TL;DR
The authors introduce POLIPHONE, a new open-source dataset containing over 50 hours of audio recordings from 20 recent smartphone models to advance the field of audio source forensics. By capturing diverse audio types in a strictly controlled anechoic environment, the dataset provides a reliable, reproducible benchmark for identifying the specific smartphone model used to make a recording.
What Problem Does This Paper Address?
In audio forensics, determining the "source attribution"—or exactly which device model recorded a specific audio track—is crucial for verifying digital integrity and providing legal evidence. While modern machine learning models excel at this multi-class classification task, they are highly data-hungry.
Existing datasets suffer from three major flaws:
- Obsolescence: They feature smartphones released years ago, failing to reflect modern hardware trends.
- Inconsistency: They are often recorded in uncontrolled environments with varying background noise, making it hard to isolate the microphone's actual acoustic signature.
- Low Sampling Rates: Many older corpora cap audio at 8 or 16 kHz, ignoring high-frequency data.
POLIPHONE addresses these gaps by providing a standardized, high-fidelity dataset of modern smartphones (released from 2018 onwards) across multiple audio domains.
Methods
To eliminate confounding environmental factors, the researchers recorded all data in a standardized, ISO3745-compliant anechoic chamber.
The core methodology includes:
- Controlled Playback: Audio tracks were played through a high-quality Genelec 8020C loudspeaker, with smartphones suspended 3 meters away to ensure far-field recording conditions.
- Raw Capture: Audio was captured in its raw form (44.1 kHz, 16-bit WAV PCM) using the internal mono microphone at the bottom edge of each phone, avoiding any software compression or automatic gain control.
- Impulse Response (IR) Extraction: The authors recorded a 5-second sine sweep (20 Hz to 20 kHz) on each device. By convolving this sweep with its mathematical inverse, they isolated the unique linear Impulse Response of each phone's microphone hardware.
- Post-Processing Pipeline: To prevent classifiers from simply exploiting trivial differences in volume or dynamic range, the authors created a normalized version of the dataset using a standardization step followed by a non-linear
tanhtransformation to compress extreme values.
Datasets
The paper introduces the POLIPHONE dataset, which spans 10 brands and 20 distinct smartphone models (e.g., iPhone 13, Huawei P30, Xiaomi 12). It includes approximately 150 minutes of audio per smartphone, divided into three distinct versions:
original_recordings: Unaltered raw data at 44.1 kHz.normalized_data: Uniformly scaled data, available at both 44.1 kHz and 16 kHz.convolved_data: Clean source audio mathematically convolved with the extracted smartphone IRs to simulate recording pipelines.
The audio content itself draws from three well-known open-source corpora to cover diverse acoustic domains:
- Speech: 13 utterances from 30 different speakers from the VCTK corpus (balanced for gender and accents), recorded in both clean conditions and noisy conditions (with 30 dB Additive White Gaussian Noise).
- Music: 64 symphonic music excerpts from the ORCHSET dataset.
- Environmental Sounds: The complete ESC-10 subset, covering 10 distinct classes of environmental noise.
Key Insights and Ideas
- The Value of High Frequencies: The most critical hardware signatures for identifying smartphone models reside in the high-frequency bands above 8 kHz. Standard downsampling to 16 kHz severely limits a classifier's accuracy and data efficiency.
- Hidden Hardware Anomalies: The controlled setup exposed unexpected manufacturer design choices. For example, the Xiaomi 12 and 12 Pro completely clip all audio content above 8 kHz, despite recording at a 44.1 kHz sampling rate. Meanwhile, the Moto G9 Power features an aggressive, un-bypassable Voice Activity Detector (VAD) that zeroes out the audio signal entirely during moments of silence.
- Cross-Domain Generalization is Hard: Classifiers trained purely on one audio domain (like clean speech) struggle when tested on another (like music). However, training models on environmental noise (the ESC-10 data) yielded the most robust, versatile classifiers across all test domains.
Key Results
The authors benchmarked POLIPHONE using a state-of-the-art Convolutional Neural Network (CNN) classifier that processes log-spectrograms:
- Sampling Rate Impact: When trained on 44.1 kHz clean speech, the baseline CNN achieved nearly 100% accuracy, even when utilizing only 10% of the available training data. Conversely, at 16 kHz, the maximum accuracy topped out at 92% and dropped sharply as training data volume decreased.
- Cross-Domain Heatmaps: Testing across different audio domains revealed strong diagonal performance (exceeding 92% accuracy when training and testing domains matched). Noisy speech proved to be the easiest domain to classify during testing, regardless of what the model was trained on.
- IR Simulation Efficacy: Training models on the synthetic
convolved_dataresulted in an average accuracy drop of only 9% compared to real recordings. This proves that extracted Impulse Responses can successfully scale up training sets when physical devices are unavailable.
Limitations and Future Work
While POLIPHONE provides an excellent foundation for model identification, it currently features only one physical device per smartphone model. This means the dataset cannot yet be used for device fingerprinting (distinguishing between two different physical units of the exact same phone model).
The authors plan to address this in future iterations by expanding the corpus to include multiple physical devices per model, integrating newer smartphone releases, and introducing additional audio domains.
Summary updated