Where the numbers come from
We analyzed the privacy policies, terms of service, data processing agreements, sub-processor lists and AI-data pages of 53 transcription and meeting-notes services, taken from Internet Archive captures and coded twice, independently, against a written codebook. Method and data are in the paper. The data describe what the documents say, not what any service does.
Audio retention: usually 'as long as necessary'
The documents of 22 services (41.5%) say audio is kept as long as necessary for stated purposes. Ten (18.9%) say until the user deletes it, and 14 (26.4%) give a fixed period. Four say nothing and 2 conflict.
Transcripts follow the same pattern: open-ended at 22 (41.5%), until the user deletes at 16 (30.2%), and a fixed period at 8 (15.1%). 'As long as necessary' is common in privacy law and not wrong, but it does not let a researcher write a retention period into a data management plan. Services that state a fixed period, or delete audio after processing, give users something they can rely on.
Deleting it yourself
Twenty-nine services (54.7%) state how long deletion takes after a user deletes content or closes an account, most often 30 days. Twenty-two (41.5%) mention deletion without giving a time.
Two timings are worth separating when you read: how long the service keeps content if you do nothing, and how long it takes to remove content after you ask. Documents often answer only one.
Who can see the recording
Eleven services (20.8%) describe routine human access to customer content, for example annotation, quality review or transcription by human contractors. Twenty-four (45.3%) describe access only under conditions such as support requests, user consent or legal obligation. Three say staff never access content, and 10 say nothing.
Sub-processors and AI providers are the next layer. Twenty-five services (47.2%) publish a named list of sub-processors and 26 (49.1%) describe them only by category. Twenty-two (41.5%) name the third-party AI or speech providers that process content, 11 describe them by category and 18 do not mention them.
Location and voiceprints
Forty-four services (83.0%) state where data is stored or processed. Twenty (37.7%) mention voiceprints, voice data as biometric data or biometric identifiers: 13 together with consent or notice terms and 7 without.
If your recordings include identifiable voices, those two lines matter more than the headline retention figure.
A short list to take away
Look for a number, not a purpose. Note who is named as having access and under what conditions. Check whether a sub-processor list exists and names the AI provider. Keep the date you checked.
The pre-upload checklist lays this out as eight questions, whether services train on your recordings covers the training side, and the guides on confidential transcription and HIPAA and AI transcription go into handling sensitive audio. Pepys is one of the 53 services in the study, coded the same way as the rest. How Pepys Transcribe itself handles a recording is set out in its privacy and data handling reference.

