Speech Recognition Techniques Help Predict Volcanoes’ Behavior

Speech Recognition
In the last decade, the application of machine learning to pattern identification has been integral in speech recognition, but researchers are now using it to forecast volcanoes’ behavior. “Although these fields vary significantly in terms of context and source, the object of the analysis is the same – the study of their harmonics over time in search of patterns,” Zuccarello said.

The project’s main output will be a set of algorithms – set to be complete when the project ends later this year – and he hopes that they will be used widely in the scientific community to monitor volcanoes on a day-to-day basis.

“Speech and seismic signals share important properties,” said Dr. Guillermo Cortés, a specialist in signal processing and machine learning at the University of Udine in Italy. He ran a project called VULCAN.ears, which also used speech recognition technology to understand what volcanoes are saying.

Cortés and colleagues developed a real-time volcano monitoring system, which automatically detects and labels volcanic ‘events’ in the data streams coming from monitoring stations that detect seismic signals. This system then creates catalogues of activity in order to find patterns of behavior.

Dr. Roberto Carniel, a geophysicist at the University of Udine and the project’s scientific supervisor, says: ‘The arrival of machine learning and applied deep-learning techniques is uncovering new solutions for old problems. (Now) it is easier to mix results from several monitoring areas involving the study of seismic signals, infrasonic signals, magnetic signals, geochemical analysis of gases and fluids, deformation, thermal and video cameras, to produce more robust and reliable predictions.’

The team developed a volcanic seismic recognition system based on supervised machine learning, in which they analyzed data that had already been labelled by other experts, teaching the software to identify volcano events such as volcanic tremors, ashfall, or explosions within the volcano. This approach is similar to finding words in a conversation, labelling their parts of speech and finding the patterns of language unique to each volcano.

This is a break from the classic methods for building catalogues of volcano behavior, Cortés says. These methods involve the automatic detection of events and manual classification by experts. “Usually they perform this task on a daily basis, which could be too slow in a situation involving a population at risk due to an unexpected eruption,” he said.

Time can be of the essence when it comes to volcanoes, particularly in the event of ashfall, collapses and landslides, he says. In those cases, ‘the detection and classification in real-time operation is critical’ in order to reduce decision-making time if nearby communities need to be evacuated.

Cortés’ ultimate aim was to develop a system that is universal and volcano-independent that could be easily embedded at any volcano observatory. To build this, the researchers have created a universal database from dozens of volcanoes around the world and used their machine-learning techniques to build universal models. A preliminary version of this is available online.

However, forCarniel, what’s important now is that volcanic observatories around the world take the work forward. “They are the real key to advancing the volcano-independent idea, installing the volcanic seismic recognition system in their own observatories, sharing resources, and giving valuable feedback,” he said.

These observatories are, after all, the front line of countries’ efforts to protect their citizens from the volcanoes within their borders – and scientists need to be able to hear volcanoes’ whispers to predict when they are going to start shouting.

This article is published courtesy of Horizon, the EU research and innovation magazine