Vocal Authenticity Resistance Index (VARI) Dataset: YouTube Audience Discourse on AI Voice Aesthetics — Vocal Resistance vs Vocal Acceptance Corpus (n=2,423, April 2026)

This dataset contains 2,423 YouTube comments collected viathe YouTube Data API v3 (April 2026) across five vocalidentity discourse domains targeting naturalistic consumeraesthetic responses to AI-generated voices in educationaland podcast content: (1) AI voice educational video annoying robotic 2025(2) AI generated voice podcast natural or fake listener(3) AI voice vs human voice which is better education(4) AI narrator voice uncanny emotional connection 2025(5) AI voice accent bias western voice 2025 The corpus is the primary empirical dataset for the VocalAuthenticity Resistance Index (VARI) — the firstcomputational metric measuring the ratio of VocalResistance to Vocal Acceptance in naturalistic consumerdiscourse about AI-generated voices, providing empiricalgrounding for decolonial aesthetic critique of vocaluniversalism in AI systems. VARI = Vocal Resistance / (Vocal Resistance + Vocal Acceptance)1.0 = pure Vocal Resistance (maximum aesthetic rejection)0.5 = contested aesthetic space0.0 = pure Vocal Acceptance (aesthetic embrace) Three signal types — distinct from all prior metrics: VOCAL RESISTANCE (VR): aesthetic refusal of AI voice → "robotic", "soulless", "creepy", "can't connect", "unnatural", "lifeless", "uncanny", "impersonal" Distinct from CAABI (deception anger), ASAI (surveillance fear) — THIS: sensory-aesthetic rejection of non-human vocal presence in educational mediation VOCAL ACCEPTANCE (VA): aesthetic embrace of AI voice → "natural", "soothing", "professional", "warm voice", "feels connected", "impressive", "love the voice" VOCAL INDIFFERENCE (VI): content-over-voice orientation → "doesn't matter", "just here for content", "ignore voice" Theoretically critical: Biesta (2015) identifies instrumental orientation as the failure mode of educational mediation — VI empirically operationalises this critique Grounded in:- Wenzel et al. (2025) accent bias in AI voice services- Mignolo (2013) coloniality of sensing / geopolitics of aesthesis — vocal universalism as epistemic silencing- Haraway (1988) situated knowledges- Noddings (2013) care ethics — relational vs transactional- Biesta (2015) education as relational risk- Zuboff (2019) surveillance capitalism- Mori (1970) uncanny valley — extended to AI voice KEY FINDINGS:- Total corpus: n=2,423 comments, ~50 videos- Categorised (VARI computed): 59 (2.4%)- Mean VARI: 0.9237 — near-maximum Vocal Resistance- Vocal Resistance cluster: 54 (91.5% of categorised)- Vocal Acceptance cluster: 4 (6.8%) — critically absent- Vocal Indifference cluster: 3 (5.1%) — epistemic instrumentalism present but marginal- VR density: 0.1752/100 tokens- VA density: 0.0127/100 tokens (13.8:1 ratio)- VI density: 0.0049/100 tokens (minimal) DOMAIN-LEVEL VARI:- Educational video / robotic: VARI=1.0000 (pure resistance) VR=9.01/100 | VA=0.00/100 — maximum aesthetic rejection- Uncanny/emotional connection: VARI=1.0000 (pure resistance) VR=5.20/100 | VA=0.00/100 — zero acceptance markers- AI vs human voice / education: VARI=0.8971 (resistance) VR=7.24/100 | VA=0.88/100 — only domain with VA>0- Podcast natural/fake: VARI=0.8571 (resistance dominant) VR=5.01/100 | VA=0.10/100 KEY THEORETICAL FINDING — EMPIRICAL VALIDATION OF SPDF:Mean VARI=0.9237 and VR:VA ratio of 13.8:1 empiricallyconfirm the Sensory-Phenomenological Design Framework(SPDF) prediction that vocal universalism in AI systemsgenerates overwhelmingly negative aesthetic response innaturalistic consumer discourse. The near-total absenceof Vocal Acceptance (4 comments, 6.8%) demonstrates thatAI voice design has not achieved relational presence ineducational contexts. The most-liked Resistance comment(6,555 likes): "Nah Kai AI voice got it" — and (218likes): "1000% AGREE, as a consumer, as soon as I hearthat AI voice (we all know the one I'm talking about)" —confirm that Vocal Resistance is culturally shared andsocially reinforced, not individually idiosyncratic. Files:- vari_videos.csv: ~50 unique videos metadata- vari_comments.csv: 2,423 raw comments- vari_results.csv: VARI scores + signal annotation Method: YouTube Data API v3. Python 3.12, April 2026.