PII redaction

PII redaction is the process of automatically detecting personally identifiable information, such as names, phone numbers, and card details, in a transcript and replacing it before the transcript is stored or processed. Doing it at ingestion, as soon as audio becomes text, is what keeps call transcripts PCI compliant, because unredacted data never reaches your database.

It depends on the regulation you’re under. GDPR covers any data that identifies a person, PCI DSS covers card and payment data, and HIPAA lists 18 identifiers for health data. Gladia groups names, contact details, government IDs, financial data, and health information into 13 presets such as GDPR, HIPAA Safe Harbor, PCI, CPRA, and APPI. You can also target individual entity types like CVV or PASSPORT_NUMBER, which helps contact centers meet GDPR, SOC 2, and ISO 27001 requirements without over-redacting.

Enable PII redaction in your async transcription request and choose a preset or a list of entity types. Gladia then returns the transcript with each detected entity replaced, in the same API call. Because redaction happens at transcription time, you don’t need a separate redaction tool or a post-processing pipeline.

Look for a provider that redacts at ingestion, offers presets mapped to specific regulations, and holds the certifications your buyers will ask for. Gladia holds HIPAA, GDPR, SOC 2 Type 2, and ISO 27001 compliance, and its HIPAA Safe Harbor preset de-identifies clinical transcripts automatically. These are the main criteria to check when choosing HIPAA-compliant speech-to-text for healthcare products.

The PCI preset removes card numbers, expiration dates, CVVs, and bank details from the transcript, so payment data never spreads into QA, analytics, or CRM tools. Redaction applies to the transcript text, not the audio file, so call recording compliance under GDPR, PCI DSS, and HIPAA still depends on how you store and restrict access to the recordings themselves.

Not yet. PII redaction currently runs on async transcription only, so a common setup is to stream the call with real-time transcription and send the recording through the async API afterward for the redacted version you store.

Yes. With MARKER output, each entity gets a consistent label such as [NAME_1], so an LLM can follow the conversation without ever seeing the raw data. Audio-to-LLM runs your prompts on that redacted transcript in the same request, which is a key part of making AI transcription legal and safe for support calls.

No. PII redaction is included in Gladia’s per-hour async rate, with no add-on fee on any pricing plan. Many providers sell PII redaction as a separate add-on, which adds up quickly at contact center volumes.