Skip to content

Voice Conversion

Definition

A signal-processing or deep-learning technique that transforms the vocal characteristics of a source speaker's utterance to match a target speaker, while preserving the linguistic content. The input is real speech; the output is modified real speech with a different perceived identity.

Field
Speech signal processing, audio forensics
Input
Genuine recorded speech from a source speaker
Output
Modified real speech carrying a different perceived voice identity
Preserves
Linguistic content and timing of the original utterance
Contrast
Text-to-speech synthesis generates speech from text, not from a real recording

Common questions

How is voice conversion different from voice cloning?+

Voice conversion transforms an existing real utterance to sound like a different speaker, while voice cloning typically trains a text-to-speech model on a target voice and then generates entirely new synthetic utterances from arbitrary text, with no original recording as the input.

What artifacts do examiners look for to detect voice conversion?+

Common indicators include unnatural prosody or pitch contours inconsistent with the claimed speaker's known speech, spectral discontinuities at conversion boundaries, and a mismatch between vocal tract characteristics and formant patterns expected from the target speaker's physiology.

Why does voice conversion matter for evidentiary audio authenticity?+

Because the underlying speech content and timing are real, conversion can be harder to flag through content or timing analysis alone than fully synthetic audio, pushing examiners toward lower-level acoustic and spectral consistency checks.

Related terms

Anti-Spoofing Countermeasure (CM)
A classifier, also called a CM system, trained to output a score indicating the probability that a given audio segment is genuine...
ASVspoof
A recurring evaluation campaign and dataset series that benchmarks anti-spoofing countermeasures against corpora of genuine and spoofed utterances. Editions in 2015, 2017,...
Diffusion Model
A generative neural network architecture (Ho et al., 2020; Stable Diffusion, Rombach et al., 2022) that learns to reverse a noise-addition process...
Encoder-Decoder
A neural architecture where an encoder compresses an input into a compact latent representation and a decoder reconstructs an output image from...
Equal Error Rate (EER)
The point on a classifier's detection error tradeoff curve where the false accept rate equals the false reject rate. Lower EER indicates...
GAN
Generative Adversarial Network. A framework with two neural networks, a generator that creates synthetic data and a discriminator that tries to distinguish...
Latent Space
The compressed, lower-dimensional representation of data learned by a neural network's internal layers. Generative models sample from or navigate this space to...
NeRF (Neural Radiance Field)
A neural representation that encodes a 3-D scene as a continuous volumetric function, allowing novel viewpoints to be rendered. In talking-head systems,...
Neural TTS Cloning
A text-to-speech system that adapts to a target speaker using a short enrollment recording, generating new utterances in that speaker's voice from...
Tandem Detection Cost Function (T-DCF)
The primary evaluation metric in ASVspoof from 2019 onward. It measures the cost of errors when a countermeasure is integrated with an...

Explained in these topics

Your journey to becoming a forensic professional starts here.

Practice with mock tests, learn from structured notes, and get your questions answered by a global forensic community, all in one place.