💬 NLP

Quantum Vision Theory Applied to Audio Classification for Deepfake Speech Detection

This paper proposes Quantum Vision (QV) theory, a quantum-inspired framework that transforms audio spectrograms into information waves to enhance deep learning models, demonstrating that QV-based CNNs and Vision Transformers significantly outperform standard architectures in detecting deepfake speech on the ASVSpoof dataset.

Original authors: Khalid Zaman, Melike Sah, Anuwat Chaiwongyenc, Cem Direkoglu

Published 2026-04-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Khalid Zaman, Melike Sah, Anuwat Chaiwongyenc, Cem Direkoglu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to catch a master forger who can perfectly mimic your voice. You have a security system (a computer) that listens to audio recordings and decides: "Is this a real human, or is it a robot?"

For a long time, these security systems have looked at audio like a photograph. They take a sound wave and turn it into a flat, static image called a spectrogram (a visual map of sound frequencies over time). The computer then looks at this "photo" to spot the forgery.

The Problem: A photograph is a "collapsed" version of reality. It shows you what the object looks like right now, but it loses all the hidden potential and the subtle "vibrations" that existed before the picture was taken. In the world of deepfakes, forgers are getting so good that they can make these "photos" look almost identical to real ones.

The New Idea: "Quantum Vision"

This paper proposes a radical new way of thinking, inspired by Quantum Physics.

In quantum physics, there's a famous concept called Particle-Wave Duality.

The Particle: When you look at an electron, it acts like a solid dot (a particle). It's fixed and definite.
The Wave: When you don't look at it, it acts like a ripple in a pond (a wave). It exists in many places at once, carrying a huge amount of hidden information about where it could be.

The Analogy:
Imagine you are trying to identify a person in a crowd.

Old Method (Standard AI): You take a single, frozen photo of the person. You look at their face and say, "That's Bob." If the forger puts on a mask that looks exactly like Bob's photo, you get fooled.
New Method (Quantum Vision): Instead of just taking a photo, you imagine the person as a living, breathing wave of energy before you even snap the picture. You analyze the "ripples" of their movement, the way their voice might vibrate, and all the possibilities of who they are. Even if the forger copies the face (the particle), they can't perfectly copy the wave (the hidden energy and potential).

How the Paper Does It

The researchers built a special "magic lens" called a QV Block (Quantum Vision Block). Here is how it works in simple steps:

The Input: They take the audio recording and turn it into a standard spectrogram (the "photo").
The Transformation: Instead of feeding this photo directly to the AI, they pass it through the QV Block.
- Think of this block as a machine that takes the static photo and turns it into a dynamic, vibrating wave.
- It mathematically shifts the image slightly in different directions (up, down, left, right) and calculates the differences. This creates a set of "wave maps" that highlight the edges, boundaries, and hidden structures of the sound.
The Detection: The AI (either a CNN or a Vision Transformer) then looks at these wave maps instead of the original photo.
- Because the wave maps contain more "potential" information, the AI can spot tiny, subtle glitches that the forger missed. It's like seeing the vibration of a fake voice rather than just its shape.

The Results

The team tested this on the ASVspoof dataset, a giant library of real and fake voices. They compared their new "Wave" AI against standard "Photo" AI.

The Winner: The Quantum Vision AI won almost every time.
The Best Combo: The most powerful combination was using MFCC features (a specific way of breaking down sound) with the Quantum Vision CNN.
The Score: It achieved 94.20% accuracy and a very low error rate. This means it is much harder to fool than the old systems.

Why This Matters

Think of it like upgrading from a still camera to a 3D motion scanner.

Old deepfake detectors were like security guards looking at a 2D ID card. A really good forger could print a fake card that looked perfect.
This new Quantum Vision approach is like a guard who can see the ID card vibrating and check if the person's "energy" matches the card. Even if the card looks real, the vibration gives the forger away.

In a nutshell: By treating sound not just as a static picture, but as a dynamic "information wave," this new theory helps computers detect deepfake voices with much higher precision, making our digital world safer from voice scams.

1. Problem Statement

The rapid advancement of speech synthesis and voice conversion technologies has led to the proliferation of deepfake speech, posing significant threats to security, privacy, and trust in biometric authentication and financial systems.

Current Limitations: Existing deep learning approaches for deepfake detection typically treat audio signals as fixed, "collapsed" time-frequency representations (e.g., spectrograms, MFCCs) and process them directly using standard Convolutional Neural Networks (CNNs) or Vision Transformers (ViTs).
The Gap: These conventional methods may discard latent information contained in the original signal by treating the spectrogram as a static image, rather than exploring the signal's underlying dynamic characteristics. The authors argue that current models focus heavily on classifier architecture design while neglecting the fundamental representation of the input data itself.

2. Methodology: Quantum Vision (QV) Theory

The paper proposes Quantum Vision (QV) Theory, a novel framework inspired by the particle-wave duality in quantum physics.

