Impact of Phonetics on Preserving Identity in Adversarial Voice Attacks

Research Overview

This website accompanies our ICASSP paper and provides interactive access to all materials referenced in the manuscript. For each target phrase, we include two sets of normalized speaker–similarity graphs—one generated using an ECAPA‐TDNN backbone and one using a ResNet backbone. These graphs visualize how the similarity distributions shift under adversarial perturbations. We also provide scatter plots illustrating the relationship between signal‑to‑noise ratio (SNR) and similarity scores for each phrase.

Target Metrics Table

ID	Target Text	Phonetic Short Note	Model	#Samples	#Genuine	#Impostor	TMR@0.1%FMR	d′
T1	yes	mono-syll.; 1:2 V:C; glide+fric stop	ECAPA	11881	109	11772	1.0000	9.68
			ResNet50	11881	109	11772	1.0000	9.43
T2	open the door	4 syll.; 4:6 V:C; dental fric. + stops	ECAPA	11881	109	11772	1.0000	9.11
			ResNet50	11881	109	11772	1.0000	8.89
T3	call emergency services	fricative-heavy; 8 syll.; 8:16 V:C	ECAPA	11881	109	11772	0.9725	5.59
			ResNet50	11881	109	11772	0.9817	6.22
T4	the quick brown fox jumped over the lazy dog	pangram; 11 syll.; broad coverage	ECAPA	11881	109	11772	0.9083	4.80
			ResNet50	11881	109	11772	0.9817	5.14
T5	shhh she sees the sea fish	fricatives /ʃ s z/ cluster; 6 syll.	ECAPA	11881	109	11772	0.9908	7.46
			ResNet50	11881	109	11772	1.0000	7.74
T6	do go big bag dig	voiced stops chain; minimal vowels; 5 syll.	ECAPA	11881	109	11772	1.0000	8.44
			ResNet50	11881	109	11772	1.0000	8.35
T7	two tall teachers talk to tim	/t/ alliteration; alveolar bursts; 7 syll.	ECAPA	11881	109	11772	1.0000	7.79
			ResNet50	11881	109	11772	1.0000	7.59
T8	i whisper while walking wildly	approximants + /w/ clusters; 8 syll.	ECAPA	11881	109	11772	1.0000	7.17
			ResNet50	11881	109	11772	1.0000	7.33
T9	pack my box with five dozen liquor jugs	pangram; many consonant clusters	ECAPA	9025	95	8930	0.8632	4.63
			ResNet50	9025	95	8930	0.9474	5.02
T10	glib jocks quiz nymph to vex dwarf	pangram; high fricative/affricate load	ECAPA	9025	95	8930	0.8421	4.75
			ResNet50	9025	95	8930	0.9579	5.34
T11	a mad boxer shot a quick gloved jab to the jaw of his dizzy opponent	mixed plosives/fricatives; many unstressed vowels	ECAPA	9025	95	8930	0.7474	3.79
			ResNet50	9025	95	8930	0.9263	4.45
T12	just before twilight the wizard quickly jabbed five boxes of hazy quartz to vex a plump knight’s jovial frog	very long pangram; many clusters; vowel centralization	ECAPA	6561	81	6480	0.4444	3.07
			ResNet50	6561	81	6480	0.7160	3.63
T13	twelve jolly grizzlies briskly danced over waxy benches while a flighty kitten kept humming jazz tunes in the background	fricative+affricate mix; many unstressed vowels	ECAPA	3025	55	2970	0.5273	3.06
			ResNet50	3025	55	2970	0.6364	3.46
T14	quantum driven flux engines jam beneath zigzagging vortex panels as cryptic bioforms whisper behind polymorphic glass domes	dense consonant clusters; many fricatives/affricates	ECAPA	2209	47	2162	0.6809	3.10
			ResNet50	2209	47	2162	0.7447	3.78
T15	while whispering winds wander westward jittery jackals jiggled jellies above velvet jars beyond flickering bonfires in a frozen jungle	repeated /w/ /ʤ/; long with many approximants	ECAPA	3025	55	2970	0.6545	3.39
			ResNet50	3025	55	2970	0.7091	3.54
T16	kindly expedite bizarre frozen jumpsuits for victors whirlwind gala to maximize xenon emissions before daybreak	‘x/ʒ/ks’ clusters; mixed stops/fricatives; multi-syllabic	ECAPA	5184	72	5112	0.5278	3.09
			ResNet50	5184	72	5112	0.6806	3.30

ECAPA-TDNN Normalized Results

ResNet Normalized Results

SNR vs Similarity Analysis

Scatter plots showing the relationship between Signal-to-Noise Ratio (SNR) and similarity metrics for each target text, providing insights into the trade-offs between attack strength and audio quality.