# SDKs and models

Compare network requirements and data flows for transcription and translation, then choose compatible SDKs and models.

## Where transcription and translation run [#where-transcription-and-translation-run]

| Feature | What it needs | Where the data goes |
| --- | --- | --- |
| Offline transcription | A compatible device, the recognition SDK, a matching model and a valid license. Once installed and licensed, speech recognition can run offline. | The local engine reads the audio and returns a transcript to the app on the device. |
| Offline translation | A model for the required language direction and a valid license. Once installed and licensed, translation can run offline. | The app sends the source text to the model on the device and receives the translation locally. |
| Cloud translation API | A network connection from your server, an API key for an authorized project and a configured inference service. | The platform forwards the source text to the inference service and returns the translation to your server. |

When choosing speech packs for ObtuneAI, check the [offline language directory](/en/docs/languages?feature=offline). When integrating an SDK, check the release package for supported devices and language directions. See [offline licensing](/en/docs/license) for license validity checks.

The table follows the data for a single recognition or translation request. Check the app settings and release documentation to see whether the app also syncs recordings, transcripts or translations. For cloud integrations, see the [API reference](/en/docs/api) and [data and privacy notes](/en/docs/privacy).

## Platform delivery [#platform-delivery]

Android, iOS, and C++ are planned integration targets. The resource list determines which platforms are currently available for download. Each release should include installation steps, an API reference, a sample project, dependency declarations, and a compatibility matrix.

## ASR integration checks [#asr-integration-checks]

Confirm the sample rate, channel count, encoding, streaming chunk size, threading model, and result callbacks. Test initialization, cancellation, resource cleanup, and error paths on the target device.

## Translation integration checks [#translation-integration-checks]

Confirm language directions, model versions, text-length limits, and how models are loaded and unloaded. Initially, availability is determined by the model list actually released.

## Versions and integrity [#versions-and-integrity]

SDKs and models are not necessarily interchangeable. Verify SHA-256 checksums after downloading and select matching models according to the release notes. Validate device performance and license behavior before a production release.

## Not yet included [#not-yet-included]

This repository contains the developer platform, not an actual ASR or NMT engine. Specific SDK APIs, language coverage, package sizes, processing times, and hardware-acceleration capabilities must be documented with the real delivery packages.
