Related Experiment Video
Updated: Aug 4, 2025
09:09
Foreign Accent and Forensic Speaker Identification in Voice Lineups: The Influence of Acoustic Features Based on Prosody
Published on: September 27, 2024
483
Mel frequency spectral domain defenses against adversarial attacks on speech recognition systems
Nicholas Mehlman1, Anirudh Sreeram1, Raghuveer Peri1
1Ming Hsieh Department of Electrical and Computer Engineering, University of Southern California, Los Angeles, California 90089, USA nmehlman@usc.edu, asreeram@usc.edu, rperi@usc.edu, shri@usc.edu.
JASA Express Letters
|April 1, 2023
Summary
This study introduces mel domain noise flooding (MDNF), a novel defense against adversarial attacks on automatic speech recognition (ASR) systems. MDNF enhances ASR robustness by adding noise to speech features before re-synthesis.
Area of Science:
- Speech Processing
- Machine Learning Security
- Cybersecurity
Background:
- Automatic speech recognition (ASR) systems are susceptible to adversarial attacks.
- Existing defenses often adapt image-domain methods, neglecting speech-specific vulnerabilities.
Purpose of the Study:
- To explore speech-specific defenses in the feature domain for ASR systems.
- To introduce and evaluate a novel defense method called mel domain noise flooding (MDNF).
Main Methods:
- Developed MDNF, a defense that injects additive noise into the mel spectrogram representation of speech.
- Re-synthesized the audio signal with modified features as input to the ASR system.
- Evaluated MDNF against strong white-box adversarial threat models.
Main Results:
- MDNF demonstrated competitive robustness against sophisticated adversarial attacks.
- The proposed speech-specific defense shows promise in enhancing ASR security.
Conclusions:
- Mel domain noise flooding (MDNF) offers an effective, speech-specific defense strategy for ASR systems.
- Feature-domain defenses are a viable approach to improving ASR resilience against adversarial manipulation.
Related Concept Videos
01:03IR Frequency Region: Fingerprint Region
987
IR spectra are divided into two main regions: the diagnostic region and the fingerprint region. The diagnostic region of the spectrum lies above 1500 cm−1. The absorptions resulting from single-bond vibrations of the N–H, C–H, and O–H stretch at higher wavenumbers and appear on the left side of the spectrum. The stretching absorptions of the C≡C and C≡N occur between 2100–2300 cm−1. In contrast, those arising from stretching absorptions of the...
987
01:18Aliasing
178
Accurate signal sampling and reconstruction are crucial in various signal-processing applications. A time-domain signal's spectrum can be revealed using its Fourier transform. When this signal is sampled at a specific frequency, it results in multiple scaled replicas of the original spectrum in the frequency domain. The spacing of these replicas is determined by the sampling frequency.
If the sampling frequency is below the Nyquist rate, these replicas overlap, preventing the original...
If the sampling frequency is below the Nyquist rate, these replicas overlap, preventing the original...
178
01:24Frequency-Domain Interpretation of PD Control
155
Proportional-Derivative (PD) controllers are widely used in fan control systems to improve stability and performance. A fan control system can be effectively represented using a Bode plot to illustrate the impact of a PD controller through its transfer function. The Bode plot visually conveys how PD control modifies the fan's response across various frequencies, providing a frequency domain interpretation of the controller's behavior.
The proportional control gain, combined with the...
The proportional control gain, combined with the...
155
01:02¹H NMR: Interpreting Distorted and Overlapping Signals
1.1K
Spin systems where the difference in chemical shifts of the coupled nuclei is greater than ten times J are called first-order spin systems. These nuclei are weakly coupled, and their chemical shifts and coupling constant can generally be estimated from the well-separated signals in the spectrum.
As Δν decreases and the signals move closer, the doublets appear increasingly distorted. The intensities of the inner lines increase at the cost of those of the outer lines as the signals are...
As Δν decreases and the signals move closer, the doublets appear increasingly distorted. The intensities of the inner lines increase at the cost of those of the outer lines as the signals are...
1.1K
01:26Linear Approximation in Frequency Domain
120
Linear systems are characterized by two main properties: superposition and homogeneity. Superposition allows the response to multiple inputs to be the sum of the responses to each individual input. Homogeneity ensures that scaling an input by a scalar results in the response being scaled by the same scalar.
In contrast, nonlinear systems do not inherently possess these properties. However, for small deviations around an operating point, a nonlinear system can often be approximated as linear....
In contrast, nonlinear systems do not inherently possess these properties. However, for small deviations around an operating point, a nonlinear system can often be approximated as linear....
120
01:20¹³C NMR: Distortionless Enhancement by Polarization Transfer (DEPT)
1.1K
When proton-coupled carbon-13 spectra are simplified by a broadband proton decoupling technique, structural information about the coupled protons is lost. Distortionless enhancement by polarization transfer (DEPT) is a technique that provides information on the number of hydrogens attached to each carbon in a molecule. While the DEPT experiment utilizes complex pulse sequences, the pulse delay and flip angle are specifically manipulated. The resulting signals have different phases depending on...
1.1K