How Truthring detects synthetic speech
Truthring runs three analyses on a submitted recording: whether the audio carries the signature of a real microphone in a real room, whether it matches a known generator, and how the voice behaves at conversational edges. Each result names the model version that produced it and carries a reference code.
Every result Truthring returns comes from the pipeline described on this page. It is versioned and dated, so a report issued last month can be checked against the model that actually produced it.
The short answer. Truthring measures a recording against signatures left by known speech-generation systems, and against a model of how real telephone audio degrades. It returns a probability with a stated confidence — not a proof. When the signal is insufficient it returns unclear rather than guessing.
What Truthring measures
Speech generators leave traces. Some are artefacts of how the audio is synthesised; others are absences — the small irregularities of a real vocal tract and a real room that a model does not reproduce. Truthring examines the following. [VERIFY: keep only the categories your engine genuinely uses. Delete the rest — an unused entry here is the claim that gets you caught.]
- Spectral structure. Distribution of energy across frequency bands, and discontinuities where a vocoder has reconstructed a band.
- Prosody. Rhythm, stress and intonation contours across the clip, and whether variation falls inside the range of natural speech.
- Vocal-tract consistency. Whether formant behaviour stays coherent with a single physical speaker across the recording.
- Generation artefacts. Signatures characteristic of specific synthesis architectures.
- Splice and concatenation. Boundaries where separately generated segments have been joined.
- Codec and compression history. What the file has been through — and whether that history is consistent with the claimed source.
- Background and room. Whether ambient noise behaves like a real acoustic space.
The pipeline
| Stage | What happens |
|---|---|
| 1 · Intake | File accepted, format validated, duration and channel count read. Rejected if outside supported formats or 25 MB. |
| 2 · Conditioning | Resampled to a common rate, normalised, and segmented. Codec history recorded before any processing that would erase it. |
| 3 · Quality assessment | Signal-to-noise, compression depth, speaker count and usable duration measured. This determines the confidence ceiling for the clip. |
| 4 · Feature extraction | The measurements described above are computed across the segmented audio. |
| 5 · Inference | Features evaluated against generator signatures and against the phone-degradation model. |
| 6 · Attribution | Where the signature matches a system we hold, it is named. Where it does not, the result reads unknown generator — never a guess. |
| 7 · Confidence | Raw score adjusted down by the quality ceiling from stage 3. A strong signal in a poor recording is reported as a weaker result, because it is one. |
| 8 · Verdict | One of three states, with the reasons stated in plain language. |
Median analysis time [VERIFY: not measured]. Audio will be deleted once the final stage completes. [VERIFY: retention not yet set] A one-way SHA-256 hash is retained so a report can be tied back to the exact file.
What each verdict means
Generation signatures were found, and the recording's quality was sufficient to trust them. The confidence number states how strongly. This does not establish who created the clip, when, or why.
No generation signature was found and the recording behaves like captured speech. This is the weaker of the two positive verdicts: absence of a known signature is not proof of authenticity, and a generator we have not seen would produce this result.
The clip did not carry enough usable signal to answer. Too short, too compressed, too noisy, multiple overlapping speakers, or heavily edited after generation. This is a real verdict, not a failure — and it is the one most detectors will not give you.
Why phone audio is the hard case
Most published detection accuracy is measured on clean studio recordings. That is not the audio anyone actually brings to a detector. A voicemail has passed through a codec designed to discard exactly the high-frequency detail that detection depends on, often twice, and usually with background noise on top.
Truthring's degradation model is trained on audio that has been through that path. It is why our accuracy figures on clean audio are lower than some competitors quote, and why our figures on telephone audio are higher. The phone-audio note sets this out in detail.
What this does not do
Truthring returns a probability, not proof. It cannot establish that a recording is authentic, identify who produced a clip, or determine intent. A detector can miss a well-made clone and can flag genuine speech in a poor recording.
For decisions involving money, employment, legal claims or safety, treat the result as one data point and verify through a channel you already trust.
The full limitations page lists every condition under which we know the result is unreliable.
Version and changes
This page describes Truthring Detection Engine v2.4, in service since [VERIFY: date]. The model is retrained monthly as new generation systems appear.
Every report carries the engine version that produced it. The changelog records each revision with its date and what changed, so a result issued under an earlier version can still be interpreted correctly.
[VERIFY: Written by — real name, real role.] Reviewed [VERIFY: date]. Method version 2.4.
Corrections and challenges to anything on this page: method@aivoicedetctor.com. We publish corrections rather than quietly editing.
Questions
How does Truthring detect AI voices?
It runs three analyses: the recording chain, which asks whether the audio was captured by a real microphone in a real room or rendered directly to a file; the generator signature, which looks for regularities left by a specific synthesis system; and prosody under stress, which examines how the voice behaves at interruptions and restarts. The pipeline is versioned and every result names the version that produced it.
Does Truthring analyse the words that were spoken?
No. The analysis is acoustic. It does not transcribe speech, assess what was said, or evaluate whether a claim in a recording is true. It estimates whether the speech was produced by a machine.
Can the method be independently reproduced?
The pipeline, the verdict definitions and the versioning scheme are published here so the approach can be argued with. Quantitative results are not yet available, and Truthring has had no independent third-party audit.
Reviewed