Gaussian-Filtered High-Frequency-Feature Trained Optimized BiLSTM Network for Spoofed-Speech Classification

Area of Science:

• Speech processing
• Machine learning
• Cybersecurity

Background:

• Voice-controlled devices require robust security against speech spoofing attacks.
• Existing spoof detection methods need improvement due to unknown attack algorithms.
• High-frequency speech features show promise in distinguishing genuine from spoofed speech.

Purpose of the Study:

• To develop an effective spoof speech detection model using high-frequency features.
• To investigate the efficacy of Gaussian filters for extracting inverted Mel-frequency cepstral coefficients (iMFCC).
• To optimize a Bidirectional Long Short-Term Memory (BiLSTM) network using a Bayesian algorithm for improved spoof classification.

Main Methods:

• Extraction of high-frequency iMFCC features using a Gaussian filter.
• Integration of complementary features with iMFCC to enhance discrimination.
• Optimization of a BiLSTM network architecture and hyper-parameters via a Bayesian algorithm.
• Training and evaluation on the ASVSpoof 2017 dataset.

Main Results:

• The optimized BiLSTM model achieved 99.58% validation accuracy with minimal epochs.
• The proposed algorithm attained a 6.58% Equal Error Rate (EER) on the evaluation dataset.
• A relative improvement of 78% was observed compared to a baseline spoof-identification system.

Conclusions:

• High-frequency features, particularly iMFCC extracted with Gaussian filters, are effective for spoof speech detection.
• Optimized deep learning models, like the Bayesian-tuned BiLSTM, significantly enhance spoof classification accuracy.
• The proposed method offers a substantial advancement in securing voice-controlled systems against sophisticated spoofing attacks.