The replacement detector computes RMS over 20 ms frames at a 10 ms hop, then applies one voiced test: rms > max(rms.max() * 0.06, 0.004) [10]. The relative arm sits 24.4 dB below the loudest frame in the clip [18]. The absolute arm sits at roughly -48 dBFS [19]. Taking the larger of the two carries the design, because a relative threshold alone lets the noise floor read as voiced on a quiet render, and an absolute threshold alone promotes breath and lip smacks on a loud one [11]. The twelve voices spanned 5.5 dB of stated level, -13.3 to -18.8 [12][20], and the post does not say which level measure that is.
Voiced blocks under 30 ms are then discarded, because otherwise lip and quantisation noise fragments into many tiny blocks and the sequence logic downstream stops meaning anything [13]. That 30 ms is half the 0.06 s minimum tail [21], which is the line I would check first in anyone else's copy of this code: a block filter set above your artifact threshold quietly deletes the thing you are trying to detect.
The artifact test needs at least two voiced blocks, a gap before the last one of 0.25 s or more, and a last block of 0.06 s or longer [14]. Measured artifacts had gaps of 0.26 to 0.91 s and tails of 0.07 to 0.36 s [15]. So the smallest observed gap clears its threshold by 0.01 s, exactly one hop [22], and the shortest tail clears its threshold by the same 0.01 s [23]. Those two constants are fitted to four failures of one model family, according to the dev.to write-up.
The false positive is the useful part of the story. A comma pause is correctly rendered script audio, and it satisfied both conditions with margin to spare [16]. GAP_MIN is a property of your probe script's punctuation, not of TTS artifacts [14][16]. The copy of the post I have cuts off as the author begins to state the discriminator he added, so I cannot tell you what replaced the bare gap rule.
Two arithmetic notes on the defect report itself. Body plus silence plus artifact comes to 1.52 s against a stated 1.61 s total, and 1.56 s against 1.65 s, leaving 0.09 s unaccounted for in both clips, which the post does not place [4][24].
The retention order in the transcripts runs against duration: the shorter artifact left a trace and the longer one, nearly twice its length, left none [7][25]. The stated reason is that short sounds which are not meaningful words may not appear in STT output at all [8]. Before the waveform work, the author read his own null result as evidence about the model, recording that STT scored 0/6 on extra sounds and concluding there were none [9].
For these constants to transfer to your pipeline, two things have to be true. Your probe text must contain no internal silence longer than 0.25 s [14], and your artifacts must sit above roughly -48 dBFS [19]. The first one you write yourself.
