What happens to the words

A session is a talk, typed out and translated while it happens. This page says where the audio and the text go, and when they are gone.

Sound

The speaker's microphone audio goes from the speaker's browser straight to Deepgram, the service that turns speech into text. It does not pass through our server and we do not record or store it.

To correct a sentence, the speaker's browser sends the sound of that sentence to our server, which passes it to Deepgram and to Microsoft's speech service for a second reading. The browser keeps the last minute of sound in memory for this; our server does not store it.

Text

The sentences Deepgram returns are stored on our server for as long as the session runs, so listeners can read them. Fifteen minutes after the session ends, the session, its transcript and its cached translations are deleted from our server. Our logs do not contain the text.

A few seconds after a sentence is shown, its text, the three sentences before it and the session's list of names are sent to Claude, an AI model run by Anthropic, which corrects punctuation and words the speech service misheard. The corrected sentence replaces the first one on every screen. A word Claude adds because the speech service dropped it is marked as added. When the organiser asks for a clean transcript, the text of the session is sent to Claude once more, to be put into paragraphs.

When a listener reads in a language other than the one being spoken, each sentence is sent to Google Cloud Translation to be translated.

Listeners

Joining needs no account, name or email. We count how many browsers joined a session and nothing else about them. The page remembers a few things in the listener's own browser: the language they picked, the light or dark theme, that feedback was already sent for a session, their answers to polls, and a random number that lets a poll count each phone once.

While the moderator has questions open, a listener can send a written question, with a name if they choose to give one. Only the moderator sees it. A question written in another language than the one being spoken is sent to Google Cloud Translation. Questions are deleted with the session.

A listener can also ask to speak. If the moderator gives them the floor, the sound of their phone's microphone goes straight from the phone to the speaker's laptop and is played in the room. It does not pass through our server and is not recorded. If they give a name, their words appear under it for everyone in the session; otherwise under “Audience”. While the speaker's microphone is on, that voice is sent to Deepgram to be written down, like the speaker's own. To set up the direct link, both devices ask a public address service run by Google (a STUN server), which sees their network addresses.

When the session ends, or earlier if the moderator opens it, a listener can leave a rating from 1 to 5 and a comment, without a name. A comment in another language than the one spoken is sent to Google Cloud Translation. The organiser can download the feedback; on our server it is deleted with the session.

When the moderator launches a poll, a listener can pick one answer. We store the answer next to that random number, not next to a name, and delete both with the session. Poll questions are translated by Google Cloud Translation for listeners who read in another language.

Visits

On the first page, the room guide and this page we use Google Analytics to see how many people visit. It starts only if you accept it in the bar at the bottom of the page, and “Cookie settings” in the footer lets you change your mind. It is the same choice as on socialit.ro: answer once and both remember it. The pages of a session, where the speaker talks and listeners read, do not load Analytics at all.

Other services

• Deepgram receives the speaker's audio and returns text.
• Microsoft (Azure Speech) receives the sound of single sentences, for a second reading, and returns text.
• Anthropic (Claude) receives sentences as text, to correct them and to write the clean transcript. It receives no sound. If the speaker pastes notes to fill the list of names and terms, Claude reads them once to pick the terms; the notes are not stored.
• Google Cloud Translation receives sentences to translate.
• Google Analytics counts visits to the pages named above, only after you accept.
• Google Fonts serves the typefaces, so it sees the address of each browser that loads a page.
• QR Server (api.qrserver.com) draws the QR codes, so it receives the session's join link.

These services run on their own infrastructure, which may be outside the European Union. Their own terms govern what they keep and for how long.

Who can start a session

Someone holding an invite link: a link starts one session of a fixed length and cannot start a second.

Anyone can also start a short public demo from the first page. It runs for three minutes, takes up to five listeners, and its text is deleted two minutes after it ends instead of fifteen. To keep the demo from being overused we count how many each network address starts in a day. For that we keep the time of each demo next to a scrambled form of the address, from which the address cannot be read back, for 24 hours.