End-to-end recurrent denoising autoencoder embeddings for speaker identification

Speech ‘in-the-wild’ is a handicap for speaker recognition systems due to the variability induced by real-life conditions, such as environmental noise and emotional state of speaker. Taking advantage principles representation learning, we aim design recurrent denoising autoencoder that extracts robust embeddings from noisy spectrograms perform identification. The end-to-end proposed architecture uses feedback loop encode information regarding into low-dimensional representations extracted spectrogram autoencoder. We employ data augmentation techniques additively corrupting clean speech with in database containing real stressed speech. Our study presents joint optimization both denoiser identification modules outperforms independent components under stress distortions well handcrafted features.