May 2025. I gave a talk at The One About AI Agents and Imposters — a community event at Lorong AI (WeWork, 22 Cross St) in Singapore. The session covered voice synthesis, deepfake detection, and the surprisingly thin line between party trick and large-scale fraud.
the talk
The title was Real-time Voice Cloning in the Age of GenAI. I went in at technical level 200–300: enough to explain how the technology actually works, without losing the room.
The structure was simple:
- What is a deepfake? — ML-generated synthetic media. Covers audio, video, images, text. Not new, but the barrier to entry has collapsed.
- Live demo — cloned a voice in the room in real time. Twice. The second time convinced people who weren't convinced the first time.
- So what? — why this matters beyond the "cool tech" angle. Voice fraud, social engineering, impersonation attacks.
- What's stopping mass fraud at scale — for now — this is the part I find most interesting.
the temporary reprieve
The reason we haven't seen a wave of real-time AI voice fraud isn't that the technology doesn't exist. It's that the pipeline isn't seamless yet:
- Real-time speech-to-text (STT) exists and is good
- High-quality text-to-speech (TTS) exists and is good
- But real-time speech-to-speech (STTS) — where you speak, the AI transforms your voice into someone else's voice, with matching cadence and latency, in real time — is still brittle
That gap is the moat. It's closing.
are we doomed?
Not yet. There are still enough imperfections in cloned voices — artefacts, unnatural cadence, response latency — that a sufficiently paranoid listener can catch them. And there are practical countermeasures: verification challenges, call-back protocols, out-of-band confirmation.
But the honest answer is: the current state is a temporary window, not a permanent defence. The tech will get better faster than most organisations' security posture will adapt.
The talk ended at "what to do about it" rather than "how scary this is." That felt like the right note.