Security and Deployment

This page explains where Hecttor SDKs run, what they need from the host environment, and how to think about data handling during integration.

Where processing happens

Both Hermes and Orpheus process audio locally, inside your application. Audio chunks are passed to the SDK and the enhanced result is returned in-process. Audio data does not leave your environment during processing.

This is true for evaluation and production deployments alike. The unit of work (a single chunk) never leaves the host running your code.

What the SDK needs from the host

A typical integration needs:

• A supported platform. Linux x86_64, macOS (Intel and Apple Silicon), and Windows x86_64 are supported across the SDK language bindings. Specific runtime and version requirements depend on the binding you choose; other targets are available on request.
• Compute. Both SDKs are CPU-friendly for real-time use; specific model choice affects CPU and memory footprint.
• An SDK key. Issued with your license. Loaded by your application at startup.
• Configuration matching your audio. Sample rate, chunk size, and (for Hermes) which functionalities you want enabled.

Data handling

Connectivity, telemetry, and data-handling behavior depend on the license and deployment configuration being evaluated. The right setup for your team depends on:

• Whether your environment is allowed to make outbound network calls at all.
• What internal compliance commitments apply to audio data.
• Whether you need fully air-gapped operation, a self-hosted control plane, or a managed setup.

We discuss these requirements during onboarding and configure the SDK accordingly. If you have specific deployment constraints (no outbound network, on-prem only, regulated industry), tell us up front so we can scope the right configuration before you start integrating.

Operational guidance

• Treat the SDK as a real-time component. Process chunks on a dedicated audio thread or task; don't block the audio path with unrelated work.
• Manage initialization once. Initialization is the costly step (model loading, key validation). Re-use a single enhancer instance across the lifetime of a session.
• Handle initialization failure explicitly. Both SDKs surface initialization failure through whatever idiom matches your binding (return pair, result type, or thrown error). Catch it and report instead of silently degrading.
• Surface processing errors. Per-chunk processing surfaces errors the same way; never assume the enhanced buffer is valid without checking the error channel for your binding.

For platform support and compliance posture matching your environment, contact us.