An hour of speech, transcribed whole

new

Ramble is long-form stream-of-consciousness interaction: one recording of up to 60 minutes, transcribed whole. Crash-proof: a reload costs five seconds at most. Capture is raw: echo cancellation, noise suppression and automatic gain are requested off. The transcription service hears what your microphone heard.

Ramble is part of Free and Pro only. On Open, voice input is the standard browser microphone, which transcribes live as you speak (see Voice: talk to it).

Start a take

Four ways in, all opening the same panel:

Where	What you click
Desktop left nav	the Ramble icon
Mobile drawer	the Ramble menu item
Attachments menu	Ramble · Record audio and transcribe it
Mobile, when minimized	the red bubble, back into the running take

While a take runs the panel also carries a pause: Pause & mute - paused time is not recorded. The trash arms rather than acts: one tap turns it into Keep recording beside a red Confirm discard recording.

Crash-proof recording

The recording is saved every five seconds. An application reload, a closed browser or a dead battery costs at most the last five seconds. The next start notifies you of recovered takes, chipped recovered: Recovered from an interrupted recording - the tail may be missing. You can use them right away, and a recovered take always keeps its audio.

A soft tick every minute says the take is up and listening. The bell shown while recording is its toggle.

Recording in the background

Minimize the panel and move anywhere in the application - the take keeps recording and saving, on desktop and on mobile. Come back, stop it, and the transcript is ready.

While recording, the close button becomes Minimize. On desktop, the left nav carries a red face that pulses with your voice. On mobile it is a red bubble you can drag anywhere on the screen.

Transcription runs on API keys

Transcription needs an engine of your own: a Deepgram or an OpenAI key, set at Settings › Voice › Input. An OpenAI service you already added counts as one, as does a custom transcription endpoint. Audio goes from your browser straight to that AI service, and the text comes back. Big-AGI's servers are not in that path; there is no server-side transcription.

One engine transcribes a take - the one you pinned, otherwise the first with valid credentials - and a failure is not retried elsewhere.

The first transcript also triggers a title. It is written once, and a title you typed is never overwritten. Titling sends the transcript's opening and closing lines to a fast model. That is the one part of Ramble that uses your chat models.

What the engine can label

Words are the floor. Each engine has its own section at Settings › Voice › Input, most of it behind Advanced...:

Setting	What it adds
Smart Format	punctuation and paragraphs - on by default, Deepgram
Label Speakers	Speaker 0: prefixes through the transcript, one per voice
Label Topics	the topic chips on the card; hovering one highlights its span in the transcript - Deepgram
Label Sentiment	one chip - positive, neutral or negative - carrying the score in its tooltip, Deepgram
Personal Dictionary	up to 100 names, brands and jargon terms, sent with every transcription

Detect Topics in the Ramble menu asks for topics take by take, whatever the engine section says. An engine that cannot label them returns words only. Some models ignore the dictionary, and its label then reads Unused by this model.

Where the recording lives

Audio bytes stay in this browser's IndexedDB, on the machine that recorded them. They never reach a Big-AGI server, and they never sync. Everything else travels: the transcript with its language and topics, the title, dates and review state. The recording's description travels too - duration, format, capture time, device, and the chat it started from.

A ramble reaches your other devices once two things are true: a first non-empty transcript exists, and you are on Pro pro. A take with no words in it never leaves the device. A copy that arrived by sync can be read but not replayed.

Ramble options

The overflow menu holds five toggles:

Menu item	What it does	Default
Transcribe	Transcribes each take as soon as it stops.
Preserve Audio Locally	Keeps the audio file after a transcript lands; available while Transcribe is on.	off
Detect Topics	Asks for topic labels; Deepgram only.
Copy When Done	Puts each fresh transcript on the clipboard.	off
Show Size Bar	Draws each recording's relative size on its collapsed card.	off

The rest of the menu acts rather than toggles. Import Audio... transcribes an existing file the same way, taking its file name as the title. There is no length or size limit. Transcription Options opens Settings › Voice › Input. Remove All Audio names the megabytes it would free, and Delete All empties the library. While a take is running, Continue minimized joins the top of the list.

By default, audio is deleted from the device the moment a transcript lands. Silence, failed transcriptions and recovered takes keep their audio. Deleting a ramble deletes its audio immediately.

Once the audio is gone, the transcript can be edited but never re-transcribed. To re-transcribe local audio, clear the transcript in the editor: Transcribe re-arms for another pass.

Troubleshooting

What you see	What it means	What to do
Recording stopped: reached 60 minute cap	The hard limit. Everything up to it is saved.	Start the next take.
Recording stopped: reached 100 MB cap	The size limit. An hour at this quality is about 55 MB, so the time cap normally trips first.	Start the next take.
Recording interrupted: no audio received	The recorder stopped delivering data for 15 seconds - a suspended tab, or an input device that died. Silence does not trip this.	Check the input device, then record again.
Microphone disconnected	The input device went away mid-take.	Reconnect it; what was recorded is kept.
Audio channel busy	The audio channel was already in use.	Start the take again, or reload the application.
error, with the service's own message	The transcription request was refused.	Fix what the message names, then press Transcribe.
no speech	The transcription ran and found no speech.	Play the take back to check the input, then record again.

A take under a second is discarded - Recording under 1s - discarded - and dictation keeps anything over 0.3 seconds.

Related

• Talk to it instead of typing
• What syncs between devices
• What leaves the browser
• Open, Free, or Pro