AI Voices

An AI voice is a speaking voice that no human has spoken. A computer program receives a written text and computes an audio recording from it. The result today often no longer sounds like a robot, but like a person with breathing pauses, emphasis, and mood. This became possible through programs that had previously analyzed thousands of hours of real speech recordings. From these examples, they learned what speech sounds like. Some systems can even recreate the voice of a specific person if they receive only a few minutes of recorded material from them.

Why voices are suddenly copyable at will

A voice was long considered something personal and hard to forge. On the phone, you recognize friends and family instantly. It is precisely this sense of security that has now become fragile. Anyone who has a short voice message or a video of someone can generate a passable copy from it.

Criminals use this for a modern version of the so-called grandchild scam. They call, sound like a relative, and urgently ask for money. Companies have also suffered losses because a faked boss’s voice ordered a bank transfer. Banks that wanted to identify customers by their voice have had to rethink their procedures.

At the same time, a lot of money is at stake. Voice actors and narrators earn their living recording commercials, audiobooks, and dubbed versions. When software can do this in minutes, it transforms an entire profession. That is why there is currently intense debate about who legally owns a voice and what it may be used for without permission.

From letter to sound wave

The generation process runs in several steps. First, the program breaks the text down into sounds, i.e., into the smallest audible units of speech. In doing so, it must decide how abbreviations and numbers are spoken. “12 €” becomes “twelve euros,” not “one two.”

After that, the model calculates the sound. For each sound, it determines how long it lasts, how high the pitch is, and how loud it is spoken. This is the part that determines naturalness. A question must rise at the end, a sentence ending must fall. Finally, another component converts these specifications into an actual audio track that can be played back.

When recreating a specific person’s voice, an additional step is added. The system extracts a kind of voice profile from a short sample recording: timbre, pitch, speaking tempo. It attaches this profile to every new text. For this, the model usually is not retrained at all; it simply receives the voice as additional information alongside the text. That is why a voice clone today takes minutes instead of weeks.

Where AI voices are already speaking

Most often, people encounter them without noticing. Navigation announcements, train station announcements, and the read-aloud function on smartphones almost always come from software. Voice assistants like Siri or Alexa also respond with a computed voice. Many audiobooks from major providers are now produced without a human narrator.

In the news, AI voices usually appear in connection with deepfakes. This is the term for faked media content designed to seem real. Faked calls from politicians shortly before elections and songs in which well-known musicians sing something they never actually recorded have become well known.

A common misconception is that fakes can be recognized by their sound. With good systems, this is barely possible anymore. It makes more sense to be wary of the situation itself: anyone who urgently demands money over the phone should be called back on a known number. Some providers also embed inaudible watermarks into the audio track so that artificial recordings can later be technically verified.