Reconstruction history, not synthesis: benign processing moves frozen detectors on genuine human speech

{"Synthetic-speech":[0],"detectors":[1,30,75,138,212,229],"are":[2,230,249,335,347],"deployed":[3],"on":[4,19,71,109],"the":[5,103,114,140,210,236,247,264,269,276,280,285,288,319,350],"premise":[6,17],"that":[7,16,38,228,283],"their":[8],"output":[9],"tracks":[10,148],"synthetic":[11],"provenance.":[12],"This":[13,156],"work":[14],"tests":[15],"causally":[18],"genuine":[20,55],"human":[21,56],"studio":[22],"recordings":[23],"using":[24],"thirteen":[25,211],"frozen,":[26],"publicly":[27],"available":[28],"pretrained":[29],"and":[31,60,108,184,246,259,272,279,287,302,308,332,358],"a":[32,49,110,167,201,205,314],"set":[33],"of":[34,54,73,86,92,136,170,209,235,318],"pre-registered":[35],"benign":[36],"transformations":[37],"generate":[39],"no":[40,63,131,185],"speech":[41,57],"tokens.":[42],"On":[43],"47":[44,85,87,192],"utterances":[45,88],"from":[46,128,214,298,349],"three":[47,80,99,154],"speakers,":[48,191],"neural-codec":[50],"analysis–synthesis":[51],"round":[52],"trip":[53],"—":[58,65,198],"encode":[59],"decode,":[61],"with":[62,89,130],"generation":[64],"moved":[66,134],"detector":[67,147,267],"scores":[68],"toward":[69],"\\"synthetic\\"":[70],"9":[72],"13":[74,137],"surviving":[76],"Holm":[77],"correction,":[78],"including":[79],"fully":[81],"prospective":[82],"checkpoints":[83,164],"at":[84],"matched-pairs":[90],"rank-biserial":[91,289],"exactly":[93],"−1.000.":[94],"The":[95,223,256,325],"effect":[96,169],"persisted":[97],"across":[98,153],"professional":[100],"microphones":[101],"capturing":[102],"same":[104],"physical":[105],"performance":[106],"simultaneously,":[107],"second":[111],"speaker":[112],"outside":[113],"corpus.":[115],"It":[116],"was":[117,182],"not":[118,204,220,226],"confined":[119],"to":[120,296,303],"codecs:":[121],"linear":[122],"Griffin-Lim":[123],"phase":[124],"retrieval,":[125],"an":[126,160],"algorithm":[127],"1984":[129],"neural":[132],"network,":[133],"12":[135],"in":[139,232,251],"synthetic-ward":[141],"direction.":[142],"Which":[143],"transformation":[144],"moves":[145],"which":[146,218],"its":[149],"documented":[150],"training":[151,171],"exposure":[152],"architectures.":[155],"is":[157,177,188,219],"reported":[158],"as":[159,338],"association":[161],"among":[162],"frozen":[163,265,270],"rather":[165],"than":[166],"causal":[168],"data.":[172],"Stated":[173],"limits.":[174],"No":[175,179],"mechanism":[176],"identified.":[178],"listening":[180],"test":[181],"run":[183,297],"perceptual":[186],"claim":[187,227],"made.":[189],"Three":[190],"utterances,":[193],"neutral":[194],"speech,":[195],"one":[196,215],"language":[197],"enough":[199],"for":[200,316],"paired":[202],"effect,":[203],"population":[206],"claim.":[207],"Seven":[208],"come":[213],"research":[216],"group,":[217],"seven":[221],"replications.":[222],"paper":[224,257],"does":[225],"broken":[231],"general.":[233],"Thirteen":[234],"author's":[237],"own":[238],"claims":[239],"were":[240,294],"withdrawn":[241],"during":[242],"independent":[243],"adversarial":[244],"audit":[245],"withdrawals":[248],"published":[250],"Section":[252],"9.4.":[253],"Deposit":[254],"contents.":[255],"(PDF":[258],"Markdown":[260],"source),":[261],"citation":[262],"metadata,":[263],"per-utterance":[266],"scores,":[268],"result":[271],"pre-registration":[273],"freeze":[274],"records,":[275],"five":[277],"figures,":[278],"two":[281],"scripts":[282,293],"regenerate":[284],"figures":[286],"bootstrap":[290],"intervals.":[291],"Both":[292],"verified":[295],"this":[299],"package":[300],"alone":[301],"reproduce":[304],"every":[305],"deposited":[306],"figure":[307],"interval":[309],"file":[310],"byte-for-byte.":[311],"DEPOSIT_MANIFEST.json":[312],"carries":[313],"SHA-256":[315],"each":[317],"other":[320],"25":[321],"files.":[322],"Not":[323],"included.":[324],"11.5":[326],"GiB":[327],"audio":[328],"evidence":[329],"estate.":[330],"Speaker":[331],"sample":[333],"identifiers":[334],"retained":[336],"deliberately":[337],"provenance":[339],"under":[340],"signed":[341],"contract;":[342],"workstation-specific":[343],"absolute":[344],"path":[345],"prefixes":[346],"stripped":[348],"released":[351],"trial":[352],"record,":[353],"whose":[354],"relative":[355],"dataset":[356],"paths":[357],"filenames":[359],"remain":[360],"traceable.":[361]}

