Linear Prediction-based Part-defined Auto-encoder Used for Speech Enhancement

Autoregressive Parameter Estimation with Dnn-Based Pre-Processing

“…An alternative to these classical methods is to use a purely data-driven approach for computing the AR-spectrum. Although originally introduced for speech enhancement, the recently proposed PAE in [5] can easily be modified to be an example of such a data-driven approach. For the modified PAE, the functions by training a DNN convert the log-periodogram of a signal segment x into a set of p reflection coefficients (to ensure stability).…”

“…The AR coefficients play an important role in many speech applications such as speech recognition [1], coding [2,3], and enhancement [4,5]. Therefore, the estimation of the AR parameters from an observed signal segment has been a classical signal processing problem, and many different estimators have been proposed over many decades.…”

“…For voiced speech, however, the excitation signal does not resemble a white, Gaussian excitation signal as in the autoregressive process, but is much more accurately modelled by an impulse train [3]. As a consequence of this, many alternative ways of estimating the AR-parameters have been proposed based on the prior knowledge on the power spectral density (PSD) [10,11,5] or the excitation signal [3,12,13]. For example, El-Jaroudi and Makhoul proposed in [10] the discrete all-pole (DAP) approach in which the AR-parameters are estimated by minimising the Itakura-Saito (IS) divergence for a discrete set of points, leading to better performance for voiced speech.…”

“…In addition to classical signal processing methods, datadriven approaches to estimating the AR-parameters have also been proposed [5,14]. One recent example of this is the part-defined auto-encoder (PAE) [5] in which the analysisby-synthesis (ABS) strategy [15] is combined with an autoencoder [16].…”

“…In addition to classical signal processing methods, datadriven approaches to estimating the AR-parameters have also been proposed [5,14]. One recent example of this is the part-defined auto-encoder (PAE) [5] in which the analysisby-synthesis (ABS) strategy [15] is combined with an autoencoder [16]. In this approach, a DNN is trained to learn the mapping from the raw data to the reflection coefficients which can easily be translated into the corresponding ARparameters.…”