Not sure what
his source is, but from the article:
Andrew Clement, a computer scientist and expert in surveillance studies, has been mapping the NSA’s warrantless wiretapping activities since before Snowden’s disclosures. He strongly believes the agency would not be restrained in their uses of speaker recognition on U.S. citizens. The agency has often chosen to classify all of the information collected up until the point that a human analyst listens to it or reads it as metadata, he explained. “That’s just a huge loophole,” he said. “It appears that anything they can derive algorithmically from content they would classify simply as metadata.”