Enhancement of ADPCM Speech by Adaptive Postfiltering

Abstract

Adaptive Differential Pulse Code Modulation (ADPCM) systems can provide high‐quality digitizations of telephone‐bandwidth speech at a bit rate of 32 kb/s. At a lower bit rate such as 24 kb/s, the quality of the speech is limited by an easily perceptible level of quantization noise. This paper proposes an adaptive postfiltering procedure that can significantly enhance the quality of lower bit rate ADPCM. The coefficients of the postfilter are easily derivable from the predictor coefficients in the ADPCM decoder. In a subjective test involving 18 listeners and two sentence‐length test inputs, the enhanced 24‐kb/s speech with an optimized postfilter design ranks very close to conventional 32‐kb/s speech. A suggested application of the postfiltering procedure is in packet voice or mobile radio systems where substandard bit rates such as 24 kb/s or 16 kb/s are sometimes necessary. The postfiltering algorithm has also been successfully tested in non‐DPCM situations, such as in the enhancement of speech degraded by additive white Gaussian noise.

Convolutional Neural Networks to Enhance Coded Speech

“…All results turned out to be significant. We can see a clear 1.76 7 Note that this manipulation naturally also degrades the instrumental values as given in Section V-B1. For structure III in Tab.…”

“…First of all, windowing of the coded speech s(n) is implemented to form frames for processing, which can be denoted as (7) where N s is the frame shift, N w is the length of window function, w is the window function vector, and • denotes the sample-wise multiplication. As shown in Fig.…”

“…The fourth is the In a preliminary informal subjective listening test we observed that an ideally very low 0-th cepstral coefficient turns out to assume slightly higher values after the CNN estimation, resulting in somewhat noisy speech pauses. Therefore, for the subjective listening test, we very slightly manipulate the CNN output as follows 7 ĉenv (ℓ, 0) → ĉenv (ℓ, 0), if ĉenv (ℓ, 0) C 0 ĉenv (ℓ, 0) − γ 0 , else,…”

“…However, as the bitrates go down for most of the modern codecs, it becomes more difficult for the classical Wiener filter to effectively suppress the quantization noise, while maintaining the speech perceptually undistorted, since the SNR drops and more importantly, only the mean squared error (MSE) is minimized in the Wiener filter [7]. Therefore, some perceptually-based postfilters have been proposed to reduce the perceptual degradation caused by low bitrate codecs.…”

Speech Coding: Fundamentals and Applications

“…Despite the use of perceptually weighted error minimization, the synthesized speech coming from an LPC-AS coder may contain audible quantization noise. In order to minimize the perceptual effects of this noise, the last step in the decoding process is often a set of adaptive postfilters [11,80]. Adaptive postfiltering improves the perceptual quality of noisy speech by giving a small extra emphasis to features of the spectrum that are important for human-to-human communication, including the pitch periodicity (if any) and the peaks in the spectral envelope.…”