Synthetic Intelligence (AI) constructed for speech is now decoding the language of earthquakes, Nvidia mentioned in a weblog put up, noting that researchers have repurposed an AI mannequin constructed for speech recognition to analyse seismic exercise, providing new insights into how faults behave earlier than earthquakes. A crew at Los Alamos Nationwide Laboratory used Meta’s Wav2Vec-2.0, a deep-learning AI mannequin initially designed to course of human speech, to review seismic alerts from Hawaii’s 2018 Kilauea volcano collapse. Their analysis, revealed in Nature Communications, reveals that faults produce distinct, trackable alerts as they shift—much like how speech consists of recognisable patterns.
Additionally Learn: Nvidia Accelerates AI Integration in Medical Imaging with MONAI
AI Listening to the Earth
“Seismic data are acoustic measurements of waves passing via the strong Earth,” mentioned Christopher Johnson, one of many examine’s lead researchers. “From a sign processing perspective, many comparable strategies are utilized for each audio and seismic waveform evaluation.”
By coaching the AI on steady seismic waveforms and fine-tuning it with real-world earthquake information, the mannequin decoded complicated fault actions in actual time—a job the place conventional strategies, like gradient-boosted bushes, typically fall quick. The challenge leveraged Nvidia’s GPUs to course of huge seismic datasets effectively.
“The AI analysed seismic waveforms and mapped them to real-time floor motion, revealing that faults would possibly ‘communicate’ in patterns resembling human speech,” Nvidia mentioned in a put up.
Additionally Learn: CES 2025: Nvidia AI Bulletins, Launches and Partnerships Throughout Industries
Can AI Predict Earthquakes?
Whereas the AI confirmed promise in monitoring real-time fault shifts, it was much less efficient at forecasting future displacement. Makes an attempt to coach the mannequin for near-future predictions — basically, asking it to anticipate a slip occasion earlier than it occurs — yielded inconclusive outcomes. Johnson emphasised that enhancing prediction would require extra numerous coaching information and physics-based constraints.
“We have to broaden the coaching information to incorporate steady information from different seismic networks that comprise extra variations in naturally occurring and anthropogenic alerts,” he defined.
“So, no, speech-based AI fashions aren’t predicting earthquakes but. However this analysis suggests they may sooner or later — if scientists can train it to pay attention extra fastidiously,” Nvidia concluded.
Additionally Learn: Nvidia and Companions Develop AI Mannequin to Predict Future Glucose Ranges in People
Meta’s Wav2Vec-2.0
Meta’s Wav2Vec-2.0, the successor to Wav2Vec, was launched in September 2020. It makes use of self-supervision and learns from unlabeled coaching information to boost speech recognition throughout quite a few languages, dialects, and domains. In keeping with Meta, this mannequin learns fundamental speech items to sort out self-supervised duties. It’s skilled to foretell the right speech unit for masked parts of audio.
“With only one hour of labeled coaching information, wav2vec 2.0 outperforms the earlier cutting-edge on the 100-hour subset of the LibriSpeech benchmark — utilizing 100 instances much less labeled information,” Meta mentioned on the time of its announcement.
