Eliz­a­beth Gon­za­les

To­ba­go Cor­re­spon­dent

Two AI an­a­lysts say a leaked au­dio clip, al­leged­ly of a con­ver­sa­tion be­tween THA Chief Sec­re­tary Far­ley Au­gus­tine and Op­po­si­tion Leader Kam­la Per­sad-Bisses­sar, is most like­ly re­al.

Af­ter the clip was post­ed on so­cial me­dia last week­end, Au­gus­tine de­nounced it as “fool­ish­ness” and an at­tempt to link his To­ba­go Peo­ple’s Par­ty with the Unit­ed Na­tion­al Con­gress, while Per­sad-Bisses­sar de­scribed the au­dio as “fake news.”

How­ev­er, ac­cord­ing to one ex­pert, the tech­nol­o­gy tells a dif­fer­ent sto­ry.

“Ac­cord­ing to its syn­the­sis mod­el . . . it’s a two per cent pos­si­bil­i­ty (it’s AI-gen­er­at­ed)” said Steven Williams, an AI ap­pli­ca­tions con­sul­tant and ex­pert, with more than 30 years of ex­pe­ri­ence in tech­nol­o­gy and a ro­bust back­ground in cy­ber­se­cu­ri­ty, da­ta pri­va­cy, and dig­i­tal me­dia based in Bar­ba­dos.

Williams analysed the 577-sec­ond au­dio file us­ing three sep­a­rate tools, in­clud­ing 11Labs, which he de­scribed as “the lead­ing speech syn­the­sis mod­el.”

“It is very un­like­ly that this au­dio is gen­er­at­ed ei­ther by them or the au­dio was ma­nip­u­lat­ed in any way,” he ex­plained. “I found no tool that gave me any­thing above two to five per cent that it is AI-gen­er­at­ed.”

His re­port al­so high­light­ed the nat­ur­al sound and rhythm in the record­ing — fea­tures that are dif­fi­cult for AI to repli­cate.

“The au­dio was clean and con­sis­tent, in terms of things such as nat­ur­al sound or am­bi­ence, am­bi­ent noise. Noise in the nat­ur­al world is ran­dom. A bird chirp­ing, a car pass­ing by . . . AI could put cer­tain el­e­ments in, but it would have a cer­tain con­sis­ten­cy to it,” Williams said.

He al­so not­ed that “there was no abrupt clip­ping” which is a red flag that could sug­gest au­dio splic­ing or ma­nip­u­la­tion.

When eval­u­at­ing voice pat­terns, Williams said he looks for emo­tion, vari­a­tion, and rhythm, all of which ap­peared in the clip.

“Voice dy­nam­ics vary in loud­ness and in­flex­ion, so we find that AI hard­ly has the abil­i­ty to do this and that and be ex­pres­sive with sound and go up and down,” he ex­plained.

“There was no ex­is­tence of a pre­tence or an ar­ti­fi­cial in­stance of that. The tone was nat­ur­al, right, and the hu­man emo­tions con­veyed right was that of a hu­man be­ing.”

He added: “Paus­es and speech time ap­peared nat­ur­al and re­spon­sive, not over­ly pol­ished or ro­bot­ic.

Williams said AI sys­tems strug­gle with Caribbean di­alects, es­pe­cial­ly Trinida­di­an and To­bag­on­ian tones.

“Most AI tools are not trained on the Caribbean tonal­i­ty and ac­cent­ing. The like­li­hood that you would have some per­son train AI to ob­fus­cate a sin­gle con­ver­sa­tion is rel­a­tive­ly low,” he said.

“You have to train in a com­plete­ly new lan­guage.”

In the leaked record­ing, a male voice is heard dis­cussing funds al­leged­ly sent for two elec­tions, a pro­posed po­lit­i­cal al­liance, and what ap­peared to be a plan to un­der­mine con­trac­tors in To­ba­go. While no speak­er has been of­fi­cial­ly con­firmed, the con­tents of the al­leged record­ing have trig­gered de­bate and con­fu­sion in To­ba­go.

Williams was care­ful to note that his re­sults were not a foren­sic rul­ing.

“Even AI, there’s no one or ze­ro in terms of ei­ther yes or no. We talk about per­cent­ages,” he said.

How­ev­er, based on the tests he con­duct­ed, he con­clud­ed that the nat­ur­al tim­ing, tone and speech pat­tern in the clip point away from AI in­volve­ment.”

Travis Sookoo, a cy­ber­se­cu­ri­ty ad­vi­sor and ma­chine learn­ing re­searcher had a more cau­tious but in­sight­ful per­spec­tive on how to de­tect AI-gen­er­at­ed voic­es, fo­cus­ing on pat­terns the pub­lic can learn to recog­nise.

“There’s a sort of per­fec­tion that comes with it, even in speech,” he said.

“Even with di­alect, you know, you will hear some sort of per­fec­tion.”

“Re­al speech has flaws. If the voice sounds too smooth, too per­fect, that’s when you should be sus­pi­cious.”

Sookoo said with most AI tools, the rhythm and de­liv­ery might be too pol­ished.

“If we look at how AI writes and how it says things back, it is usu­al­ly in a per­fec­tion­ist base,” he said.

Sookoo added that at­ten­tion should be paid to tone, move­ment, di­alect, and small de­tails.

“There may be some­thing that will catch your ear if you are lis­ten­ing to it, the in­to­na­tion, how they use the ac­cent, the di­alect,” he said.

Both an­a­lysts called for more at­ten­tion to be paid to dig­i­tal lit­er­a­cy.

So far, no one has con­firmed or de­nied who is speak­ing in the clip but with no signs of voice merg­ing, no au­dio ma­nip­u­la­tion de­tect­ed, and clear in­di­ca­tors of nat­ur­al hu­man speech, the tools seem to point in one di­rec­tion.

“This is ei­ther a re­al con­ver­sa­tion, or it’s the most ad­vanced fake I’ve ever seen,” Williams said.