How Computers Learn to Sound Human
A computer engineer explains how text-to-speech technology has evolved to sound more natural and human-like through artificial intelligence.
Computer engineer Tam Nguyen explains the process by which digital assistants like Siri and Alexa have achieved human-like voices. The evolution of text-to-speech technology, which simulates human speech production, has led to increasingly natural-sounding digital speech.
The technology works by shaping sound waves. Computer software converts text into phonemes, the smallest units of sound in speech, which are then assembled into words and sentences. Early attempts at machine speech began in the 1700s with mechanical devices producing crude sounds. The first electronic speech synthesizers emerged in the 1930s, but computer-generated speech was initially stiff and robotic.
Modern systems utilize machine learning, a form of artificial intelligence. Programs are trained by analyzing thousands of hours of human speech, including variations in breathing, emotion, and pitch. This allows for the replication of subtle nuances and human characteristics in spoken language.
Advanced AI can even learn to mimic an individual's unique speech patterns from a short audio sample. While text-to-speech technology offers significant benefits, such as aiding the visually impaired or those unable to speak, it also presents risks of misuse, including realistic voice deepfakes. Researchers are developing tools to identify these fabricated voices.