Research‑Driven
Software Engineering

asr-redact — Thai Speech PII Redaction

A pipeline that finds personally identifiable information in Thai speech and censors it in the audio itself, not just in the transcript. Silero VAD packs speech into ≤30 s windows so cuts land in silence, Typhoon ASR (a NeMo RNNT model, driven directly rather than through its convenience wrapper) returns subword timestamps, and two detectors — PyThaiNLP NER plus pattern rules for identifiers dictated digit by digit — mark the spans. Because an RNNT emits a token at or after the sound it describes, each token is widened back toward the previous emission so a redaction covers the whole word rather than its tail. Every input yields a beeped/silenced/noised copy and a JSON report of what was removed and when.