Orator

Web Privacy Notice

Updated 29 October 2026

This notice describes the Orator web app at orator.org.uk. It updates the web-specific information in the general Orator privacy policy, particularly the hosting and speech-analysis sections. The iOS app may use a different service configuration.

Information used by the web app

Your account includes your username, email, optional name, password hash and points. Practice requests can include audio, a transcript, word timings, speech context, practice mode and presentation text. Saved practice records include generated feedback, scores, selected speech excerpts and progress information.

For the new emphasis estimate, the browser also measures pitch, recording level, voiced-audio coverage and clipping in timed sections of your recording. These measurements are associated with individual words. They are not voiceprints and are not used to identify you.

What happens when you request feedback

  1. Your browser requests microphone permission when you start recording. You can instead select an audio file. Recording alone does not submit it for AI analysis.
  2. After you agree to AI processing and request feedback, your browser uploads the recording over HTTPS to private Vercel Blob storage. Orator retrieves it for transcription by OpenAI and deletes the stored upload before transcription begins. Uploading PowerPoint slides uses the same private storage before text extraction.
  3. Where the browser can decode the recording, it calculates word-level pitch and recording-level measurements locally. It sends those measurements and the transcript to Orator's server, which checks them against the word timings and calculates relative emphasis evidence.
  4. OpenAI receives the transcript, context, practice mode, relevant presentation text and derived evidence to generate coaching. The language model does not listen directly to the audio in this step.
  5. For signed-in users, Orator saves the resulting feedback and scores. This includes an emphasis summary and up to six highlighted words with their timings, relative emphasis estimates and available pitch and recording-level measurements. Guest feedback is returned to the browser without creating an account or a practice-history record.

Recording level is measured in dBFS, not calibrated real-world loudness. Results can be affected by noise, music, microphone processing, the speaker's range and transcription errors. Missing or unreliable measurements fall back to timing-only feedback. These estimates do not establish emotion, confidence, pronunciation, personality or objective speaking ability.

Hosting and service providers

Vercel hosts this web app, its API and private temporary file storage. Account, session and practice records are stored in the existing Neon-hosted PostgreSQL database. These providers and their infrastructure providers process the information needed to run and secure the service; processing may take place outside the UK.

OpenAI processes audio and coaching requests as described above. Its API data use and retention depend on the endpoint and account settings; see OpenAI's data controls. The support and general policy pages are hosted separately on Vercel. See also Neon's privacy policy and Vercel's privacy policy.

Storage and deletion

Guest mode does not require a username, email or password. It creates a random guest identifier in a server-side session and uses the same HTTP-only cookie as account sessions, expiring after 24 hours. Guest audio, transcripts, context and measurements still pass through the processing services described above; guest mode is not anonymous to those providers. Guest feedback remains in the current page's memory until you leave, reload or start a new session. Downloaded feedback remains on your device.

To limit abuse and AI costs, guest requests are counted per network and across the service. The network counter uses a secret-keyed hash of the IP address that changes each UTC day, not a raw IP address in the counter record. These counter records expire at the next UTC midnight and are removed by session-store cleanup; removal can be delayed if cleanup fails. Hosting and security logs may separately contain IP addresses under the providers' retention policies. Creating another guest session does not reset the daily allowance.

The browser keeps the current recording, transcript and measurements in memory for the session, including retrying a failed analysis. Leaving the practice screen releases this state; it does not delete an original file you selected from your device or a feedback download you saved.

Private upload access is limited by short-lived upload permissions and an authorization bound to the account or guest session. Orator removes a private stored upload after it has downloaded it for processing. The API removes its working audio copy after transcription and its PowerPoint copy after text extraction, and attempts cleanup on failures. If a browser closes, an upload is abandoned or deletion fails, a daily cleanup task targets stored uploads older than 24 hours. Such uploads can normally remain for up to about 48 hours, or longer if cleanup fails or service limits interrupt the task. This is not a guarantee of instant removal from provider logs, temporary infrastructure copies or backups.

Account details and saved practice feedback remain until account deletion; there is no automatic expiry for practice history in the app. Profile & Progress contains Delete account, which requires your password and removes the account, practice records, progress, matching support-form records and active web sessions from the live database. Provider logs and backups follow their own retention schedules.

Cookies, consent and security

The web app uses an HTTP-only session cookie with a 24-hour expiry and secure transport in production. It stores your AI-processing choice in local browser storage, separately for your account. Guest consent is held only in the current practice page and is requested again on a new visit. The new measurement workflow asks for agreement again. Turning AI processing off in Profile prevents new audio submissions until you agree again; it does not recall requests already sent or delete saved feedback.

Passwords are hashed, and production connections use HTTPS. Operational records may include request paths, status codes, timestamps, errors, upload information and security-related network information. Do not include unnecessary sensitive information in recordings, context or uploaded documents. No online service can guarantee absolute security.

Questions and choices

You can decline microphone access or AI processing, review saved feedback, download a feedback report and delete your account. For questions about access, correction, deletion or other privacy rights, use the Orator support contact. Further information about rights and the developer's responsibilities appears in the general policy linked above.

Return to Orator