Identifying phone models from audio recordings using multilinear discriminant analysis

Every smartphone leaves an invisible fingerprint on the audio it records. The microphones inside mobile phones, together with their analog-to-digital converters, amplifiers, and anti-aliasing filters, each introduce tiny, device-specific distortions into a recorded signal—distortions far too subtle for the human ear to detect, yet persistent enough to serve as forensic evidence. Now, a team of researchers from Algeria and the United Kingdom has developed a new machine learning system that can identify the exact model of mobile phone used to record a piece of speech with remarkable accuracy, reaching 99.60% on one benchmark dataset and 90.12% on a more challenging one. The work, published in Multimedia Tools and Applications, could give law enforcement and forensic analysts a powerful new tool for tracing audio evidence back to its source.

The research, led by Abdennour Alimohad of Bouira University and the University of Medea, along with Ammar Chouchane, El Ouanas Belabbaci, Abdelmalik Ouamane, and Fouad Khelifi of Northumbria University, addresses a long-standing problem in digital audio forensics. When investigators recover a voice recording from a crime scene—whether it is a threatening phone call, a ransom message, or a covert recording—they often need to know what kind of device produced it. Establishing that a recording came from a suspect’s specific phone model can help reconstruct the chain of events and corroborate other evidence. But doing so reliably is difficult, because the forensic “signature” of a device’s microphone chain is entangled with the characteristics of the speaker, the content of the speech, and the acoustic environment in which the recording was made.

The core idea behind source device identification is that each recording chain imprints characteristic artifacts on the captured signal. MEMS microphones used in smartphones differ from unit to unit in frequency response, noise floor, and nonlinear behavior. These imperfections act like a stochastic fingerprint embedded in every recording. Earlier approaches tried to capture this fingerprint directly from the audio signal using conventional acoustic features such as mel-frequency cepstral coefficients, or through spectral analysis of frequency bands where devices differ most. More recently, deep neural networks have been applied to the problem, learning discriminative patterns directly from raw audio or from time-frequency representations. The new study takes a different and visually inspired route: it treats the audio as an image and applies image-analysis techniques to extract the device fingerprint.

The conversion step is critical, and the researchers paid close attention to it. The audio signal is transformed into a Mel spectrogram, a two-dimensional representation in which one axis represents time, the other represents frequency mapped onto the perceptually motivated Mel scale, and pixel intensity encodes energy. The Mel scale compresses high frequencies and expands low ones in a way that mirrors human auditory perception, which means the resulting images concentrate informative structure in a compact visual form. The team describes their method for this audio-to-image conversion as an optimal one, tuned so that the subtle inter-device differences survive the transformation and become visible as texture patterns in the spectrogram. Once the audio is an image, the full arsenal of computer vision texture descriptors becomes available.

That is where the paper’s first major technical contribution comes in: a novel descriptor called the Local Phase Quantization with Enhanced Histogram Descriptor, or LPQ-EHD. Local Phase Quantization, originally developed for blur-insensitive texture classification, analyzes the local phase of the image’s short-term Fourier transform around each pixel. Because phase information is comparatively robust to blur and to small, spatially invariant distortions, LPQ can capture fine texture structure even when the image has been slightly degraded. In the context of spectrograms, this robustness is valuable, since recording conditions introduce exactly the kind of nuisance variation that could otherwise wash out the device fingerprint. The enhanced histogram component aggregates the quantized phase patterns across the image into a compact statistical signature, capturing both the distribution of local textures and their coarser spatial organization.

To push discrimination further, the researchers apply LPQ-EHD at multiple scales, producing Multiscale LPQ-EHD features that encapsulate texture information ranging from fine-grained local detail to broader structural patterns. This multiscale strategy matters because microphone artifacts do not live at a single frequency or time scale; they manifest differently across the spectrogram depending on the spectral characteristics of the device’s recording chain. Analyzing the spectrogram at several resolutions simultaneously allows the system to integrate evidence from all of these scales into a single, rich representation.

