On supervised LPC estimation training targets for augmented Kalman filter-based speech enhancement

Abstract

The performance of speech coding, speech recognition, and speech enhancement systems that rely on the augmented Kalman filter (AKF) largely depend upon the accuracy of clean speech and noise linear prediction coefficient (LPC) estimation. The formulation of clean speech and noise LPC estimation as a supervised learning task has shown considerable promise as of late. Generally, a deep neural network (DNN) learns to map noisy speech features to a training target that can be used for clean speech and noise LPC estimation. Such training targets fall into four categories: Line spectrum frequency (LSF), LPC power spectrum (LPC-PS), power spectrum (PS), and magnitude spectrum (MS) training targets. The choice of training target can have a significant impact on LPC estimation accuracy. Motivated by this, we perform a comprehensive study of the training targets with the aim of determining which is best for LPC estimation. To this end, we evaluate each training target using a temporal convolutional network (TCN) and a multi-head attention-based network. A large training set constructed from a wide variety of conditions, including real-world non-stationary and coloured noise sources over a range of signal-to-noise ratio (SNR) levels, is used for training. Testing on the NOIZEUS corpus demonstrates that the LPC-PS as the training target produces the lowest clean speech LPC spectral distortion (SD) level. We also construct the augmented Kalman filter (AKF) with the estimated speech and noise LPC parameters of each training target. Subjective AB listening tests and seven objective quality and intelligibility evaluation measures (CSIG, CBAK, COVL, PESQ, STOI, SegSNR, and SI-SDR) revealed that the LPC-PS training target produced enhanced speech at the highest quality and intelligibility amongst the training targets.