARED_LIBS, see above):

  • Backends release tag: -DANIRA_BACKENDS_VERSION=v2.4.0 (the built-in default).

  • Offline / reproducible builds: -DANIRA_BACKENDS_SKIP_REMOTE_CHECK=ON skips the GitHub query and reuses whatever is already in modules/.

  • Bring your own engine (no fork): -DANIRA_<ENGINE>_ROOTDIR=/path/to/prebuilt (a tree with include/ + lib/), or a custom source via -DANIRA_<ENGINE>_URL=... -DANIRA_<ENGINE>_SHA256=....

Moreover, the following options are available:

  • Build the benchmark examples and their fixture (examples/benchmark/, not part of the library): -DANIRA_WITH_BENCHMARK=ON

  • Build example applications, plugins and populate example neural models: -DANIRA_WITH_EXAMPLES=ON

  • Build anira with tests: -DANIRA_WITH_TESTS=ON

  • Build anira with documentation: -DANIRA_WITH_DOCS=ON

  • Disable the logging system (compiles every anira log call out; anira otherwise logs through tanh-lib’s thl::Logger): -DANIRA_WITH_LOGGING=OFF

Anira Web (Web / JavaScript)#

Anira is available as the @anira-project/anira package for web applications: anira compiled to WebAssembly with a thin typed layer over its C ABI. anira starts the page’s threads from C on one shared memory (the context’s inference threads are Wasm Workers, the audio path a Wasm Audio Worklet node whose render callback is C), processors and inference engines are written in TypeScript without rebuilding the module, and the model and contract files of a native host load unchanged. The 3.x pre-release is on the next dist-tag:

npm install @anira-project/anira@next

See the Anira Web guides.

Building @anira-project/anira from source#

If you want to build the WASM module and JavaScript bindings yourself, you need to provide your own Emscripten SDK (4.0.23, the version CI pins) and TypeScript on the PATH (emcc --emit-tsd runs tsc). The CMake presets expect the EMSDK environment variable to be set to the root of your emsdk installation.

git clone https://github.com/anira-project/anira.git
cd anira

export EMSDK=/path/to/your/emsdk

# Configure and build the WASM module (release); wasm-abi adds the node ABI gates
cmake --preset wasm-release
cmake --build --preset wasm-release

# Build the JavaScript package
cd web
npm install
npm run build

For packaging it locally, use

npm pack

in the web folder, which will create a .tgz file that can be installed with npm or yarn.

Then install the package in your project:

npm install path/to/anira/web/anira-project-anira-x.x.x.tgz

A debug preset is also available via cmake --preset wasm-debug / cmake --build --preset wasm-debug.

Examples#

Bundled examples#

  • Simple JUCE Audio Plugin: Demonstrates how to use anira in a real-time JUCE / VST3 plugin, with a “Try nonetheless” button for a backend too slow for the machine.

  • CLAP Plugin Example: Demonstrates how to use anira in a real-time CLAP plugin, with a “Try nonetheless” parameter.

  • Benchmark: Demonstrates how to use anira for benchmarking of different neural network models, backends and audio configurations, with the Google Benchmark fixture of examples/benchmark/fixture/ to copy into a project of your own.

  • Minimal Inference: Demonstrates how minimal inference applications can be implemented in every engine.

  • Web examples: A model file in the browser, and a processor and an engine written in TypeScript.

Other examples#

  • nn-inference-template: A more elaborate JUCE / VST3 plugin that uses anira for real-time safe neural network inference, with a more appealing GUI than the simple JUCE example.

Real-time safety#

anira’s real-time safety is checked in this repository with the rtsan sanitizer.

Citation#

If you use anira in your research or project, please cite either the paper or the software itself:

@inproceedings{ackvaschulz2024anira,
    author={Ackva, Valentin and Schulz, Fares},
    booktitle={2024 IEEE 5th International Symposium on the Internet of Sounds (IS2)},
    title={ANIRA: An Architecture for Neural Network Inference in Real-Time Audio Applications},
    year={2024},
    volume={},
    number={},
    pages={1-10},
    publisher={IEEE},
    doi={10.1109/IS262782.2024.10704099}
}

@software{ackvaschulz2024anira,
    author = {Valentin Ackva and Fares Schulz},
    title = {anira: an architecture for neural network inference in real-time audio application},
    url = {https://github.com/anira-project/anira},
    version = {x.x.x},
    year = {2024},
}

Contributors#

License#

This project is licensed under Apache-2.0.

Next Steps#