Synthetic speech

The neutral umbrella term for every kind of machine-made speech. It is the word to reach for when you want to describe how a recording was produced without also implying why.

Why careful writing prefers it to “AI voice”

“AI voice” is a phrase from product marketing, and it points in at least three directions at once. It can mean the voice an assistant answers you in. It can mean a voice sold as a product. It can mean a voice stolen from someone. A sentence built on it usually needs a second sentence to explain which one was meant.

“Synthetic” does the work in one word: the audio was generated. Nothing else is asserted. That precision matters most in the places where the description will be read adversarially — an incident report, a disclosure line under a video, a note to a regulator, a paragraph in a witness statement. A claim that a recording is synthetic can be examined. A claim that it is “an AI voice” invites an argument about what that means before anyone gets to the evidence.

The term also stretches backwards without embarrassment. Speech has been generated by machine since long before neural networks, by stitching together recorded fragments or by modelling the vocal tract directly. Those outputs were synthetic speech too. Calling everything “AI” quietly asserts a technique the writer usually has not established.

Where the boundary sits, and why the edge case is the dangerous one

The line is drawn at generation. If a model produced the waveform, the speech is synthetic. If a microphone captured a person and software was then applied to the result, it is edited human speech — however misleading the edit.

That distinction is not pedantry, because the two categories break detection in different directions. A generated clip may carry signatures of the system that made it. An edited genuine recording carries none, since every sample in it came from a real vocal tract. Someone can trim a sentence out of the middle of a real recording, splice two answers together, drop the question that was asked, and produce something thoroughly dishonest that any generation detector will pass without hesitation.

So a page like this one has an uncomfortable corollary: proving a recording is not synthetic tells you almost nothing about whether it is honest. That is a separate enquiry, and an older one. It belongs to audio forensics, which asks about originals, edits, containers and chain of custody rather than about generation.

A concrete case

A regional newsroom receives a thirty-second clip in which a council official appears to accept a payment. Two competing explanations fit the audio: the clip was generated by a model trained on the official’s public speeches, or the words are genuinely his, recorded across several meetings and reassembled.

Only the first is synthetic speech. The second is an edit. A detector can address the first and is close to useless against the second, which is why the newsroom’s first question should be about the original file and where it came from, not about the waveform. Our note on recordings produced in disputes works through that order of operations.

Commonly confused with: synthetic media, and with deepfake audio

Synthetic media is the parent category — generated images, video, music and speech together. It is a useful label for policy and platform rules, and a poor one for technical work, because the detection problem in each medium is almost entirely different. Nothing learned from spotting a generated face transfers to a voicemail.

Deepfake audio is the popular word for roughly the same thing, but it carries an accusation of deceptive intent that synthetic speech does not. An audiobook narrated by a licensed synthetic voice is synthetic speech and is not a deepfake. Using the words interchangeably means describing publishers, accessibility tools and localisation studios with vocabulary built for fraud.

How Truthring uses the term

Our product pages are named for the words people search with — AI voice detector, deepfake audio detection — and our method pages use this one, because a description of what an engine measures should not carry an implied motive. What the analysis looks for is evidence of generation. It has no view on why a clip was generated, and cannot acquire one from the audio.

That also shapes how a result should be read. Likely synthetic is the stronger verdict, because it rests on positive evidence that something in the file matched a generation signature. Likely human is weaker, because it can equally mean the evidence was destroyed on the way — a re-compressed forward through two messaging apps can leave a generated clip looking clean. The methodology sets out both directions.

Questions this term raises