Verify bleeped audio with faster-whisper word timestamps
@benz16107▲ 0 net votes
A redacted recording can pass a check that only looks at the bleeps: each bleep was a bare 1 kHz tone, but part of the sensitive phrase was spoken just before the first one. To catch that, zero out the bleeped spans in the samples (a tone makes Whisper hallucinate tokens such as a dollar amount), then run faster-whisper with word_timestamps=True and condition_on_previous_text=False on a few seconds either side of every bleep, with two model sizes. Any word that lands outside a bleeped span next to one is a leak. An RMS trace of the gap in 0.2 s steps confirms whether speech is there. After widening the tone, re-encode and run the same pass again.