The second key contribution lies in how these features are classified. Rather than flattening the multiscale features into a single long vector and applying a conventional linear classifier, the authors organize the data as high-order tensors and apply a multilinear subspace projection technique they call Tensor Exponential Discriminant Analysis, or TEDA. Classical discriminant analysis finds a linear projection that maximizes the separation between classes relative to the within-class scatter, but when applied to vectorized data it can discard the intrinsic structure of multidimensional representations and suffer from the so-called small-sample-size problem, where the number of features exceeds the number of training samples. Tensor-based methods such as multilinear discriminant analysis preserve the natural structure of the data by learning a set of projections, one per tensor mode, directly on the original multi-way arrays. The exponential variant, drawing on a broader exponential framework for dimensionality reduction introduced in earlier work, replaces the scatter matrices in the discriminant criterion with their matrix exponentials. This has the effect of amplifying the contribution of discriminative directions and mitigating the small-sample-size issue, since the exponential weighting emphasizes small eigenvalues in the within-class scatter matrix more gracefully than the classical formulation.

The experimental evaluation was carried out on two datasets designed to test different aspects of the identification problem. The MOBIPHONE dataset represents the more controlled setting, and there the proposed system achieved an identification accuracy of 99.60%, meaning that out of one hundred recordings from previously unseen test sessions, the system correctly identified the source phone model in all but a fraction of one case. The Controlled-Conditions, or CC, dataset poses a tougher challenge, and there the system reached 90.12%. Both results surpass state-of-the-art competing methods reported in the literature, and the gap is particularly meaningful on the harder dataset, where nuisance variability from speakers, speech content, and environment would normally erode performance.

The comparison against prior work highlights why the combination of Mel spectrogram imaging, multiscale texture descriptors, and tensor discriminant analysis is effective. Earlier studies relied on hand-crafted acoustic features optimized for speech rather than for device discrimination, or on single-scale texture analysis of spectrograms that captured only part of the available structure. Deep learning approaches, while powerful, typically require large amounts of training data and can be sensitive to domain shift between recording conditions. The new pipeline sidesteps some of these issues: the LPQ-EHD descriptor is compact and computationally light, the multiscale representation is comprehensive, and the tensor exponential projection squeezes maximum class separability out of the feature space even with limited training samples per class. The result is a system that is both accurate and, by the standards of forensic work, robust.

The practical implications extend beyond identifying a single phone. In forensic casework, narrowing a recording to a specific device model can exclude suspects whose phones cannot have produced the audio, guide investigators toward which devices to seize and examine, and support chain-of-custody analysis when multiple recordings must be attributed. The technique could also serve authentication purposes, helping to verify whether an audio file plausibly originated from the device it is claimed to have come from—a growing concern in an era of cheap audio manipulation tools and synthetic speech. The same underlying idea of spectrogram texture as a device fingerprint extends naturally to related problems such as microphone identification, recording environment classification, and the detection of audio forgeries and spoofing attacks, fields that the authors note have been advancing rapidly.

There remain, of course, challenges ahead. Real-world forensic audio is often compressed, transmitted over lossy channels, resampled, or partially corrupted, and each of these degradations can alter the spectrogram texture the system depends on. The authors’ testing, while demonstrating strong superiority over existing approaches on both benchmark datasets, used publicly available data under conditions that, though varied, do not fully span the chaos of evidentiary recordings. Future work will likely need to confront noise, codec distortion, and adversarial attempts to mask a device’s fingerprint. The authors acknowledge that no dataset was newly released with the study, but they provide references to the publicly available datasets they used, allowing other groups to reproduce and extend their results.

The study was conducted by researchers at Bouira University, the University of Medea, the University Center of Barika, the University of Bejaia, the University of Biskra, and Northumbria University, with sponsorship from Algeria’s General Directorate for Scientific Research and Technological Development. It appears in Volume 85 of Multimedia Tools and Applications as article number 679, after a review process that began in November 2024 and concluded with acceptance in July 2026. As audio evidence continues to play an expanding role in investigations and courts, techniques of this kind—turning the invisible physics of microphones into visible, learnable patterns—may become standard instruments in the forensic toolkit, closing the gap between what a recording says and what it can be proven to reveal about where, how, and by whom it was made.

For now, the headline numbers speak for themselves: near-perfect identification under controlled conditions and strong performance in the face of realistic variability, achieved not by scaling up a neural network but by combining a classic perceptual representation with a carefully engineered texture descriptor and a mathematically elegant tensor projection. It is a reminder that in machine learning for forensics, the most effective systems are often those that respect the structure of the data—treating a sound, quite literally, as a picture worth a thousand words.