Alan B. Bradley

Subband/Transform coding using filter bank designs based on time domain aliasing cancellation

A new, oddly stacked, critically sampled, single side-band (SSB) [7] analysis/synthesis system based on Time Domain Aliasing Cancellation (TDAC) [1],[2] is described in this paper. The specifications for the analysis and synthesis filter responses are developed and a number of designs which satisfy the reconstruction requirements are described. The application of TDAC systems to Subband/Transform coding is also discussed and the objective performance of a 32 band coder using several different window designs is presented and compared with a coder based on Frequency Domain Aliasing Cancellation (FDAC) filter banks [3]-[5].

An improved algorithm for vector quantizer design

Vector quantization is an essential tool in signal processing. Although many algorithms for vector quantizer design have been developed, the classical generalized Lloyd algorithm (GLA) is still widely used, mainly for its simplicity and relatively good performance. Recently, Lee et al. (see IEEE Signal Processing Lett., vol.4, p.2-4, Jan. 1997) proposed an intuitive modification of the K-means algorithm (MKMA). In this letter, we propose an improved algorithm that uses the standard GLA embedded in a codevector reassignment loop. The algorithm achieves better performance by targeting specifically those sections of the codebook that contribute most to the overall reconstructed signal distortion. Simulations show that the proposed algorithm outperforms both the traditional GLA and MKMA in a variety of scenarios.

Optimization of a temporal decomposition model of speech

A dynamic programming based optimization strategy for a modified temporal decomposition (TD) model of speech is presented. In previous work with the SBEL-TD algorithm, the event localization was performed based on a spectral stability criteria. Although this approach gave reasonably good results (spectral distortion of about 1.5 dB), there was no assurance on the optimality of event locations. In this present work we have optimized the event localization task using a dynamic programming based optimization strategy. An overlapping buffering technique is also proposed to ensure smooth transitions between consecutive speech parameter blocks. Simulation results show that an improved TD model accuracy in terms of the spectral distortion (0.9 dB) can be achieved. The new optimised algorithm also adds a new degree of freedom, TD resolution (events/sec), to the analysis paradigm which can be effectively used to control the TD model accuracy.

Discriminative feature extraction applied to speaker identification

Speaker recognition systems typically consist of two individual modules providing feature extraction and classification. Conventional designs utilise a fixed feature extraction algorithm while a stochastic classifier is adapted during a training phase. Data-driven feature extraction involves adaptation of the feature extraction process in addition to the classifier during training. Discriminative feature extraction (DFE) is a data-driven feature extraction technique previously applied to speech recognition. This paper reports the application of DFE to the design of a filterbank for a Gaussian mixture model (GMM) based speaker identification system. The DFE trained filter-bank is shown to outperform conventional fixed filter-bank feature extraction.

Minimising the effects of subband quantisation of the time domain aliasing cancellation filter bank

The effect of the quantisation of filter bank subbands has been analysed by incorporating quantisation noise models into the time domain aliasing cancellation (TDAC) filter bank. We have found expressions for the reconstruction error of the quantised TDAC system in terms of several signal correlated components, and an uncorrelated component. These expressions allow easy identification of subjectively annoying errors, and provide the framework for a subjective optimisation of the quantisation process. Research has been carried out on alternative quantiser models and methods of quantiser-compensation.

Overlap-save convolution applied to wavelet analysis

The discrete wavelet transform (DWT), wavelet packet transform (WPT), and M-band wavelet techniques are often implemented as a cascade of critically sampled filter banks. Many applications apply these techniques to finite frames of samples, leading to distortion in the filtered coefficients near the frame boundaries. Overlap-save convolution (OSC) eliminates boundary distortion for a single filtering process. This article reports an application of OSC for cascaded filter banks, eliminating boundary distortion in the frame-based application of the DWT, WPT, and M-band wavelet techniques.

Speech compression by vector quantization of epochs

A pitch epoch is a fundamental unit of voiced speech. This paper introduces a speech compression method based on vector quantization of epochs. Pitch determination, epoch marking, vector quantization procedure, and a technique for epoch extrapolation are described. The compression method is then evaluated and briefly compared to other waveform coders. The quality is objectively measured by the segmental signal-to-noise ratio and the results are tabulated. The (automatic) epoch vector quantization yields the following SNRseg: 10.03 dB at 12.0 kbps, 11.35 dB at 21.3 kbps, 13.31 dB at 39.7 kbps, 18.32 dB at 76.4 kbps, and 62.69 dB at 112.9 kbps.