Recordings and transcripts
Proofbell can record answered calls and turn them into searchable text. Both are optional, and neither happens until you switch it on.
Transcription is off until you turn it on, per website. Not off by accident — off deliberately, because transcription costs money per minute and nobody should be billed for a feature they did not ask for. Switch it on for a website on the Transcripts page — Admin menu, under Administration — and it applies to calls from then on; it does not go back and transcribe your history.
Only answered calls are transcribed. A missed call has no conversation to transcribe, and a ringing line is not billed.
What a transcript gives you
The text of the call, split into turns, with a timestamp on each. Open any call in your call log and the transcript opens beside it.
Speakers are numbered, never named. The transcriber can tell two voices apart but it does not know which one is the caller, and guessing from who spoke first is wrong often enough to matter. Labelling one "Caller" would put words in a member of the public's mouth in a document you may one day quote back to them, so they are "1" and "2" — less helpful and honest.
How much to trust it
Transcription is very good and it is not perfect, and the difference matters most on exactly the words you are reading a transcript to find — a name, a postcode, a price. So we show you where it was unsure rather than leaving you to assume.
- Individual words the transcriber was unsure of are underlined with a dotted line. Hover one and it says so. These are worth checking against the recording before you rely on them.
- A whole transcript that came out badly is flagged at the top. A bad line, background noise or two people talking over each other produces text that reads perfectly well and is substantially invented. When that happens we say so, rather than presenting a guess as a record.
Both exist because an average is a poor summary of a transcript: a call can score well overall while one misheard figure sits in the middle of it. A single score would hide precisely the word that mattered.
What is removed automatically
If somebody reads out a card number, we take it out of the transcript. This is not optional and there is no setting to keep it, deliberately — a switch to store unredacted text is something somebody turns off to debug a problem and forgets.
What is detected and removed:
- Card numbers
- Bank account numbers, sort codes and IBANs
- National Insurance numbers
- NHS numbers
- Email addresses
- Postcodes
Removed text is never stored, so it cannot be recovered. We do not keep a copy "just in case" anywhere — not in a log, not in an archive, not behind a permission. If a customer read out a card number and you need it, you need to ask them again. That is the intended behaviour and it is the entire point.
The transcript shows you that something was removed and what kind of thing it was, so a gap in the text reads as a redaction rather than as a transcription failure.
The recording is masked shortly after the transcript is, not at the same moment. Rewriting an audio file takes longer than rewriting text, so for a short window the transcript is clean while the recording still contains what was said. When that is the case the call says so in as many words, and asks you not to share the audio yet.
This is worth understanding rather than skipping: the incident nobody wants is a card number in a recording, not in a transcript, and a masked transcript beside an unmasked recording is exactly the state that looks finished and is not.
A masked recording is a WAV file, where the original was an MP3. Masking works on
uncompressed audio — writing silence into a compressed file corrupts it rather than quieting it —
so a recording we have had to mask is converted, and it downloads as .wav at roughly
four times the size. Recordings with nothing to mask, which is nearly all of them, are untouched
and stay as they were.
What we cannot remove, and what we do about it
Everything above has a fixed shape. A card number is sixteen digits that pass a checksum; an NHS number is ten with its own; a postcode has a form. That is exactly why they can be found in speech and taken out.
Health information has no shape. “The chest pains are back and the ramipril is not helping” carries a condition and a medication in ordinary words, and no list of terms would catch most of what people actually say. So we do not claim to remove it — a filter that catches some of a category is worse than none, because it gets reported as protection.
If your callers discuss health, or anything else UK GDPR calls special category data, say so on the Transcripts page, under Administration in the Admin menu. Transcription then becomes unavailable on your account, and if it was already on anywhere we switch it off. Recording is a separate setting and stays your decision: audio is encrypted, reachable only with a permission and an audit entry, and a practice that needs its own call recordings is not ours to refuse.
Transcripts already produced are not deleted by that switch. They come out with their normal retention period, or you can delete the call to remove its transcript immediately.
We do not turn any industry away, and this is not a HIPAA mode or a BAA. It is one control: we stop creating the artefact we cannot clean. Recordings are encrypted with a key per account before they are stored, and reaching one needs a permission and leaves an audit entry against a named person.
A transcript is different in a way worth knowing: it is stored as searchable text, because being searchable is the point of it. That is why the switch above exists at all — if your calls carry health information, the honest answer is not to make the text, rather than to make it and promise something about it.
Who can listen, and what is written down
Reading a transcript and downloading a recording are separate permissions from seeing the call log, because they are a different act: the log tells you a call happened, the recording is somebody's voice. Your team's roles decide who has them — see your team.
Every time somebody opens a transcript or downloads a recording, we write down who and when. It appears in the audit log against a named person. That is there so the question "who listened to that call" has an answer, which is what a complaint or a subject access request actually turns on.
Recordings and transcripts are never sent anywhere by us. Not in an email alert, not in a webhook, not as a link that happens to work without signing in. Everything that leaves Proofbell automatically carries facts about the call — when, how long, which campaign — and never the conversation. If you want a transcript somewhere else, someone with the permission fetches it, and that fetch is logged like any other.
Who else sees the audio
Speech-to-text is done by Deepgram, in the United States, and only for projects with transcription switched on. They are named in our data processing agreement along with everyone else who touches your data.
No sub-processor may use your calls to train its models. Deepgram runs a programme under which submitted audio can be used for training and taking part is the default — every request we send opts out of it. Turning that round would be a change to the sub-processor list and would take the notice period, not a setting somebody could flip.
What we work out from a call
Where transcription is on, four more things come back with the transcript: the overall tone of the call, the topics that came up, what the caller seemed to want, and a short summary. They appear above the transcript on the call itself.
These come from the same supplier that produces the transcript, computed from audio it already has, in the same request. No further company is involved and no extra copy of your calls is made — see the data processing agreement for the full position. They are switched on and off by the same transcription setting; there is nothing separate to enable, and nothing happens on a project with transcription off.
Treat them as a starting point, not a verdict. Intent in particular is a guess with a confidence attached — hover a label to see how sure it is. On a clear enquiry it is usually right; on a call where somebody rings off after ten seconds there is nothing to be sure about. It is there to help you find the calls worth listening to, not to replace listening to them.
Only the summary contains any of the words. Tone, topics and intent are stored as scores and labels against positions in the recording, not as text. The summary is written from the audio, so it can repeat something sensitive a caller said out loud — it goes through the same automatic removal the transcript does before it is saved, and the call says so when something was removed.
How long they are kept
Recordings and transcripts are deleted on the retention period your plan allows, and from Pro upwards you can set a shorter one. A shorter setting is always honoured; a longer one than your plan permits is refused when you set it rather than quietly reduced, so the figure on your screen is the figure things are actually deleted on.
New accounts start at 90 days for recordings and 365 for transcripts, inside whatever your plan's maximum is. The two differ on purpose: a recording is somebody's actual voice, and a transcript is text that every report here is built on — so the sensitive one gets the shorter default and the useful one keeps the longer.
If you stop subscribing
Recordings and transcripts are kept for 30 days after your subscription ends and then deleted. We email your account administrators before that happens, naming the exact date, and nothing is deleted until that email has gone out — the deletion is gated on the warning rather than running alongside it, so an account we could not reach keeps everything.
Starting a subscription again before the date cancels the deletion entirely. Your call records, reports and attribution data are not affected by any of this; they stay for the life of the account.
Deletion is permanent. A call whose transcript has been deleted says so, rather than showing the same empty panel as a call that was never transcribed — the two mean different things, and only one of them is worth waiting for. See what each plan includes.