Your ears can be fooled.Veracust won't.
AI can imitate a voice. But can it reproduce the physical world hidden inside a sound? Veracust explores a different way to detect synthetic speech.
A familiar voice
is no longer proof.
Generative voice tools can reproduce tone, cadence and identity with startling realism. In a phone call, the human ear may not have enough evidence to tell what is genuine.
Veracust is exploring a complementary signal-based approach: examine the acoustic traces of how a voice was recorded and produced, not only what was said or how it sounded.
Synthetic voices
Modern models can create convincing speech from surprisingly little source audio.
Physical signals
Real recordings carry room tone, microphone behaviour and subtle biological variation.
A different question
Can the signal reveal clues that the voice itself conceals?
Three signal layers.
One richer picture.
The prototype combines deterministic digital signal processing with acoustic and phonation analysis. Each layer contributes evidence; no single feature is treated as a magic answer.
Inspect the digital edges.
A first-pass scan looks for known container and synthesis artefacts before deeper signal analysis.
Listen for the room around the voice.
Natural recordings often contain ambient room energy, microphone noise and channel characteristics. Unusually clean pauses can be one clue among many.
Study the mechanics of speech.
Pitch variation, jitter and inverse-filtered phonation features help examine the subtle irregularities associated with biological vocal production.
Engineering for
real-world audio.
The prototype combines deterministic digital signal processing with a lightweight API. The work is focused on the messy conditions where voice authenticity matters: compressed audio, phone channels and generated speech.
VIEW THE SOURCEEdge-aware processing · API-first integration · privacy-conscious handling
Trust is earned
through testing.
Detection performance depends on the dataset, model families, recording conditions and evaluation method. Veracust is an evolving prototype; experimental scores should be read in the context of the tests that produced them, not as a guarantee of fraud prevention or regulatory compliance.
Built from
a personal reason.
It started with a voice I thought I could trust.
I was a high school student watching AI advance at an incredible pace — and seeing the new risks that came with it. Then I encountered a scam involving an AI-cloned voice. It made the threat feel real: something as personal as a familiar voice could be imitated and used to manipulate someone.
I didn't want to just worry about the problem. I wanted to try building something that could help address it.
So I started Veracust. No funding. No team. Just the technology, information and time I could access. I began learning, experimenting and building an acoustic authenticity engine from the ground up.
Veracust is still evolving. But the reason behind it hasn't changed: make it harder for a synthetic voice to pass as someone real.
Question what
sounds real.
Explore the prototype, inspect the code, or get in touch about research and potential design partnerships.