Source-linked AI summary
RawBoost: A Raw Data Boosting and Augmentation Method applied to Automatic Speaker Verification Anti-Spoofing
Hemlata Tak, Madhu Kamble, Jose Patino, Massimiliano Todisco, Nicholas Evans
TL;DR
ASVspoof 2021 required spoofing detectors to generalize to telephony encoding and transmission conditions absent from available training data, while existing augmentation methods did not support raw-waveform systems. RawBoost addresses this by perturbing raw audio with signal-processing-based convolutive and additive noise; it improves a RawNet2 baseline by up to 27% relative without external data or model-level intervention.
Problem
ASVspoof 2021 required reliable spoofing detection under unseen telephony encoding and transmission conditions, but available training data lacked those effects and common augmentation methods targeted two-dimensional representations.
Method
RawBoost boosts raw-waveform training data by applying established linear and non-linear convolutive and additive signal-processing perturbations without additional data sources.
Results
Up to 27% relative improvement was obtained over a raw end-to-end baseline spoofing detector, with RawBoost strategies outperforming the baseline across all seven evaluation conditions.
Takeaways & Limitations
RawBoost is data, application, and model agnostic and can generate telephony-like variability from an existing source database for related audio classification tasks.
Takeaways & Limitations
The literature comparison is restricted to single systems, and differences reflect both data augmentation and underlying model or classifier choices.
Abstract
from arXiv · showhide
This paper introduces RawBoost, a data boosting and augmentation method for the design of more reliable spoofing detection solutions which operate directly upon raw waveform inputs. While RawBoost requires no additional data sources, e.g. noise recordings or impulse responses and is data, application and model agnostic, it is designed for telephony scenarios. Based upon the combination of linear and non-linear convolutive noise, impulsive signal-dependent additive noise and stationary signal-independent additive noise, RawBoost models nuisance variability stemming from, e.g., encoding, transmission, microphones and amplifiers, and both linear and non-linear distortion. Experiments performed using the ASVspoof 2021 logical access database show that RawBoost improves the performance of a state-of-the-art raw end-to-end baseline system by 27% relative and is only outperformed by solutions that either depend on external data or that require additional intervention at the model level.
1. INTRODUCTION
ASVspoof 2021 required reliable spoofing detection under unseen telephony encoding and transmission conditions, while available training data lacked those effects. RawBoost addresses this gap with raw-waveform augmentation that is model- and data-agnostic.
- ASVspoof 2021 evaluated logical-access spoofing detection when bona fide and spoofed utterances underwent telephony encoding and transmission.
- Training and development data lacked encoding and transmission effects, motivating augmentation to compensate for missing in-domain conditions.
- Generalised countermeasures had to handle unseen TTS and voice-conversion attacks as well as unseen encoding and transmission conditions.
- RawBoost applies data boosting and augmentation directly to raw audio to improve reliability under telephony nuisance variation.
- RawBoost requires no additional data sources or model-level intervention and is data, application and model agnostic.
2. DATA AUGMENTATION
Data augmentation generates additional utterances from source data and is widely used to improve machine-learning systems. However, spectral methods such as SpecAugment are not easily applicable to raw waveforms, creating a need for raw-level augmentation.
- Data augmentation generates new utterance samples from a source database to expand the training pool.
- Augmented variability can help reduce overfitting and bias and improve classification performance.
- Data augmentation has been applied across image processing, speech recognition, speaker verification, and anti-spoofing.
- ASVspoof 2021 participants used speed perturbation, SpecAugment, and codec augmentation to address unknown codecs in evaluation data.
- SpecAugment masks spectral or temporal regions but cannot be applied easily at the waveform level.
3. RAWBOOST DATA BOOSTING AND AUGMENTATION
RawBoost operates on raw waveforms by perturbing training utterances with three independent nuisance-simulation processes. These processes model convolutive distortion, impulsive signal-dependent noise, and stationary signal-independent noise using established signal-processing techniques.
- RawBoost boosts or augments training data at the raw waveform level using established linear and non-linear signal-processing techniques.
- Linear and non-linear convolutive noise: Linear and non-linear convolutive noise models stationary channel distortion and higher-order harmonic distortion from encoding and transmission processes.
- Linear and non-linear convolutive noise: Multi-band notch filters generate convolutive noise with randomly selected center frequencies, filter widths, gains, and coefficient counts.
- Linear and non-linear convolutive noise: The convolutive-noise process applies multi-band filter coefficients to a raw waveform, with j indexing the order of the linear or non-linear component.
- Impulsive signal-dependent additive noise: Impulsive signal-dependent noise represents non-stationary disturbances associated with clipping, device operation, synchronization, overflow, and computational limitations.
- Stationary signal-independent additive noise: Stationary signal-independent additive noise models effects such as cable faults, transmission channels, electromagnetic interference, and thermal noise.
4. EXPERIMENTS AND RESULTS
Experiments evaluate RawBoost with an end-to-end RawNet2 baseline on ASVspoof 2021 LA, using codec-specific and pooled detection metrics. RawBoost improves the baseline across all seven evaluation conditions and compares favorably with competing single systems, though comparisons across systems also reflect model differences.
- RawBoost configurations: RawBoost combines three augmentation families: linear and non-linear convolutive noise, impulsive signal-dependent additive noise, and stationary signal-independent additive noise.Configurations apply techniques individually and in selected series or parallel combinations, with parameters randomly drawn from specified ranges.
- Dataset, protocols and metrics: The ASVspoof 2021 LA evaluation measures RawNet2 performance using min t-DCF for each codec, pooled min t-DCF, and pooled EER.The evaluation includes multiple encoding and transmission conditions, including known and unknown telephony codecs.
- Results: All RawBoost strategies outperform the baseline across all seven evaluation conditions.The baseline pooled min t-DCF is 0.4257, compared with 0.3527, 0.3260, and 0.3372 for the three individual augmentation families, respectively.
- Results: 27% relative reduction in pooled min t-DCF is achieved by RawBoost (1)+(2), reaching 0.3099 versus the baseline.The same configuration achieves an EER of 5.31%, a 44% relative reduction over the baseline.
- Comparison to competing systems: RawBoost (1)+(2) gives the third-best result among compared state-of-the-art single systems.The comparison excludes more complex ensemble systems, and differences reflect both augmentation and underlying model or classifier choices.
5. CONCLUSIONS
RawBoost augments existing training data with telephony-relevant waveform variability and improves raw end-to-end spoofing detection without external data or model intervention.
- 27% relative improvement was achieved over a raw end-to-end baseline spoofing detection solution.