Skip to main content

​
Overview

AirTen executes ai-coustics models locally on the CPU. Applications use it through the SDK; they do not need to integrate a separate inference framework such as ONNX Runtime, PyTorch or TensorFlow.

​
Characteristics

  • CPU inference: A GPU is not required for SDK audio processing.
  • SDK-managed runtime: Install the binding and its supported native library for your platform. See the language bindings.
  • Stateful processing: Each independent audio stream needs its own processing state. See streams and state.
Local inference describes where audio processing happens. Model downloads, authorization and telemetry have separate network requirements, described in SDK telemetry.

​
Model support

AirTen executes .aicmodel files. Match the model to the core SDK version using the model catalog and compatibility matrix. A language binding version and a core SDK version are different identifiers.

​
Performance

Processing time depends on the model, CPU, audio configuration and competing workload. A fixed algorithmic delay does not guarantee a fixed execution time or that every call finishes within its audio frame budget. Measure processing duration, memory use and concurrent-stream capacity on the hardware and resource limits you intend to deploy. Include resampling, application buffering and other pipeline stages in end-to-end measurements. Use the latency guide to interpret SDK delay and the performance guide to evaluate runtime cost. The deployment guide connects these measurements to readiness and overload handling.