> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ai-coustics.com/llms.txt
> Use this file to discover all available pages before exploring further.

# SDK telemetry

> Separate local inference, authorization, usage reporting, error reporting and optional metrics export.

## Overview

The ai-coustics SDK processes audio locally. In core SDK 0.25.0, the authorization, usage and SDK error-reporting payloads described here do not include audio samples, recordings, transcripts or model input/output features.

Each network channel is controlled separately:

| Activity | Destination and data | Control |
| :- | :- | :- |
| Model provisioning | Model download service; model identifier and download request | Provision compatible model files before deployment, then load them from a local path. |
| Authorization and metering | ai-coustics backend; credentials, session identity, SDK/model/system metadata and usage | Governed by the credential and entitlement. Self-service keys use online authorization and usage reporting. |
| SDK error reporting | ai-coustics error tracking; error category and SDK/model/system/account metadata | Set `DO_NOT_TRACK=1` before loading the SDK to disable this channel. |
| OpenTelemetry metrics | Your configured collector; processing metrics and session attributes | Disabled by default; enable through `AIC_SDK_OTEL_ENABLE=1` or the binding's per-instance configuration. |

This description covers SDK traffic. Your application, speech-to-text provider and agent framework may send audio or other data independently.

## What we collect

Product and usage records include:

* Core SDK version and wrapper type, such as Python or Node.js
* Model identifier, such as `quail-vf-2.2-l-16khz`
* Operating system and CPU architecture
* Processing-session identity and processed or disallowed audio duration
* Whether the session uses a dedicated voice activity detection (VAD) model

## Authentication and metering

A session associates processing activity with the authorized account. Treat SDK keys, decoded API keys and bearer tokens as secrets.

Network failure behavior depends on the credential and session requirements. Core SDK 0.25.0 requires activation for JWT credentials; failed activation can disable processing. Explicit authorization rejection can also disable processing. Where usage reporting is required, sustained reporting failure can disable processing too.

Do not use a temporary grace period or tolerated failure as an offline deployment contract. For disconnected operation, contact ai-coustics about an appropriate offline license and verify its validity period and deployment conditions. Preprovision model files as well. An offline entitlement does not automatically prevent your application or an enabled customer metrics exporter from making network requests.

See [authentication](/models/get-started/authenticate-apps) for token handling and [deployment](/production/deployment) for restricted-network validation.

## Error reporting

SDK error reporting was added in [core SDK 0.24.0](/changelog). It covers failed session activations, failed usage reports and rejected bearer-token updates.

Reports contain the error class and message, core SDK version, wrapper type, model identifier, operating system and CPU architecture. Account identity is included when available and trusted. The reporting path excludes credential material and audio.

To disable SDK error reporting:

```bash theme={null}
export DO_NOT_TRACK=1
```

Set this before your process loads the SDK. The setting is read once per process. Licenses with an offline entitlement and WebAssembly builds do not use this error-reporting channel.

<Note>
  `DO_NOT_TRACK=1` does not turn off authorization or usage reporting, grant an offline entitlement or disable your OpenTelemetry exporter.
</Note>

## What we do not collect

The SDK reporting channels to ai-coustics described above do not send:

* Audio samples, recordings or buffered audio
* Transcripts or speech content
* Model input/output feature data

Customer-controlled [OpenTelemetry export](/models/get-started/opentelemetry) includes derived processing and audio metrics. Hashes and session IDs remain correlatable identifiers: apply access controls, retention limits and appropriate redaction in your own observability system.

## Network review and contractual sources

For a deployment with restricted outbound access, validate model provisioning, activation, usage reporting and optional exporters independently. Obtain the current destination requirements for your SDK release and entitlement from ai-coustics before finalizing firewall rules. This page is not an exhaustive hostname allowlist.

The [Privacy Policy](https://ai-coustics.com/legal/privacy-policy) and your applicable license agreement govern data and privacy terms. This technical reference does not establish retention periods, certifications, service levels or contractual offline rights.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.