English Español Français Deutsch
← Back to the list
Speech and multimodal Methodologists Directly upstream of ai100's methodology

Junichi Yamagishi

Whether human listeners — or automated detectors — can tell AI-generated speech apart from real human speech, and how that distinguishability changes as speech synthesis improves.

Yamagishi has been the primary organizer of ASVspoof, the international challenge that tracks how well automated systems can detect spoofed or synthetic speech, since the first edition in 2015. The accumulated record from the challenge series is one of the few longitudinal datasets in AI-content-detection that covers an entire technology generation — from early concatenative synthesis to current neural voice cloning. He co-organizes the parallel VoiceMOS challenge for evaluating synthetic-speech quality, which makes him uniquely positioned on both sides of the question: how good has synthesis gotten, and how detectable does that goodness remain. For ai100, as AI engines move toward voice output, this is the methodology stack for evaluating what it sounds like when an LLM speaks.

Worth following when
you need to evaluate AI-generated speech — either its quality or its detectability — and want longitudinal benchmark methodology that has tracked the technology through multiple generations.
Topics
anti-spoofing for voice (ASVspoof); synthetic-speech quality evaluation (VoiceMOS); the longitudinal tracking of speech-synthesis progress through annual challenges.
Key works
ASVspoof challenge series organization (2015 onward); VoiceMOS challenge co-organization; long body of work on speech synthesis at NII and Edinburgh CSTR.