AI voice tech is now so good you almost certainly won't be able to distinguish between real and deepfake voices - here's what you need to know (www.techradar.com)

A new study from Queen Mary University of London shows AI-generated speech has reached human-level realism: listeners could not reliably distinguish between real recordings, voice clones trained on a few minutes of someone’s speech, and voices produced by large‑language‑model–driven TTS systems that weren’t tied to any specific speaker. The team found cloning can be done quickly, cheaply and with minimal expertise, and—strikingly—participants sometimes rated synthetic voices as more dominant or even more trustworthy than real human voices. For the AI/ML community this signals both opportunity and urgency. Technically, it confirms rapid improvements in neural vocoders, speaker encoding and LLM‑conditioned speech generation that remove the old “synthetic” artifacts; practically, it means detection and provenance methods (robust watermarks, authenticated metadata, forensic classifiers) must evolve to keep pace. Benefits exist for education, accessibility and personalized interfaces, but the ease of cloning from short samples heightens risks around consent, impersonation and fraud—prompting calls for better technical safeguards, policy, and dataset/annotation standards to govern voice identity and misuse.