Privacy Policy

01Who runs this site

Auslan / English is a research showcase maintained by Preet Prasad for an Honours project at the University of Western Australia, supervised by Dr Atif Mansoor. This policy covers this project website; it is not a privacy policy for the University as a whole.

For privacy questions or requests, email 23857493@student.uwa.edu.au.

02When you browse

The walkthrough loads published video clips and saved model results from this site. It does not ask for camera or microphone access; only the translator asks for the camera, and only when you choose to enable it. There are no accounts, advertising trackers or application analytics in this release, and the application sets no cookies or visitor identifiers. The translator stores only the model files and the download choice described below.

Vercel hosts the website. Delivering and protecting the service involves technical request information, such as your IP address, browser or device information, requested URLs and request times. Hosting records are separate from the translation examples and are governed by Vercel’s Privacy Notice. This project does not set a separate retention period for those provider records. Vercel may process information in the United States and other countries.

Your browser may cache public page assets and clips. External links open services with their own privacy practices.

03When you use live translation

On this site the translator runs in your browser. Nothing is downloaded until you select Download and start in the translator; before that, the translator page contacts only this site.

• Downloads: the model files come from Hugging Face (huggingface.co and its download hosts under hf.co) and the ONNX Runtime Web runtime from jsDelivr (cdn.jsdelivr.net). As with any download, those providers receive your IP address, browser user agent and the addresses of the requested files, under their own privacy policies. The first download is 401.7 MB for the default single model; the six-model ensemble adds 391.7 MB and is downloaded only if you choose it. Every file is checked against a published SHA-256 fingerprint before use.
• Stored in your browser: the checked files are kept in this site’s Cache Storage so later visits need not download them again, and your choice to download is remembered in one local-storage entry (auslan-live-consent-v1) that holds no identifier. On later visits the translator fetches the small model list (manifest) from Hugging Face again and loads the models from browser storage. Pose estimation speed is checked after the download; if this device is too slow, the downloaded files are removed again automatically.
• Camera: access starts only after you choose to enable it. The microphone is not requested. During a signed recording, each video frame is converted to an image inside the page and processed by a background worker in the same browser tab, which extracts movement coordinates and, after Finish, produces the English text on your device. Camera frames, coordinates and translations are not sent to this site, Hugging Face, jsDelivr or anyone else.
• Processing and cleanup: frames, coordinates and timing information are held in memory. The single model keeps only the coordinates; the six-model ensemble keeps the recorded frames until its image features have been computed after Finish. Capture buffers are released after completion or cancellation, or when you leave the page, except as described under People in view. The application does not save recordings or add them to training data.
• People in view: the translator follows everyone it detects, using their movement coordinates (not images), and names them Person A, Person B and so on, so that the result can say who was translated. It does not use faces or appearance to tell people apart: people are matched from frame to frame, and from one utterance to the next, only by where their shoulders are, so labels can change if people move or swap seats. This happens even when only one person is in view: each person’s label and last shoulder position are kept in this browser tab’s memory until you turn the camera off, select Clear conversation or close the page, so that the translator can tell later that someone new has appeared. When two or more people were in view, that utterance’s movement coordinates (and, for the six-model ensemble, its recorded frames) are also kept until your next utterance, Clear, turning the camera off or closing the page, so that you can switch to another person; with one person in view, nothing from the utterance is kept after Finish. Everyone in view is processed, so please ask other people in the room first. Nothing is saved or sent anywhere.
• Results: the current result and recent conversation history remain in the open browser page. They are not written to an application database or browser storage. Reloading or closing the page removes this application state. Copying text creates a separate copy on your clipboard.

The software also supports a separate, server-backed research deployment (the project’s Python translation service), which is not used by the public Vercel site. In that mode, during a signed recording, video frames and capture times are sent to the connected translation server to extract movement coordinates and produce English text. The server holds frames, coordinates and timing information in memory; image-based models keep frames until image features have been computed. Capture buffers are released after completion, cancellation or disconnection, although a computation already running may finish before its buffers can be released. Temporary session identifiers, connection information and rate-limit counters support operation and abuse prevention. Results remain in the open browser page as described above.

Hosting and server infrastructure may keep technical or error logs. The application is designed to omit session content from its error messages; this is not a claim that all infrastructure logging is disabled.

04Your controls

In the translator, Finish or Cancel stops the current capture. Turning the camera off stops the live feed. Clear removes the current result; Clear conversation removes the visible history. You can also revoke camera permission in your browser.

Remove downloaded models (under Session details in the translator) deletes the stored model files and forgets your download choice; the translator then asks again before any new download. Clearing this site’s data in your browser settings has the same effect. Your browser’s ordinary HTTP cache may also hold copies of downloaded files under its usual rules; clearing browsing data removes them.

Speech is optional. The application uses only voices reported by the browser as local to your device; it does not send text to a project-operated speech service. If no local English voice is available, speech remains unavailable. Device and browser services have their own privacy settings.

05Enquiries and requests

If you email the project, your address, message and any information you choose to include are processed through the University email service to respond to you. Email is separate from the website’s temporary translation session, and has no project-configured automatic deletion schedule. Please avoid sending signing videos or sensitive personal information unless we have first agreed how to handle them.

You can ask what information the project holds about you, request access, correction or deletion, or raise a privacy concern by emailing the address above. We will explain what can be located and what action is possible; we cannot retrieve a translation session that was never saved. Provider-held records may require a request to that provider.

This policy will be revised when the public service’s data handling changes. The date at the top identifies this version.