Core Concept

Particle-Wave Duality: In quantum physics, unobserved particles exist as probability waves containing all possible information, collapsing into a definite state only upon measurement.
Application to Audio: The authors propose that audio spectrograms should not be treated merely as static "collapsed" images (particles). Instead, they should be transformed into information waves that preserve richer, latent characteristics of the data before being fed into a neural network.

The QV Block (Technical Implementation)

The core innovation is the QV Block, a deep learning module that converts standard spectrogram images into quantum-inspired wave functions.

Basis Wave Generation: Inspired by Schrödinger's wave equation and quantum numbers ( $n, l, m$ $n, l, m$ ), the block generates basis wave functions by calculating the difference between the original spectrogram image $I(x,y)$ $I (x, y)$ and its shifted versions.
- For spatial directions $x$ and $y$ , shifts of $m = \pm 1, \pm 2$ are applied.
- Formula: $\psi_{x,m} = I(x+m, y) - I(x, y)$ (and similarly for $y$ ).
- This subtraction emphasizes boundaries and transitions, capturing spatial position information.
Superposition and Non-linearity:
- The basis waves are combined via linear and non-linear superposition.
- The paper utilizes Convolutional Neural Networks to learn the optimal non-linear combination of these basis waves.
- The process involves multiple convolutional layers with ReLU activations, generating 128 distinct wave-function representations for a single input spectrogram.
Integration: These 128 wave feature maps replace the original spectrogram image as the input to the downstream deep learning models.

Model Architectures

The QV block is integrated into two state-of-the-art architectures:

QV-CNN: A sequential CNN (QV-CNN-Heavy) with six convolutional layers, batch normalization, max-pooling, and ReLU activations.
QV-ViT: A Vision Transformer (ViT-8/8) where the 128 wave feature maps serve as input tokens, utilizing multi-head self-attention to capture global context.

3. Experimental Setup

Dataset: ASVspoof 2019, a standard benchmark for audio spoofing and deepfake detection.
Input Features: Three types of time-frequency representations were tested:
1. Short-Time Fourier Transform (STFT)
2. Mel-spectrograms
3. Mel-Frequency Cepstral Coefficients (MFCC)
Preprocessing: Audio was resampled to 16 kHz. Features were resized to $32 \times 32$ pixels for uniformity.
Metrics: Classification Accuracy and Equal Error Rate (EER).

4. Key Results

Extensive experiments demonstrated that QV-based models consistently outperformed their non-QV counterparts across all feature types and architectures.

Best Overall Performance:
- QV-CNN with MFCC features achieved the highest robustness with 94.20% accuracy and an EER of 9.04%.
- QV-CNN with Mel-spectrograms achieved the highest raw accuracy of 94.57% (EER 10.84%).
Comparison with Baselines:
- Standard CNNs and ViTs showed lower accuracy and higher EERs across all configurations.
- For example, the standard CNN with MFCC achieved 91.67% accuracy, while QV-CNN improved this to 94.20%.
- The QV-ViT model significantly reduced the EER of the baseline ViT (e.g., from ~24% to ~10% on MFCC).
Feature Analysis:
- MFCC and Mel-spectrograms proved more discriminative than STFT when combined with QV theory.
- The QV mechanism was particularly effective at enhancing cepstral features (MFCC), suggesting it successfully extracts subtle spoofing artifacts.
Robustness: The QV models showed consistent performance across different batch sizes (8, 16, 32, 64), with batch size 64 generally yielding optimal results.

5. Key Contributions

Novel Perspective: Introduced Quantum Vision (QV) Theory to the domain of audio classification, shifting the paradigm from processing static spectrograms to processing "information waves."
The QV Block: Proposed a specific architectural module that mathematically transforms spectrogram images into quantum-inspired wave functions using spatial shifting and convolutional superposition.
State-of-the-Art Performance: Demonstrated that integrating QV blocks into CNNs and ViTs yields new state-of-the-art results on the ASVspoof 2019 dataset, outperforming existing deepfake detection methods (including CQCC-GMM, LCNN, and standard Transformers).
Cross-Modal Validation: Successfully extended a theory originally designed for image recognition (visual objects) to the audio domain (speech spectrograms), suggesting a unified framework for multi-modal perception.

6. Significance and Future Directions

Impact on Deepfake Detection: The results confirm that rethinking the representation of data (from collapsed images to information waves) is as critical as designing better classifiers. This approach offers a new pathway to improve the robustness of anti-spoofing systems.
Quantum-Inspired Learning: The paper validates that concepts from quantum physics (wave functions, superposition) can be effectively simulated in classical deep learning to capture richer feature representations.
Future Work: The authors plan to extend QV theory to larger-scale datasets, other audio representations, and multimodal deepfake detection frameworks (combining audio and video).

In conclusion, this paper establishes that Quantum Vision Theory is a highly effective tool for deepfake speech detection, with the QV-CNN model using MFCC features currently setting a new benchmark for accuracy and error rate reduction in the field.