What is PII?
Personally Identifiable Information, anything that can identify a specific individual, directly or indirectly.
Names, phone numbers and ID numbers are obvious PII. So is the voice itself. A voiceprint can be used to identify someone.
That is why the compliance chain on a speech dataset has to be complete: the speaker's informed consent, the agreed scope of use, and de-identification where it is required.
What counts as PII
Direct identifiers name the person on their own: a full name, phone number, national ID number, email address or home address. In speech data these can arrive in the metadata, in the filename, or spoken aloud in the recording. A batch with clean filenames can still carry a speaker reading out an account number, which is why the audio itself is part of any PII review.
Indirect identifiers identify only in combination. A rare job title plus a small town, a birth date plus a postcode, a distinctive event described in the audio. None of these is a name, and together they can point at one person. That is why de-identification which only removes names and numbers is not enough, and why a scrubbed transcript can still be personal data.
A third group gets special treatment. A voiceprint, or any recording processed to identify a person by voice, is biometric data, and most privacy regimes give it extra protection. Health information and children's recordings attract their own extra rules. What the recording is used for decides which treatment applies, not the file itself, so the same audio can be ordinary data in one project and regulated data in another.
How it affects a data purchase
Ask how PII was handled, and where. There are three common treatments: scrubbing, where identifying content is removed from the transcript and the metadata; masking, where the audio itself is edited by bleeping or muting; and retention, where it is kept and the buyer's use is covered by consent and a contract instead. Each leaves the data in a different state.
Masking is the one to look at closely for speech. Editing the audio removes the identity and damages the signal around it: a bleeped phone number is a hole in the acoustics, and if the target model has to learn numbers, those files may be worthless for that purpose. Scrubbing the transcript keeps the audio intact but leaves the words in the recording.
The practical check is a sample. Listen to the files and read the metadata before the full delivery. Two failures are common: a set sold as de-identified that ships with speaker names in the filenames, and a transcript that still contains phone numbers because only the audio was cleaned. Ask which of the two was cleaned, and whether the original was kept.
Related terms
-
Data Compliance
Where the data came from, whether it may be used that way, and whether there is paper to prove it. That whole chain is compliance.
-
Consent
The speaker knows how their recording will be used, and has signed to say so.
Keep reading
-
Full glossary
Every term we explain, in one list.
-
How to buy training data
Where these terms actually show up, and which ones change a quote.
-
Help center
How a project runs, from specification to delivery.
Submit a sourcing request
Tell us the language, the hours, and what the data needs to look like. You will get a real number and a real timeline, not a range. If we cannot source it well, we will tell you that instead.
- Pilot batch before the full run, so problems surface early.
- Consent documentation delivered with the data.
- No medical or clinical data. No recorded telephone calls.
- You send the spec Language, hours, recording conditions, and the deadline.
- We reply within two business days A real number and a real timeline, not a range.
- You approve a pilot batch A small run first, so problems surface before the full order.
- Full run, then delivery with documents Audio, transcripts, and the consent record for every speaker.