Trust, plainly
A rehearsal recording is personal. This page says what happens to yours in plain language: what is analyzed, what never is, who processes it, and the rules the product is built around.
What is analyzed, and what never is
Only the audio of a take is analyzed: what you said and how it sounded. Nothing visual in your recording is processed, even when you record video. Appearance, gaze, posture, gesture, facial expression: none of it is read, in this version, at all.
Nothing runs without your consent
Before each analysis you see exactly what will happen: what is analyzed, which processors touch the audio, and how long it is kept. This isn’t a banner to click away; the pipeline refuses to start without that consent, and the refusal is enforced where the data lives, not just in the interface.
Optional slides
When slide review is available, you can attach a PDF and separately agree to send it with your transcript to Anthropic Claude. The whole PDF is sent, including any images and charts on its pages. It is used to compare what the slides say with what you say, never to judge how the slides look, and never anything about you. Feedback compares content and structure, citing slide text and speech passages. It preserves your original slides and personal style. Slide files and slide feedback share your recording’s retention window and are removed when you delete the rehearsal.
Who processes your audio
| Service | Role |
|---|---|
| Supabase | Account, database, and media storage, hosted in the EU (Frankfurt). |
| Railway | The analysis worker that processes each take, hosted in the EU (west). |
| Google Speech-to-Text | Transcription of the audio, using Google’s EU endpoint. |
| Anthropic | The coaching analysis runs on Anthropic’s Claude models via API. |
| ElevenLabs | Voices the finished report as a spoken walkthrough, from the report’s text only, never the audio. Sessions that did not consent to it keep the browser’s voice. |
| OpenAI | Standby for the coaching analysis: used only if Anthropic’s API is unavailable, on the same transcript and measured signals, never the audio. Listed on the consent screen; jobs that did not consent to it never reach it. |
| Vercel | Hosts this website and the app you’re reading now. |
We name processors instead of making blanket residency promises: storage, queue, and the worker run on EU-hosted infrastructure, and the model analysis is an API call to Anthropic, or to OpenAI as the named standby when Anthropic is unavailable.
How this website is counted
Separate from everything above, which is about your audio: this site counts visits so we can see what to improve. Basic traffic counts are cookieless, carry no identifier, and are served from this domain, so they run for everyone. Google Analytics is a different thing: it sets cookies, so it is only ever loaded if you accept it in the notice at the foot of the page. Decline, or simply ignore it, and nothing is requested from Google at all.
You can change your mind here, whenever you like. Turning it off stops the script and deletes the cookies it set, so the identifier doesn’t outlive the permission that created it.
Neither one can see a rehearsal. Recordings, transcripts, and reports are never part of website analytics.
Your data, your window
- LiveYou choose the retention window for each session when you consent: 30 or 90 days.
- LiveEvery row and file is isolated to your account, and media is served only through short-lived signed links.
- LiveDeletion cascades through the data model: removing a session removes its media and every derived artifact with it.
- LiveSelf-serve delete with a receipt: every rehearsal has a delete control in your library and at the foot of its report. Files and records go immediately, the receipt exists only in your browser, and platform backups age out on the provider’s schedule. Full account deletion is not self-serve yet; the contact on the privacy page handles it.
Fairness rules the product is built around
- Moments, never you. Notes describe timestamped moments in a recording. Nothing scores or judges you as a speaker, and no overall person-score exists anywhere in the product.
- Accent is never a finding. The system doesn’t label, infer, or coach your accent, and “accent reduction” does not exist here. Pacing rubrics account for speaking in a second or third language.
- Measured and inferred are kept apart. Numbers like pace and pauses are reported as measurements. Anything interpretive is labeled as an inference, carries its uncertainty, and stays off by default.
- Accommodations are respected, never guessed. If you declare one, like recalibrated pause norms or a speech difference, the analysis rules adjust. The system never diagnoses or detects anything about you.
- It serves the speaker only. Reports are for you. There is no ranking of people, no comparison between users, and no screening or surveillance use: those are refused, not just unbuilt.
Honest status
The end-to-end loop is live and has produced a real evidence-checked report. The pilot is small, free, and invite-only on purpose. You’ll find no user counts, testimonials, or improvement percentages on this site because none have been measured yet; when they exist, they’ll come with their methodology.
This page is a plain-language description of how the product behaves. The formal version, with legal bases, processors, retention, and your rights, is the privacy policy.