Resemble AI’s Detect-2B: A Leap Forward in Combating Deepfake Audio

Resemble AI has released Detect-2B, a next-generation AI audio detection model that can identify deepfake audio with 94% accuracy, marking a significant advancement in the fight against misinformation and erosion of trust in an era of increasingly sophisticated generative AI.

Key features of Detect-2B: The new model utilizes a series of pre-trained sub-models and fine-tuning techniques to analyze audio clips and determine whether they were generated using AI:

• The sub-models consist of a frozen audio representation model with an adaptation module inserted into its key layers, allowing the model to focus on artifacts that often distinguish real audio from fake ones without requiring retraining for each new clip.
• Detect-2B’s architecture is based on Mamba-SSM or state space models, which use a stochastic, probabilistic approach to better respond to different variables and capture dynamics in audio signals, even in poor-quality recordings.

Rigorous testing and evaluation: Resemble AI subjected Detect-2B to a comprehensive test set, including unseen speakers, deepfake-generated audio, and various languages:

• The model accurately detected deepfake audio in six different languages with an accuracy of at least 93%, demonstrating its robustness and adaptability.
• Detect-2B will be made available through an API, allowing integration into various applications to help identify and combat the spread of deepfakes.

Growing importance of deepfake detection: As generative AI capabilities continue to advance, the need for reliable detection tools has become increasingly crucial, particularly in light of the potential impact on the 2024 U.S. Presidential Elections:

• AI-generated voices and videos could facilitate the spread of misinformation and mislead voters, making tools like Detect-2B essential in identifying and proving deepfakes before they reach the public.
• Other companies, such as McAfee and Meta, are also developing solutions to detect AI-generated audio and add watermarks to help distinguish authentic content from deepfakes.

Ongoing research and development: While Detect-2B represents a significant leap forward in deepfake detection, Resemble AI acknowledges that their work is far from over:

• As generative AI capabilities continue to evolve, so must detection capabilities, and the company has several research directions planned to further improve Detect-2B.
• These efforts will focus on areas such as representation learning, advanced model architectures, and data expansion to stay ahead of the ever-advancing generative AI landscape.

The release of Resemble AI’s Detect-2B marks a crucial step in the ongoing battle against deepfakes and the erosion of trust in an increasingly AI-driven world. As the technology continues to evolve, the development and refinement of robust detection tools will be essential in maintaining the integrity of information and fostering public trust in the digital age.