Text-to-speech

The oldest and by far the largest category of machine-made speech, and the one where a detection result is least likely to mean anything is wrong.

Detecting TTS is a different problem from detecting a clone

Both are synthetic speech, so both can carry generation signatures. What differs is how well characterised those signatures are, and what a match is worth once found.

A commercial TTS voice is used by an enormous number of people, unchanged, at scale. Its output is consistent, obtainable, and can be studied directly, which makes attribution to a particular system comparatively tractable. That is why we publish a page for each generator we track, from the widely used commercial systems to open-source models people run themselves.

A clone is closer to a one-off. It may come from a fine-tuned or locally modified system, built for one target and used a handful of times. There may be nothing to compare it against. Detection there leans harder on general traces of machine generation rather than on recognising a specific system, and attribution is more often reported honestly as unknown generator than named.

The error that costs someone something

Because TTS is ordinary, treating “synthetic” as a synonym for “suspicious” produces a specific and unpleasant harm: it lands on people who speak through a machine because that is how they speak.

Someone using a speech-generating device sends a voice note. A fraud queue flags it as synthetic, correctly. The system was right about the audio and the process that acts on the flag is wrong about the person. The same trap catches localisation studios, publishers who narrate with a licensed voice, and anyone whose employer records announcements through a synthesis tool.

The lesson is a design one. A synthesis verdict is a fact about a file, and it should never sit alone at the end of a workflow that decides something about a human being. Where a control has to be built, build it on a written verification procedure in which the audio is one input among several.

A concrete case

A charity publishes its annual report as audio, narrated by a licensed synthetic voice, and says so in the description. A supporter runs the file through a detector, gets likely synthetic, and posts that the charity has faked something.

Every part of that is technically accurate and the conclusion is still false. The detector reported generation, which was disclosed; the reader supplied the accusation. This is the routine failure mode of synthesis detection in public hands, and it is far more common than a missed clone.

Commonly confused with: voice cloning

The confusion runs in one direction: people say TTS when they mean any machine voice, including one built to impersonate a named person. The clean split is whose voice comes out.

TTS produces a voice designed to belong to nobody, or to a speaker who licensed it. Cloning produces the voice of a specific identifiable person. The mechanism can be nearly the same; the consent question is completely different, and it is the consent question that decides whether a result is worth acting on. A third case, voice conversion, is not TTS at all: it starts from speech rather than from text.

What a TTS result looks like from Truthring

Where the signature matches a system we hold, the report names it. Where it does not, it says unknown generator rather than guessing, because a wrong attribution is worse than none — it invites someone to go and accuse a company. The method changelog records when coverage of a new system is added, so an old report can be read against the engine that produced it.

None of that tells you whether the file was published honestly. A named generator plus a disclosure line is a legitimate audiobook. A named generator plus an urgent request for a bank transfer is something else. The audio is identical in both.

Questions this term raises