Deep Learning Based Audio Watermarking with Controlled Distortion and Reliable Bit Recovery Using CRNN-BiLSTM Architecture

Abstract

This work introduces a deep learning based audio watermarking system comprising a single head attention mechanism, a Convolutional Recurrent Neural Network (CRNN) encoder, and a Bidirectional Long Short Term Memory (BiLSTM) decoder. The method uses RMS-scaled, hidden perceptual embedding to embed a 32-bit binary payload into the low-frequency sub band of a derivable 1-D Haar discrete wavelet transform (DWT). Over 80 training epochs, a four-phase progressive learning strategy noise layer approach spanning 19 distinct attack types is applied, covering AWGN, echo, low-pass filtering, time stretching, pitch shifting, dynamic range compression (DRC), MP3/Opus/VoIP codec simulations, and ambient reverberation. Trained on 12,000 FMA-Medium samples and evaluated on FMA-Medium and GTZAN with 999 samples each, the proposed framework achieves 98.54% bit-recovery accuracy on FMA-Medium and 97.83% on GTZAN under clean conditions at a signal-to-noise ratio (SNR) of 31.52 dB. Twenty four of twenty nine evaluated attack scenarios surpass the 80% robustness threshold, demonstrating the effectiveness of the curriculum based training strategy across diverse real world distortions.