Ask a navigation app to read a street name aloud and you hear text-to-speech at work. The system takes written characters and produces a waveform a human can understand. Early versions sounded robotic because they stitched together recorded fragments. Modern neural systems generate the audio directly, and the gap between synthetic and human speech has narrowed sharply.

The pipeline usually starts with text normalization: expanding "Dr." to "Drive" or "doctor" depending on context, converting numbers and currency into spoken words. Next comes linguistic analysis, which predicts pronunciation, stress, and intonation. Finally, an acoustic model and a vocoder turn those features into sound.

Where it shows up

• Screen readers for blind and low-vision users
• Voice assistants and smart speakers
• Audiobook production at scale
• Accessibility tools for people with speech impairments

Quality varies by language. English and Mandarin have enormous recorded corpora; many African and indigenous languages have almost none, so synthetic voices for them sound stilted or simply do not exist. Voice cloning adds another layer of concern. A convincing replica of a person's voice can be misused for fraud, which is why some vendors now embed watermarking or require consent recordings.