2/ #Deepfake'd speech has quirks.
For now.
Providing people with reference audio can help a bit.
For now.
But training about #deepfakes? Or listening again and again? Not so much.
Our limited ability to reliably spot deepfaked speech = plenty of need to improve automated #deepfake detectors.
As deepfakes proliferate, finding ways to augment human detection (we're probably still better at spotting contextual quirks) will be key.
John Scott-Railton ☕ /
npub18r…e9j7t
2023-08-08 20:59:18
in reply to nevent1q…p9t9
Author Public Key
npub18r8efsp3d9r79v42huwlyjstzr3y92s99zyfur2q2utnqpyzsrcqce9j7tPublished at
2023-08-08 20:59:18Event JSON
{
"id": "6d1bea1cffdb4130c22b0fcfff454780d88eb98cc82e9bb86c953f6381c4a281",
"pubkey": "38cf94c0316947e2b2aabf1df24a0b10e242aa0528889e0d40571730048280f0",
"created_at": 1691528358,
"kind": 1,
"tags": [
[
"e",
"0d0aec97c64e3bea0d3e319046b130e1fad930f6792f8da056d39a8abdc5f4c4",
"wss://relay.mostr.pub",
"reply"
],
[
"t",
"deepfake"
],
[
"t",
"deepfakes"
],
[
"proxy",
"https://mastodon.social/users/jsrailton/statuses/110856002517873723",
"activitypub"
]
],
"content": "2/ #Deepfake'd speech has quirks. \n\nFor now. \n\nProviding people with reference audio can help a bit. \n\nFor now.\n\nBut training about #deepfakes? Or listening again and again? Not so much.\n\nOur limited ability to reliably spot deepfaked speech = plenty of need to improve automated #deepfake detectors.\n\nAs deepfakes proliferate, finding ways to augment human detection (we're probably still better at spotting contextual quirks) will be key.",
"sig": "3da7afcb4f04b4c9e19a7681f835eecaaa80a86f407ab9ed0a0b61d1c659d81708c18900fbe865fbb963bb912c7d8663913ba80af38f9d08c545f0fc598b4f13"
}