VConShield - PII Redaction

Redactify: Smart, Secure, and Seamless PII Removal from Conversations

Comment

🛡️ VConShield – PII Redaction

Protecting Conversations, Preserving Privacy

🔍 Inspiration

In a world driven by digital communication, protecting user privacy is more important than ever. With the rise of virtual conversations and voice data sharing, personally identifiable information (PII) is often unknowingly exposed. We were inspired to build VConShield to address this critical concern by automatically identifying and redacting PII from conversational data (vCons), ensuring compliance and user trust.

💡 What it does

VConShield takes audio conversations (e.g., .wav, .mp3) and returns redacted .vcon files by:

  • Identifying PII such as names, addresses, phone numbers, emails, etc., using advanced Named Entity Recognition (NER) models.
  • Redacting the PII in transcripts by replacing it with [REDACTED].
  • Optionally processing the audio to either beep, silence, or cut the identified PII segments.
  • Providing a secure conserver link to access both detected and redacted outputs.

🛠️ How we built it

  • Transcription: Converted audio input to text using Whisper from OpenAI for accurate speech-to-text conversion.
  • PII Detection: Used spaCy’s NER and fine-tuned transformer-based models (like BERT and RoBERTa) to detect PII entities.
  • Redaction: Replaced PII in the transcript with [REDACTED] tags.
  • Audio Redaction: Used audio processing libraries (pydub, ffmpeg) to modify the original audio by beeping or silencing the PII segments.
  • vCon Integration: Packaged output into .vcon format using open-source vCon schema.
  • Frontend: Created a simple upload interface using Flask and integrated conserver links for download.

🚧 Challenges we ran into

  • Aligning time-stamped audio segments to specific text tokens for precise redaction.
  • Detecting less-obvious PII like gender indicators, religious beliefs, or voice metadata.
  • Balancing redaction accuracy without over-scrubbing useful context.
  • Processing long conversations without exceeding model or memory limits.

🏆 Accomplishments that we're proud of

  • Successfully built an end-to-end pipeline that works for both transcript and audio redaction.
  • Achieved high precision in PII detection using fine-tuned NER models.
  • Integrated with vCon format and delivered results via conserver link – clean, simple, and secure.
  • Created an intuitive demo that non-technical users can interact with.

📚 What we learned

  • How to build robust NER pipelines tailored for spoken conversations.
  • How to work with audio segmentation, timestamps, and waveform redaction.
  • The importance of addressing edge-case PII like voice, city, or inferred metadata.
  • Practical implementation of the vCon standard and working with conserver links.

🚀 What's next for VConShield - PII Redaction

  • Integrating multilingual PII detection and redaction.
  • Supporting real-time redaction for live call centers.
  • Adding customizable redaction rules per compliance frameworks (e.g., HIPAA, GDPR).
  • Improving the UI/UX and deploying as a cloud-native SaaS tool.
  • Collaborating with telehealth, customer support, and legal teams to adopt privacy-first audio processing.

Built With

Share this project:

Updates