The landscape of AI speech generation and deepfake detection

Authors: Simon King

Published: 2026-11-18 13:55:38+00:00

Comment: 80 pages including definitions of all acronyms, a glossary of technical terms, and a bibliography

AI Summary

This report provides a comprehensive overview of AI speech generation and deepfake detection, fields that have advanced to a point where synthetic speech can deceive human listeners. It covers the historical development, current state-of-the-art technologies, and the challenges in both generating highly natural speech and detecting artificially created audio.

Abstract

Synthetic speech has reached a level of naturalness that can deceive listeners into believing they are hearing natural speech. This opens up the possibility of using synthetic speech deepfakes to mislead both human listeners and automated systems. One response to this threat is to attempt to automatically detect deepfakes. This report offers the reader a self-contained overview of the landscape of synthetic speech generation and speech deepfake detection. It is intended to be accessible to a general audience and therefore includes brief histories of both fields, definitions of all acronyms, a glossary of technical terms, and an extensive bibliography for those who wish to read more.


Key findings
The naturalness of synthetic speech has reached a level capable of deceiving human listeners, posing significant threats. Deepfake detection is a less mature field that heavily relies on detecting low-level waveform artifacts, making generalization to novel deepfake methods challenging. Future research will focus on explainability, attributing deepfakes to their source, and developing adversarial methods to evade detection.
Approach
This paper is a survey and review article rather than presenting a novel approach. It analyzes existing methods for speech deepfake detection, categorizing them by features (classical, deepfake-specific, SSL-derived, meta-features), direct waveform input, and ensemble classifiers. It also discusses 'one-class' classification as an anomaly detection approach.
Datasets
UNKNOWN
Model(s)
UNKNOWN
Author countries
United Kingdom