Authors

Institutions

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-09-17
DOI
https://doi.org/10.5281/zenodo.22819222
Primary Topic
Memory Processes and Influences
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
preprint

Reconstruction history, not synthesis: benign processing moves frozen detectors on genuine human speech

Panagiotis Gkilis
Zenodo (CERN European Organization for Nuclear Research)
Memory Processes and Influences
preprint

Reconstruction history, not synthesis: benign processing moves frozen detectors on genuine human speech

Panagiotis Gkilis
preprint en

Abstract

Synthetic-speech detectors are deployed on the premise that their output tracks synthetic provenance. This work tests that premise causally on genuine human studio recordings using thirteen frozen, publicly available pretrained detectors and a set of pre-registered benign transformations that generate no speech tokens. On 47 utterances from three speakers, a neural-codec analysis–synthesis round trip of genuine human speech — encode and decode, with no generation — moved detector scores toward "synthetic" on 9 of 13 detectors surviving Holm correction, including three fully prospective checkpoints at 47 of 47 utterances with matched-pairs rank-biserial of exactly −1.000. The effect persisted across three professional microphones capturing the same physical performance simultaneously, and on a second speaker outside the corpus. It was not confined to codecs: linear Griffin-Lim phase retrieval, an algorithm from 1984 with no neural network, moved 12 of 13 detectors in the synthetic-ward direction. Which transformation moves which detector tracks its documented training exposure across three architectures. This is reported as an association among frozen checkpoints rather than a causal effect of training data. Stated limits. No mechanism is identified. No listening test was run and no perceptual claim is made. Three speakers, 47 utterances, neutral speech, one language — enough for a paired effect, not a population claim. Seven of the thirteen detectors come from one research group, which is not seven replications. The paper does not claim that detectors are broken in general. Thirteen of the author's own claims were withdrawn during independent adversarial audit and the withdrawals are published in Section 9.4. Deposit contents. The paper (PDF and Markdown source), citation metadata, the frozen per-utterance detector scores, the frozen result and pre-registration freeze records, the five figures, and the two scripts that regenerate the figures and the rank-biserial bootstrap intervals. Both scripts were verified to run from this package alone and to reproduce every deposited figure and interval file byte-for-byte. DEPOSIT_MANIFEST.json carries a SHA-256 for each of the other 25 files. Not included. The 11.5 GiB audio evidence estate. Speaker and sample identifiers are retained deliberately as provenance under signed contract; workstation-specific absolute path prefixes are stripped from the released trial record, whose relative dataset paths and filenames remain traceable.

Zenodo (CERN European Organization for Nuclear Research)
Bevital (Norway) (NO)
Peace, Justice and strong institutions
Memory Processes and Influences
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.