Reproduce “Using Text Injection to Improve Recognition of Personal Identifiers in Speech” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Yochai Blau, Rohan Agrawal, Lior Madmony, Gary Wang, Andrew Rosenberg, Zhehuai Chen, Zorik Gekhman, Genady Beryozkin and 2 more
A text-injection method by Google researchers improves Automatic Speech Recognition (ASR) systems' ability to recognize personal identifiers after training on privacy-compliant, de-identified data. By introducing synthetic text with fake PII during training, the system achieved an 8% increase in Names recall and a 13% increase in Dates recall in medical speech, alongside a 3.2% increase in Sentence Accuracy for alphanumeric sequences.
Key insights
- Text injection, even without corresponding audio examples of PII, significantly restores and improves the recognition of sensitive terms (e.g., names, dates, alphanumeric sequences) in ASR models trained on de-identified data.
- Leveraging readily available text-only data containing synthetic PII provides an effective and privacy-compliant mechanism to overcome the scarcity of sensitive transcribed speech data in specialized domains.
- The approach effectively supports domain adaptation to medical ASR, improving both overall Word Error Rate (WER) and the recall of medical terms, alongside mitigating the negative effects of PII redaction and elimination.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.