Your recording and the model travel in different directions
Transcribana reads the file you select on your device and performs speech recognition in your browser. It does not send the recording or extracted audio to a transcription server. You do not need an account to use the audio transcription workspace. A model is a separate set of recognition weights: downloading it gives the browser the machinery to process your recording locally.
Local processing does not mean zero network activity. The site downloads application code, runtime assets and model files. The larger Arabic model needs about 1.33 GB of weights. Those requests expose normal connection information to the services delivering the assets; they are different from uploading your media. Read the privacy policy for the site’s stated handling of connection data.
Verify the distinction with a short, non-sensitive file
Open your browser’s developer tools and the Network panel before selecting a file. Clear the request list, select a short test recording, then start transcription. Inspect request URLs, methods and payloads rather than assuming that every request is an upload. Model and runtime downloads can be large, especially on a first run. The selected filename, recording bytes and transcript should not appear in an outbound request body.
A quiet network panel alone is not an independent security certification. Browser extensions, a compromised device or modified application code sit outside this check. If you handle regulated or confidential material, use your organization’s approval process before selecting it. A browser-only design does not itself establish legal compliance.
Treat offline use as a session capability
A model already loaded in the current page can be reused without downloading it again. This does not promise that closing and reopening the site will work offline. Browser storage may be evicted, a different model may need new files, or the runtime may not be cached. Test offline reuse with a disposable recording before relying on it during travel.
Do not clear site storage halfway through a job. Allow the first model load to finish and save the transcript before reloading. The Arabic workspace guidance explains the larger local download and desktop resource requirements.
Keep control of the files you create
Copy Text and TXT use the selected readable version; SRT, VTT and JSON use the original timestamped segments. A corrected name in the document can therefore still appear incorrectly in a subtitle export. Check what you are sharing, including quotations and personal details, before sending any downloaded file to someone else.
Downloads remain wherever your browser saves them. A local transcript can still be synchronized by your operating system’s backup or cloud-drive settings. Transcribana does not control those folders. Keep the original recording until review is complete, then follow your own retention policy. For the end-to-end workflow, see how to turn a local recording into a reviewed transcript.