Berkeley study finds listeners mistake AI voice clones for their sources

Ordinary listeners struggled to distinguish AI voice clones from real recordings in a University of California, Berkeley study, correctly identifying cloned clips 60.8% of the time. In a separate identity task, participants judged a real speaker and that person’s clone to be the same voice at a median rate of 83.3%, suggesting that synthetic speech can often pass as its source speaker.

The experiment used recordings from 220 speakers and 604 listeners. The naturalness test asked participants to decide whether individual clips were real or AI-generated, while the identity test asked whether two voices belonged to the same speaker. Listeners identified real clips correctly 67.4% of the time, and recognized two genuine recordings of the same speaker as sharing an identity at a median rate of 100%.

The researchers created the clones with ElevenLabs’ Instant Voice Cloning API, using each speaker’s first two prompts from the DeepSpeak dataset. When participants heard two different but similar real speakers, they gave a same-identity answer at a median rate of 11.1%, underscoring the contrast with the clone results. Berkeley said the findings showed that people could not consistently identify AI-generated voices and that the clones were largely through the “uncanny valley” for human listeners.

That leaves detection systems to seek signals that may not be apparent to the ear. Jitter measures cycle-to-cycle variation in a voice’s fundamental frequency, while shimmer measures variation in amplitude; research on speaker recognition has found the measures can provide complementary information. But Voice Over Herald reported that no individual acoustic measurement catches every clone, and detectors trained on one company’s tools often transfer poorly to audio made by other providers. Telephone calls further complicate analysis by stripping out much of the acoustic detail needed for forensic examination.

Watermarking is another route, placing an identifying mark in generated media so its origin can later be checked. Some AI-audio companies, including Google, embed inaudible watermarks that travel with shared files, though altered recordings can have those marks erased and some open-source tools make watermarking optional.

Research on a proposed system called VoiceMark reported more than 95% watermark-detection accuracy after zero-shot voice cloning, compared with about 50% for existing methods in its tests. The Federal Trade Commission has nonetheless cautioned that fraudsters may remove or distort watermarks, and said there is no single solution to the harms posed by AI-enabled voice cloning. It identifies prevention and authentication, real-time monitoring, and post-use evaluation as separate points for intervention.

Berkeley’s DeepSpeak project is intended to support the development of detection techniques; its second iteration contains footage from 500 people aged 18 to 75, including audio, face-swap, avatar and lip-sync deepfakes.