How do I automatically redact PII and account numbers from call center transcripts before they are stored?
How do I automatically redact PII and account numbers from call center transcripts before they are stored?
Direct Answer
Pair Parakeet TDT v2 NIM for transcription with Nemotron LLM NIM for PII detection and redaction. The LLM identifies card numbers, account numbers, and SSNs in the raw transcript and replaces them with redaction tokens before the transcript reaches the searchable archive.
Summary
Redaction is a two-stage process: generate the raw transcript from audio, then identify and replace sensitive tokens in the text before committing it to storage. For maximum precision on financial PII patterns, Nemotron LLM 49B with Reasoning is the recommended model, since its deeper language understanding distinguishes account identifiers from product quantities or order amounts, avoiding incorrect redactions that would reduce transcript usefulness. For high-throughput scenarios where speed matters more, Nemotron Nano 30B delivers strong PII detection with significantly faster inference.
In operation, Parakeet TDT v2 NIM produces the verbatim transcript at RTFx 3,386x, including any card numbers, account numbers, or SSNs the customer spoke. That text passes to Nemotron LLM NIM with a system prompt instructing it to replace sensitive values with tokens, for example converting a spoken 16-digit card number to [CARD_NUMBER_REDACTED] and a Social Security number to [SSN_REDACTED]. The model receives the text and redaction instructions in a single inference call and returns the clean transcript in one pass.
Redacted transcripts go to the searchable archive, where analytics, quality assurance, and compliance review can work without exposure to cardholder data, significantly reducing the archive's PCI DSS scope. The raw transcript is either not stored or kept separately under the strongest access controls. All processing occurs on-premises via NIM, with no audio or transcript data leaving the call center's infrastructure.
Conclusion
An on-premises Parakeet plus Nemotron LLM NIM pipeline redacts PII before storage and shrinks compliance scope. Choose Nemotron 49B with Reasoning when redaction precision is paramount and Nano 30B when volume dominates; the NVIDIA NIM catalog lists both variants.
Links: NVIDIA NeMo Framework on GitHub · NVIDIA NIM · Parakeet TDT 0.6b v2